Skip to content
Zabio is in beta: models, prices and features can change. Tell us what breaks at hello@zabio.ai.

Changelog

Improvements and fixes in each release, newest first. New features are announced on the roadmap.

  • Conversation items show their parent and response, so a client can rebuild the current thread. Each item of GET /v1/conversations/{id}/items carries zabio.parent_id and zabio.response_id, and GET /v1/conversations/{id} carries zabio.leaf_id. Branching still happens only through previous_response_id.
  • Ask Zabio on every page. A button at the bottom right of every page of zabio.ai, the docs and the dashboard opens Ask Zabio: ask a question about Zabio and get an answer from the docs, with links to the pages it comes from. Under each answer you see the model that answered and what the answer cost, and what the same answer would have cost on a frontier model. It needs no account or key. Press ? to open or close it, or write to hello@zabio.ai to talk to us.
  • A chat in the dashboard. Chat in the menu (app.zabio.ai/chat) talks to your models with your own key, at your usual prices, and keeps each chat as a stored conversation. Pick auto, auto/cheap, auto/best or any model your organization allows. Under each answer you see the model auto picked, what the answer cost, and what the same answer would have cost on the cheapest frontier model. Use this in your code shows the exact API call for the conversation. The first key made at signup now has the Conversations scope as well as Chat; a key without it gets a link to make one that has it.
  • The chat inside the dashboard. Signed in, Chat opens inside the dashboard: the menu becomes a strip of icons beside your chats, so the conversation keeps its width, and it opens out to the full menu and back. Your organization, its credit and the search stay at hand. On a phone, the menu and your chats each open as a panel of their own. Without an account, the chat stays full screen as before.
  • Easy classification jobs on auto go by our measurements of today’s models, and the price comparison counts what each side bills. Our second study of Jev measured it against the text models auto picks today on 180 easy classification jobs: Jev got every field right on 85 percent, level with gpt-6-luna (88 percent), and claude-haiku-5-5 on 81 percent. Zabio’s estimate of the routed model’s price now counts the tokens your JSON schema adds to the input and the thinking of a model whose setting reasons, and its estimate of Jev’s price follows TypeSafe’s current count, so of two models measured level the one Zabio estimates cheaper is now much closer to the one that bills less. On our test requests about 6 percent of easy jobs now go to gpt-6-luna, each one billed at or below Jev’s price, and Jev answers the rest. The same estimates apply to every job Jev can answer on auto, to auto lines inside a batch, and to POST /v1/public/analyze. See Classification jobs on auto.
  • The AI Firewall’s model check reads more of an image and of a PDF. With the paid tier, and on the AI Firewall check with the model check: text drawn a few shades from a flat background (white on white, a banner’s own colour) is made readable before the image is read. On 44 such images with a published injection the check caught 43; it caught 15 before. Benign text hidden that way is flagged more often too. A JPEG that carries an EXIF orientation is turned the way models turn it. A scanned PDF is now read by its pages’ images when they are stored as JPEG, the first 8 of a document, each counted and billed as an image; before, a scan was not read. A document block’s title and context on /v1/messages are read as text, each reported at its own path. An image checked before this release is checked once more. See the AI Firewall.
  • Medium classification jobs on auto go by our measurements. Our second study of Jev measured it against the text models auto picks today, on 56 classification jobs rated medium: Jev got every field right on 82 percent, level with gpt-6-luna (80 percent) and ahead of claude-haiku-5-5 (70 percent). A classification job rated medium is now answered by the line’s rule on those measurements, as easy ones already were. Jev was the cheaper one on every medium request we tested, so it answers these jobs as before. See Classification jobs on auto.
  • Two more hosts on Alibaba Cloud, and qwen4-max in full. deepseek-v4.1-flash-1017@alibaba-intl and kimi-k3.1-pro@alibaba-intl are new. Both are processed outside mainland China, and you call them by name. qwen4-max now takes images and tools, has its 1,000,000-token context and 131,072-token output, and bills cached input at Alibaba’s 0.075 USD per million tokens instead of the full input price. Prices and limits are on the models page.
  • A lower cache price for DeepSeek V4.1 Flash on Alibaba Cloud. Alibaba now prices cached input on deepseek-v4.1-flash at a fixed 0.006 USD per million tokens in peak hours and 0.003 off-peak, a tenth of what Zabio billed. Zabio’s hourly price check reads the new price, and calls to deepseek-flash@alibaba-intl are billed at it from about a day after this release.
  • Edit an image from its link. The url of an image made with response_format: url can be sent back as an image_url on a JSON edit, the way it came, for the 60 minutes its link lasts. Zabio does not fetch the address: it checks that the link is one Zabio signed and reads the image from your organization’s files, so your key must be able to read it. An expired link answers 400 expired; any other web address is still refused. See Edit an image from its link.
  • Spoken answers from Qwen Omni. Alibaba’s qwen3.5-omni-plus, qwen3.5-omni-plus-2026-03-15 and qwen3.5-omni-flash answer in audio on chat completions, processed outside mainland China. They answer in audio only when the request streams (an unstreamed request with the audio modality gets 400 unsupported_value on stream), in pcm16, with 56 voices. An answer that carries audio bills its audio alone: its text is free, as Alibaba bills it. qwen3.5-omni-flash-2026-03-15 answers in text only, because it sends no audio. On any model, audio tokens a vendor reports for an answer that carried no audio are not billed. See Audio answers.
  • Rotate an API key. Rotate on the Keys page, zabio keys rotate or POST /admin/v1/orgs/{org}/keys/{key}/rotate gives a key a new secret, shown once, and keeps everything else: its id, name, project, settings, limits, models and usage history. The old secret stops working at once. A connected app’s key gets a new one by reconnecting the app.
  • A failed background response returns its charge. A background response that runs outside a batch and ends failed after its answer was billed (a restart between the bill and storing the answer, or an answer that could not be stored) now has what the turn was billed returned to your balance. The response’s error says so, with the amount, and your ledger shows the returned charge as an adjustment.
  • The AI Firewall check on its own. POST /v1/guard/check checks texts, images and PDFs for prompt injection without calling a model: the patterns for free, up to 60 checks a minute per key, and the model check at 0.50 USD per million tokens checked, with an optional task it reads each item beside. The answer says what was found, where, and what could not be read. POST /v1/moderations is the same check in OpenAI’s moderations shape, so an OpenAI SDK’s moderations.create() works unchanged; it judges prompt injection only. Zabio keeps none of the text. See AI Firewall check.
  • Check for prompt injection from an AI app. Zabio’s MCP server has a check_injection tool: an app connected to https://mcp.zabio.ai/mcp can check a fetched page, a mail, a file or a tool’s output with the AI Firewall check before it acts on it, under the connection’s key and its daily limit. The patterns are free; the model check needs a connection allowed to ask. See the tools.
  • Existing keys follow the organization’s default action. An API key made before the organization’s default prompt-injection action existed now follows it too, as a new key does, unless an owner or admin had set the key’s own action, which it keeps. In an organization whose default is Tag, as it is until an owner or admin changes it, nothing changes for such a key; in one whose default is Block, the key now blocks. A key that follows shows guard_action: null. See the guide.
  • A default action for the prompt-injection guard. An organization now sets the action its keys take on a pattern match, Tag or Block, in AI Firewall, Prompt Injection or with action on PATCH /admin/v1/orgs/{org}/guard. A key follows it unless it sets its own action; a new key follows it, and an existing key keeps tag as its own action until you set it to follow (guard_action: null, or Organization in the key’s Action). A key that follows shows guard_action: null, and every key shows the action in force as guard_action_effective. See the guide.
  • Allow a text by typing it. Owners and admins can now add an allowed text before any call is flagged: Add text in AI Firewall, Prompt Injection, or POST /admin/v1/orgs/{org}/guard/allowlist/text. The patterns read the text and Zabio keeps the rule and a hash of each match, never the text. An entry can carry a note. Only a web session can do it: no token, API key or MCP app.
  • Room for an answer on auto. A request to auto with no max_tokens now goes only to models whose context window leaves at least 4,096 tokens for the answer beside the input, after the 5 percent Zabio keeps free. Before, a long prompt could go to a model whose window it nearly filled, with almost no room for the answer. When no model auto can use leaves that room, the request gets 400 context_length_exceeded, and the message says so.
  • DeepSeek models near their context window. DeepSeek now refuses a request whose input plus output limit passes the window. Close to the window Zabio now sends DeepSeek models the room the window leaves beside the input, as on Together AI models; with less than 4,096 tokens of room the request is too long for the model. Before, such a request was sent the model’s maximum output and DeepSeek refused it.
  • max_tokens near the context window. On Anthropic models the max_tokens you set is sent as you set it, within the model’s maximum output, and with none set the model’s maximum output is sent at every input size: Anthropic takes input plus max_tokens over the window and stops at the window. On other models a max_tokens too large for the window is lowered to the room the window leaves, but no longer below 4,096 tokens (or your own value, when smaller); with less room than that the request is too long for the model. Before, it could be lowered to a few tokens.
  • An empty answer at the context window says so. When a model reasons until its context window ends and writes nothing, zabio.no_answer now says the input left no room for an answer, instead of asking for a higher max_tokens.
  • Mark text you did not write. A text part of a user message can carry "zabio": {"untrusted": true} on chat completions, the Responses API and /v1/messages, for a forum post, an email or a page you paste in. The prompt-injection guard then reads it as it reads a tool result: the patterns use the tool result rules, and the paid tier’s model check reads it beside your user’s own words. The model receives the text without the mark, and the price does not change. A mark anywhere else is refused with 400 invalid_value. See Text you did not write.
  • What the guard did not read. zabio.guard.not_read and the new header x-zabio-guard-unread count the images, audio, video, files and system messages of a request, and with the paid tier the earlier messages, assistant turns and trusted tools’ results its model check does not read. The header ends with the number of strings the model check could not read, or pending when the headers were sent before it finished. A long text the model check flagged while part of it could not be read is now also listed in zabio.guard.unchecked, with partial: true.
  • The model check reads images and documents. With the paid tier, the AI firewall’s model check now looks at the PNG and JPEG images in tool results on /v1/messages and at images you mark with "zabio": {"untrusted": true} on any API, at most 8 per request, and reads the text of PDF and plain text documents on /v1/messages, hidden text included. Both are read as tool results and priced as tokens checked, as text is. An image given by URL, a GIF, a WebP and the pages of a scanned PDF are not read and are counted in zabio.guard.not_read; what could not be read is listed in zabio.guard.unchecked, with the new reasons image_limit and unsupported. See The paid tier.
  • Background responses keep their estimate set aside until they finish. A background response (background: true on POST /v1/responses) that runs outside a batch now keeps its estimated cost set aside from the moment it is queued until it is billed or ends, as a synchronous call does while it runs. A turn whose estimate does not fit the balance left after the calls in flight, queued background turns included, is refused with 402 insufficient_credits when it is queued. Cancelling a queued turn frees its estimate at once.
  • A background response whose call failed says failed. A background response whose model call failed now ends with status: "failed" and the error of the call. Before, it ended cancelled, with a message saying it was cancelled before it ran.
  • Trial credit and background responses. For an organization on trial credit alone, a model of the frontier class named with background: true is now refused with 402 trial_model_restricted when it is queued, as a synchronous call is. A background response that runs outside a batch is also held to the other trial terms when it runs: auto does not pick a frontier model, and a turn whose estimate is above the trial balance ends failed with insufficient_credits before it reaches the model. Before, such a turn was queued and run on the model it named.
  • A background response interrupted after its answer was billed is not run twice. When a restart interrupted a background response that runs outside a batch after the call that wrote its answer was billed and before the answer was stored, the call was made and billed a second time. It is no longer made again: the response ends failed.
  • A background response with a deadline of 24 hours or more needs its estimate once. When such a turn joined your organization’s batch, it was refused with 402 insufficient_credits unless the available balance covered its estimate at the standard price and its estimate at the batch price together. Now the balance has to cover each of the two alone.
  • max_output_tokens on a background response. The object of a background response now shows the max_output_tokens you set, when it is queued and each time you read it. Before, it said null.
  • A time limit per request. A call to chat completions, /v1/messages or the Responses API now ends after 20 minutes without an answer, or, when streamed, after 10 minutes without the first piece of the answer. A stream that has started sending its answer is never cut. Compaction of a Responses conversation counts within the same limit; a compaction it cuts answers request_timeout, where it answered 500 internal_error. Zabio starts no new attempt with under a minute left. You get the error of the last attempt that failed, or request_timeout when none had failed, in the body or the stream when the 200 was already sent, and nothing is billed for the attempt that was cut. Before, a request that kept failing over could run for hours, and the keep-alive bytes kept your client’s own timeout from ending it. See Errors.
  • Trial credit and batches. For an organization on trial credit alone, a batch with a line that names a model of the frontier class now fails before it runs, with trial_model_restricted and the line’s number, and auto in a batch does not pick a frontier model, as on a synchronous call. The same holds for a background response with a deadline of 24 hours or more, which runs in your organization’s batch. Before, such a line ran on the model it named, or on a frontier model auto picked, and was billed to trial credit.
  • A background response interrupted between two models pays the first one once. When auto moved a background response to a stronger model and a restart interrupted it before that model answered, the first model’s call was made and billed a second time. The response now goes on at the stronger model, and the first call is billed once.
  • Free and background responses with a deadline of 24 hours or more. On the Free limits such a response on auto, or on a model with a batch tier, was refused with 402 insufficient_credits. It is now queued and served on the Free limits: in your organization’s batch when a free model has a batch tier, as its own job otherwise.
  • Free limits count background responses and batch requests. A background response and each request of a batch served on the Free limits now count toward the account’s requests a day and its fair-use limit, as a synchronous call does. Before, they were not counted there.
  • A request refused for trial credit leaves no unfinished response. A Responses API call refused with trial_model_restricted used to leave its conversation and a response that stayed in_progress for up to an hour and then read failed. Nothing is stored for a refused call now.
  • max_output_tokens when you read a response. GET /v1/responses/{id} now shows the max_output_tokens you set on a completed response that did not run in the background too. Before, it said null.
  • The time limit covers background responses. A background response that runs outside a batch now ends failed with request_timeout when its call gets no answer within 20 minutes, counted from when the call starts, not from when the response was queued.
  • Web search on your bill. The Dashboard and Usage pages now show what web search cost, on the cost card, the Spend chart, Top keys and each call in Logs, and a key’s settings can turn web search off.
  • Web search. Claude and OpenAI models can now search the web with their own search. Send the web_search tool on the Responses API, web_search_options on chat completions, or Anthropic’s web_search_20250305 tool on /v1/messages, on a model whose price list shows a price per search. Each search is billed on its own line, and the results count as input tokens. Web search costs 0.0105 USD a search: the provider’s price of 10 USD per 1,000 searches plus Zabio’s 5 percent. See Web search.
  • Edit an image you generated in the background. A JSON edit (POST /v1/images/edits) now takes the img_ id of a completed background generation as a file_id, and "zabio": {"index": n} in the same entry picks one of its images, so you no longer download an image and send it back to edit it. An image made with response_format: url can be named by its file id while its link works. The edit is billed as any edit; an expired generation or link answers 400 expired.
  • A limit on an edit’s images in all. The images of one edit, including those read from files and generations, can be at most 100 MB together, the limit a multipart edit already had; above it the edit answers 413 media_too_large.
  • Responses usage counts the tokens written to the cache. On Claude models, usage.input_tokens and usage.total_tokens of POST /v1/responses now include the input written to the prompt cache, as prompt_tokens on chat completions always did; cached_tokens stays the tokens read from the cache. Bills do not change. A Responses turn with an image, audio or a video is no longer counted against your rate limit and balance hold as if its text were sent twice.
  • Videos over 20 MB by URL on Qwen models. An https video larger than 20 MB sent to one of Alibaba’s Qwen video models by name is no longer refused: Alibaba fetches it from your URL, up to the 2 GB Alibaba states as its limit, and Zabio reads only the first 1 MB of the file for its format and length. The call is billed on the tokens Alibaba reports, as every call is. Other models and auto keep the 20 MB limit for a video by URL; send them a larger video as a file_id.
  • Audio in streamed Responses calls and batch lines is billed like audio everywhere else. A streamed POST /v1/responses call with an input_audio part now reports its audio tokens in usage.input_tokens and usage.input_tokens_details.audio_tokens and bills them at the model’s audio price, as the same call without stream and chat completions already did. Batch lines with audio bill the same way. Earlier calls are not rebilled.
  • The credit estimate for audio. The credit estimate for a request with audio now prices the audio at the model’s audio price, and a clip whose length cannot be read is estimated from its size. Bills are unchanged.
  • Long text to speech on Gemini. Gemini speech models now answer inputs up to 4,096 characters. A model that stops before the end of the input says so with finish_reason length.
  • Stored videos are delivered. A video that was already made and billed is now delivered when Zabio hits an error while finishing it, instead of being refunded and deleted.
  • Requests close to the context window on Anthropic, DeepSeek and Together AI models. A request that names one of these models, sets no max_tokens and has an input close to the model’s context window used to be moved to a model of the same provider with a larger window, or refused with 400 context_length_exceeded, because the output limit Zabio sends by default was counted against the window. Anthropic and DeepSeek models accept such a request, so it is now served by the model you named; on Anthropic models an answer that reaches the window now ends with finish_reason length. On Together AI models Zabio now sends the room the window leaves beside the input, so the request stays on the model you named with a shorter output limit; when less than 4,096 tokens of room remain it is moved or refused as before. A request that sets max_tokens is unchanged. See Errors.
  • Call data export. Zabio can now send a record of each call your organization makes to your own observability tool over OpenTelemetry (OTLP): the model, tokens, cost, timing and outcome, and the request and answer when body logging is on for the key and you include bodies. Set it up in Settings, Call data export (owners and admins). It is off until you switch it on, and it never slows a call. See Call data export.
  • Longer audio on Gemini models. The audio you send in input_audio parts is now limited per model: Google’s Gemini models take up to 100 MB of decoded audio per request, and every other model 20 MB as before. A single request body carries about 25 MB of audio; more reaches a Gemini model on the Responses API, where a conversation’s stored clips count too. Zabio sends large requests’ clips to Google through Google’s file service and deletes them there as soon as the answer is in. A request above a model’s limit answers 413 media_too_large and now states the limit.
  • Speech as events. Text to speech now streams as events: send stream_format: sse to /v1/audio/speech and get the audio in base64 speech.audio.delta events, then a speech.audio.done event with the usage and the call’s cost, on every speech model.
  • Cached audio on two Gemini models. Audio that gemini-3-flash-preview or gemini-3.1-flash-lite reads from its cache is now billed at Google’s cache price for audio: 0.10 and 0.05 USD per million tokens at list, before your plan’s fee. Until now it was billed at those models’ cache price for text, 0.05 and 0.025. GET /v1/models shows the price as zabio.pricing.cache_read_audio. Every other model has one cache price and does not change.
  • Transcription estimates on trial credit. While an organization runs on trial credit alone, the amount set aside for a transcription uses the longer of two lengths: the one the file’s header states, and the file’s size as speech at 16 kbit/s. A file encoded above 16 kbit/s is estimated above its bill until the call ends, and a call whose estimate does not fit the balance is refused with 402 insufficient_credits. The bill is the real length, as before, and nothing changes for an organization that has added credit. See Audio.
  • The IP lists in the organization’s data download. The organization’s data, downloaded from the web app by an owner or an admin, now holds the IP allow and deny lists themselves: the organization’s, and those of each project and each key that has its own, revoked keys included. Before, it only said whether a project or a key had a list.
  • Clearer words and phone widths in the web app. The pages of the new menu settle their wording in English and Spanish, name the pages by their new names, and work at phone widths: nothing scrolls sideways but wide tables, every control shows where the keyboard focus is, and text meets the contrast the accessibility guidelines ask for. The emails about two-step sign-in and the monthly usage summary now point to My Profile. A call refused by an IP address list now says the address it came from and where the lists are set, under AI Firewall, Network Access.
  • A new tour of the web app. When you sign in, a window on the Dashboard offers a short tour in five steps: create an API key, make your first call, see it in Usage, protect it in AI Firewall and top up in Billing. Each step opens its page and points at the part it explains. Skip for now asks again the next time you sign in, and Don’t show this again does not ask again; your answer is kept with your account, on any browser. Take the tour, at the end of the menu, shows it again at any time. If you skipped or finished the earlier tour, the window does not open.
  • Background image generations and videos set their estimate aside. When a background image generation or a video is queued, its estimated cost is now set aside on your balance until it ends, as a direct call’s already was, so requests queued at the same time can no longer take the balance below zero together. A request whose estimate does not fit the balance left after the calls still in flight is refused with 402 insufficient_credits and nothing is queued; the message says the estimate, the balance and the least amount to add. The amount set aside is given back when the generation or the video fails or is cancelled, and replaced by the real charge when it is made. A video’s estimate no longer adds a quarter for safety: it is the length times the model’s highest price per second, fee included. See Images and Videos.
  • A new menu in the web app. The sidebar groups the pages under Overview, Build, Monitor, Security & Policies and Organization, and long pages are split into tabs, each with its own address. Your own settings (profile, password, two-step sign-in, sessions, access tokens and connected apps) moved from Settings to My Profile, in the menu under your initials. The AI firewall has its own page, Projects is a tab of API keys, Referrals a tab of Billing, and your account’s model access is under Models. The project switcher is now a filter: it shows all projects until you choose one, and the address keeps your choice. Old addresses and bookmarks still open the right page. What’s new, in the header, lists the latest changes and anything your account needs to act on.
  • Search in the web app. Press Ctrl K (⌘ K on a Mac), or click the search box in the header, to find any page, tab or setting, and your organization’s API keys, models and projects, by name. Use the arrow keys and Enter to open a result, and Esc to close; the pages you opened last are listed first.
  • IP allow and deny lists. An organization, a project and a key can each have a list of allowed addresses and a list of denied ones, IPv4 and IPv6, as single addresses or CIDR blocks. A deny list always wins, and a project or a key can only narrow what the level above allows. A call with a key from an address the lists refuse gets 403 ip_not_allowed: nothing reaches a model and nothing is billed. The lists are free on every plan, and with no list nothing changes. Owners and admins set them in the web app under AI Firewall, Network Access, which also shows your current address and asks before a change that would refuse it. See IP allow and deny lists.
  • A clearer Dashboard. The Dashboard opens with four figures: Spend, Saved by routing, Calls and Credits, which now says how many days your balance lasts at this month’s pace. Below them are the daily spend by model, the cost breakdown, your top models, the kind of work and your tokens; how each figure is computed is behind its info icon. A line above the figures shows one message at a time, the most important first: an action your account needs, trouble at a provider whose models you use, then what is new in the latest release. You can hide each message. Spend by key and by member moved to the Overview tab of Usage.
  • The limits of every key in one list. Under API keys, the Limits tab lists each active key’s spend limit with how much of it is used this period, its requests and tokens per minute and when it expires; Edit opens the key. The Keys tab sums them up in one line at the top: active keys, keys with a spend limit, and any key near or at its limit or about to expire. Models, API keys, Conversations, Batches and the Playground take the new look. A batch’s page opened from a link no longer shows an error.
  • Usage, Billing and Members in the new style. The Overview tab of Usage opens with four figures for the range and filters you chose: Spend with the cost per 1M tokens, Saved by routing, Calls with the share that succeeded and the latency, and Cache hits with what caching saved. The daily spend sits beside the cost breakdown. On Billing, your balance, what you added in total and your last top-up are one row of figures; while your launch match is open, Balance & Top up shows one line with where it stands, and the match itself is on Credits & Referrals. Members says how many members and pending invitations the organization has and your role, and shows each member’s spend this month against their monthly quota.
  • Image, speech, transcription and embeddings calls set their estimate aside. These calls now work like chat: before a call reaches the model, its estimated cost is set aside on your balance until the call ends, so calls sent at the same time can no longer take the balance below zero together. A call whose estimate does not fit the balance left after the calls still in flight is refused with 402 insufficient_credits; the message says the estimate, the balance and the least amount to add. Background image generations and videos are not changed yet. See Billing.
  • The AI firewall in Top keys. On the Usage page, Top keys now counts each key’s AI firewall charges in its spend and its share, as a key’s spend limit already did, so the table adds up to You paid. Top models stays the calls’ only, since the AI firewall is no model. Nothing is charged twice and no price changed.
  • The AI firewall on the Dashboard. You paid, Spend and the daily spend chart now include the AI firewall’s charges, as the Usage page and your bill already did. The chart stacks them on top of the calls’ line, and You paid shows the same figure on the Dashboard, on the Usage page and on the bill for the same period. Nothing is charged twice and no price changed.
  • The AI firewall in your usage. On the Usage page the Spend chart shows the AI firewall’s charges as a part of their own, so its total matches You paid. The monthly usage report names the AI firewall’s tokens checked and what it charged, and its bill includes them. The AI firewall’s charges count toward the 250 USD that unlocks the match. Nothing is charged twice and no price changed. See Credits, match and referrals.
  • A request too long for the model now gets a clear 400. A request that does not fit the context window of the model you named, or of any model auto can use, used to get 503 no_provider_available with Retry-After, so clients retried a request that could never succeed. It now gets 400 context_length_exceeded, with the request’s size, the output counted and the window in the message. A request whose input fits the model you named and that sets no max_tokens is now sent to the model instead of refused near the window, and one that does not fit is no longer moved to another model; on Anthropic, DeepSeek and Together models the output Zabio sends by default still counts, and a model of the same provider with a larger window still takes the request when one can. See Errors.
  • Easy classification jobs on auto go to the more accurate model. A request whose response_format is a JSON schema of fixed choices, booleans and bounded numbers was answered by Jev only when Jev was also the cheaper model, so most short requests went to the cheapest text model. In our tests Jev got every field right on 78 percent of such requests, against 51 percent for that model, so on an easy classification job Jev now answers at any input size. A short request can cost slightly more than before: at most Jev’s own price for the call, which stayed under 0.00004 USD on our test requests under 200 input tokens. The same rule applies to an auto line inside a batch, where Jev answers at its standard price at once, and POST /v1/public/analyze prices such a request the same way: its jev switch applies whenever Jev would answer, while its jev tip still appears only when Jev is the cheaper one. Other jobs keep the price comparison, and a request that names a model is unchanged. See Classification jobs on auto.
  • Cached audio on the Gemini models is billed once. When a Gemini model reads part of an input_audio clip from its cache, those tokens are billed at the cache price only: usage.prompt_tokens_details.cached_tokens counts them and audio_tokens is the audio that was not cached. No call on Zabio was billed the old way.
  • More Qwen models take video. Eleven Qwen 3.6 and 3.5 models on Alibaba Cloud (International) now take video_url parts when you call them by name. See Video.
  • zabio.cache.applied counts a cache write. On a provider that caches on its own, a call that only wrote the cache, such as the first call of a long prompt on OpenAI GPT-5.6 and later or on Moonshot, now reads vendor_automatic instead of none. read_tokens still says whether the call hit the cache, and write_tokens what it wrote. Display only: the bill did not change. See the per-provider table.
  • Image requests on trial credit. An image generation or edit whose estimated cost is above what is left of your trial is now refused with 402 insufficient_credits before it reaches the model, as a chat request already was; the message says the estimate and what is left. A background generation is checked when it is queued and again when it runs. Accounts that have added credits are not affected. See Images.
  • The prompt-injection guard reads more. It now also checks the override, extraction, persona and bypass phrases in Spanish, Portuguese, French, German, Italian, Russian and Chinese (detection is measured in English); key words broken by spaces (ign ore prev ious); text in hex with separators, percent-encoding, HTML character references, \u escapes, decimal or binary character codes, ROT13 or written backwards, and one encoding inside another; tool results that close their own wrapper and go on with orders, fake user or assistant turns, fake agent steps and fake tool calls; instructions hidden in HTML comments or in invisible text; and HTML images whose link carries data out. detections[].encoding takes the new encodings’ names. “Your new role is team lead” in a tool result is no longer reported. See Prompt-injection guard.
  • Each detection in zabio.guard now names its detector: pattern for the free patterns. See Prompt-injection guard.
  • Report what the prompt-injection guard flags. On the Usage page’s Logs tab a Guard filter shows the calls the guard flagged or could not fully read, and a flagged call’s detail has Not an injection (that exact text is no longer flagged in your organization) and Confirm attack, with the flagged text shared only if you choose. Settings gains a section for the guard: each key’s action, the texts your organization allowed and a weekly summary. A key can trust tools by name (guard_trusted_tools): what the patterns find in their results is reported with trusted: true and never redacted or blocked. The MCP feedback tool takes an injection field; such a report waits for an owner or admin in Settings. See Prompt-injection guard.
  • Webhooks when a batch ends. Add an https address in Settings, Webhooks, and Zabio sends it a signed POST when one of your batches completes, fails, expires or is cancelled (batch.completed, batch.failed, batch.expired, batch.cancelled), with the batch id, its status, the line counts and links to its files, never a prompt or an answer. The request follows Standard Webhooks (webhook-id, webhook-timestamp, webhook-signature), so its libraries verify it; the signing secret is shown once and can be rotated with a 24-hour overlap. A delivery that fails is tried 8 times over about 31 and a half hours, and Deliveries shows each attempt. See Webhooks.
  • Claude Code through Zabio. Claude Code sessions no longer stop with a 400 on their second request (messages.1.role) or when an advisor model is set. Messages with the system role between the turns are accepted, as Anthropic accepts them. Server tools, Claude Code’s advisor among them, are removed from the request and listed in zabio.ignored_parameters instead of refusing the call; the advisor is not available through Zabio. See Anthropic SDK and Claude Code.
  • Tools on Gemini models. Tools whose parameters carry JSON Schema keywords that Google’s Gemini models do not take ($schema, additionalProperties, $ref and $defs, exclusiveMinimum and others, as Claude Code and the OpenAI and Anthropic SDKs send them) failed on every Gemini model, named or picked by auto. Zabio now converts each schema for Gemini, lists the keywords that carried a rule and were left out in zabio.ignored_parameters, and keeps auto off the Gemini models for a schema that refers to itself; a named Gemini model answers such a request with 400 unsupported_parameter naming the tool. See Tools on Gemini models.
  • Calls that write OpenAI’s prompt cache are now billed at OpenAI’s cache write price, as the price list shows: 1.25 times the input price on gpt-6.1-sol, gpt-6.1-terra, gpt-6-sol, gpt-6-luna, gpt-6-astra, gpt-5.6-sol, gpt-5.6-terra and gpt-5.6-luna, at the batch and flex rates on those tiers. Before, the written tokens were billed at the input price, below what OpenAI charges, and zabio.cache.write_tokens read 0; it now shows them, and prompt_tokens is unchanged. A call on those models that writes the cache costs slightly more from this release on; calls that only read it, and every other model, are unchanged.
  • A background response (background: true on POST /v1/responses) that sets zabio.cache.ttl or cache.breakpoints to none on a vendor that caches on its own now lists the field in zabio.ignored_parameters, as the same request sent without background does. The answer and the bill were already right; only the report was missing.
  • A free prompt-injection guard on every key. Chat completions, the Responses API and /v1/messages now check the user’s messages, tool results and the assistant turns you send for text that tries to override the model’s instructions (“ignore all previous instructions”, a request for the system prompt, fake chat template markers, orders hidden in a fetched page), after undoing the usual disguises: invisible characters, lookalike letters, spaced-out or misspelled words, Base64 and hex. Each key has an action: tag (the default for every key: the call goes through and the answer’s zabio.guard lists where each detection was, never the text), redact (the span is replaced with [PROMPT_INJECTION] before the model sees it), block (403 prompt_injection_blocked, nothing billed) or off. Set it with guard_action on the key; a request can ask for a stronger action with zabio.guard.action or x-zabio-guard, never a weaker one. A phrase your organization knows is harmless can be added to its allowlist. It is pattern based and not a guarantee. See Prompt-injection guard.
  • Anthropic’s Messages API and Claude Code. The Anthropic SDKs, Claude Code and any tool that takes an Anthropic base URL work with your Zabio key: set the base URL to the API host without /v1 and send your key as x-api-key or a bearer token. POST /v1/messages takes Anthropic’s request and answers with Anthropic’s message or stream; POST /v1/messages/count_tokens counts tokens; GET /v1/models with an anthropic-version header lists the models in Anthropic’s shape; errors on these paths use Anthropic’s envelope. A Claude model named in model gets your body as you sent it (system blocks, cache_control, thinking with its signatures, documents), and auto and every other model get the request translated. Calls are billed as the same calls on chat completions. Server tools, container, mcp_servers, fast mode and data residency are refused. See Anthropic SDK and Claude Code.
  • Two-step sign-in. Turn it on in Settings: after your password, email link, Google or GitHub, Zabio also asks for a code from an authenticator app, with ten single-use recovery codes for a lost phone. While it is on, creating an API key or a personal token, approving the CLI, connecting an app, moving credit and deleting your account ask for a code if you have not entered one in the last 10 minutes; zabio keys create asks for it in the terminal or takes --code. It is off unless you turn it on. See Settings.
  • Passkeys for two-step sign-in. In Settings, add a passkey (your fingerprint, face or screen lock, a password manager or a security key) as your second step, beside or instead of the authenticator app. Signing in and the actions that ask for a second step then offer Use a passkey. A passkey cannot be phished: it works only on app.zabio.ai. The first second step you add turns two-step on and gives you the ten recovery codes; zabio keys create keeps asking for a code. See Settings.
  • The web app stays reachable while a new release starts: a page or a form sent at that moment waits a second and is answered, instead of failing.
  • An opaque id for your end users at the vendors. On Zabio’s vendor keys, every chat completion, Responses and image call now carries an opaque identifier in the field the vendor documents for one (OpenAI’s safety_identifier, Anthropic’s metadata.user_id, Google’s labels.safety_identifier, DeepSeek’s user_id, DeepInfra’s user, Moonshot’s safety_identifier), derived from your organization and from your own end user’s id when you send safety_identifier or user; never your organization’s id or your value. A vendor can then act on one end user rather than on the account all customers share. Chat completions and the Responses API now read safety_identifier, and the Responses API reads user instead of listing it in zabio.ignored_parameters. On your own vendor key, the value you sent goes to the vendor as it is. See What the vendor receives.
  • A batch line that sets zabio.cache.ttl or cache.breakpoints to none on a provider that caches on its own (OpenAI, Google and the others) lists the field in its output line’s zabio.ignored_parameters, as the same request sent synchronously does. See Batches.
  • Grok 4.7 and Grok 4.6 in batches. xAI now prices both models for its Batch API at 20 percent off every token type, so a batch line can name grok-4.7 or grok-4.6 and is billed at the batch price (1.76 USD per million input tokens, 0.44 cached and 5.28 output on Zabio’s US endpoint), and auto inside a batch can pick them. Before, such a line failed the batch with model_not_batchable. See Batches.
  • API keys show when they were last used. The Last used column of the API keys page shows the time of the key’s latest request, to the minute. Before, every key read Never, however many calls it made.
  • Usage: the cost breakdown for any range. The Stats tab of Usage shows how the spend adds up (provider price, Zabio fee, automatic model choice, sharing discount, classifier and what you paid) for the range and filters in the filter bar, a past month included, so a link with range=custom&from=...&to=... opens it.
  • A retired model answers 410. A request that names a model past its retirement date gets 410 model_retired instead of 404 model_not_found, with the same message naming the date, on every endpoint that takes a model and in the lines of a batch (a line whose model retires after the batch was created ends in the error file with model_retired and status code 410). Nothing reaches the vendor and nothing is billed. See Errors.
  • A retired model names its successor. When the vendor names a model that replaces one, GET /v1/models shows it as zabio.successor and the 410 model_retired message names it, so you know which model to call instead, before the date or after it. See Models.
  • Logs: filter by provider, model and key, and see each call’s key. The Logs tab narrows the calls by provider, by model (the model you asked for or the one that answered) and by key, and keeps the filters in the page address so a filtered view can be shared. Every row shows the name of its key and its project. GET /admin/v1/orgs/{org}/calls returns key_name and project_name on each call, and its model filter matches requested_model or served_model.
  • Responses lines in batches. POST /v1/batches takes endpoint: "/v1/responses", as OpenAI’s Batch API does: every line of the file is a POST /v1/responses request, routed, priced at the batch rates and billed exactly as a chat completion line with the same tokens, and each output line carries a Response object. A batch line has no stored conversation, so previous_response_id, conversation, store: true and background: true fail the batch with unsupported_parameter, and the Response object is not stored and does not repeat the request. A batch is all chat completion lines or all Responses lines. gpt-6-astra now takes tools in a batch and in a background response, on either endpoint: Zabio sends those lines to OpenAI’s batch for the Responses API. See Batches.
  • Batch chat lines list their ignored parameters. A chat completion line of a batch that sends seed, logit_bias, presence_penalty or frequency_penalty lists them in its output line’s zabio.ignored_parameters, as the same request sent synchronously does. Before, only Responses lines listed them. See Batches.
  • Batches in the web app. A Batches page lists your batches with their status, lines, cost and the vendors’ discount (what the batch prices took off the standard prices), filtered by status and by date. A batch’s page shows its lines one by one with their errors, the cost charged, the saving against your baselines, Cancel batch while it runs, and the output and errors files for owners and admins. The admin API gains GET /admin/v1/orgs/{org}/batches, GET .../batches/{id}, POST .../batches/{id}/cancel and GET .../batches/{id}/files/{output|errors}. See Batches.
  • Audio answers from chat models. POST /v1/chat/completions answers in audio when the request sends modalities: ["text", "audio"] and audio: {"voice", "format"}, in OpenAI’s own shape: the answer’s message.audio carries the audio as base64, its transcript, an id and expires_at, and a stream carries delta.audio chunks. The models are OpenAI’s gpt-audio-1.5, gpt-audio and gpt-audio-mini, called by name, in wav, mp3, flac, opus, pcm16 or aac (a stream in pcm16) and thirteen voices; zabio.capabilities.audio_output marks them in GET /v1/models. The audio is billed at the model’s audio output price (64.00 USD per million audio tokens on gpt-audio-1.5), the text at its output price. Zabio never stores the audio. A later turn can send the assistant’s audio: {"id"} back to the same model. auto and the Responses API answer in text; see Chat completions.
  • Every ElevenLabs and Deepgram voice and transcription model. Seven speech models join POST /v1/audio/speech: from ElevenLabs eleven-v3 (80.00 USD per million characters, 5,000 characters a request), eleven-v3-conversational (40.00), eleven-multilingual-v2 (80.00, 29 languages), eleven-flash-v2.5 (40.00, 32 languages) and eleven-flash-v2 (40.00, English only), all with the same 19 voices as eleven-v4; from Deepgram aura-1 (15.00, English only, 12 voices) and flux-tts (45.00, 266 voices in English and 41 more languages, speed from 0.5 to 1.5). Two transcription models join POST /v1/audio/transcriptions: scribe-v2-medical (ElevenLabs, for clinical audio, 0.22 USD per hour as scribe-v2) and whisper-large (Deepgram, 0.0048 USD per minute). auto on speech never picks a model that speaks English only, since a speech request does not say its language; name one to use it. The models ElevenLabs and Deepgram have deprecated, serve only in real time or do not price are not added; see Audio.
  • Gemini Omni Flash videos. gemini-omni-1.1-flash (Google) joins POST /v1/videos: 3 to 10 seconds from a prompt, an image, or a first and a last frame, at 360p, 720p, 1080p or 4k, always with sound, billed on the tokens Google reports (17.50 USD per million video tokens, about 0.10 USD a second at 720p). seconds sets the length as on the other models. It does not make videos of certain recognizable people: such a request fails with moderation_blocked and says so. See Videos.
  • Video generation. POST /v1/videos makes a video in the shape OpenAI’s SDKs knew for Sora (client.videos.create, as JSON or the multipart form the SDKs send): a prompt, a length in seconds, a size or resolution, and optionally an image to start from (input_reference). The answer comes at once with a video_ id; GET /v1/videos/{id} says when the video is ready and then carries a link that works without your key for 60 minutes, GET /v1/videos/{id}/content downloads the MP4, DELETE cancels a queued video or deletes an ended one, and GET /v1/videos lists yours. The first models: xAI’s grok-imagine-video-1.5 (0.08 USD a second, 1 to 15 seconds, up to 1080p, a first and a last frame, sound that can be turned off) and grok-imagine-video-1.5-lite (0.02 USD a second), both 0.01 USD per image sent,, made on xAI’s global endpoint, and Alibaba’s wan2.6-t2v and wan2.6-i2v (from a prompt or from an image, 2 to 15 seconds, 0.10 USD a second at 720p and 0.15 at 1080p, with sound), made in the United States. model: auto picks the cheapest model that makes the length and resolution you ask for. A video is billed per second of the video made, at the price of the resolution made, plus the plan fee; credit is checked against the video’s estimate when it is queued and again when it starts. The prompt and the images go to the vendor that makes the video; the video is stored encrypted under your organization’s retention (or sooner with zabio.expires_after) until you delete it. See Videos.
  • Video in chat. user messages on POST /v1/chat/completions take video_url parts and POST /v1/responses takes input_video parts: a video as a data URL, an https URL, a public YouTube URL, or the file_id of a file you uploaded with purpose user_data; MP4, MOV or WebM, up to 10 videos and 100 MB per request. A request with video is routed only to models that take it (zabio.capabilities.video_input in GET /v1/models): Google’s gemini-3.8-flash, gemini-3.6-flash, gemini-3.5-flash-lite, gemini-3-flash-preview and gemini-3.1-flash-lite, and Alibaba’s qwen3.8-omni-flash, qwen3.8-flash and qwen3.7-plus by name. YouTube URLs and WebM go to the Gemini models only. auto picks among the Gemini models; a named model that takes no video answers 400 before anything is sent. A video is billed as input tokens at the model’s input price, its sound at the model’s audio price where the vendor counts it apart. Zabio never keeps a video: a stored Responses turn keeps its format, size and length only, so a later turn does not see it again, and a video cannot go in a batch or a background Response. See Chat completions.
  • Images as URLs. "response_format": "url" on POST /v1/images/generations (direct or in the background) and POST /v1/images/edits returns each image as a link in data[].url instead of base64, the way OpenAI’s DALL-E models answered. The link works for 60 minutes without your API key, so it can go straight into an <img> tag; anyone holding it can open the image until then, so share it as you would the image. The images are stored for those 60 minutes, encrypted with your organization’s key, and deleted after. b64_json stays the default. See Images.
  • xAI’s Grok Imagine models. grok-imagine-image (0.02 USD per image, 0.002 per input image on an edit) and grok-imagine-image-2.0 (0.04 USD per image at quality low, 0.06 at medium, 0.01 per input image on an edit) generate and edit images: one input image on 1.0 and up to five on 2.0, no mask, sizes 1024x1024, 1248x832 and 832x1248 (the pixels xAI makes; 1536x1024 and 1024x1536 are not xAI sizes). xAI serves images only on its global endpoint, so these two models show their processing location as not guaranteed; xAI’s chat models stay in the United States. auto now picks grok-imagine-image at 1248x832 and 832x1248 and for a one-image edit without a mask. xAI charges for an image its moderation refuses, so a refused request on these models is billed what xAI charged plus the plan fee (400 moderation_blocked, the message says how much). GET /v1/models gives the input-image price as zabio.pricing.per_input_image. See Images.
  • Image edits, a FLUX model and background generation. POST /v1/images/edits takes OpenAI’s multipart form: one or more image[] files (png, jpeg or webp, 50 MB each at most) and an optional mask. The OpenAI image models take up to 16 images and a mask, gemini-nano-banana-2.1 and gemini-3-pro-image up to 14 images, gemini-3.1-flash-lite-image one; auto picks the cheapest model that takes what you send. The input images are never stored. flux-2-dev@deepinfra, FLUX.2 [dev] by Black Forest Labs served by DeepInfra, generates images billed per image (0.017858 USD at 1024x1024). "zabio": {"background": true} on a generation answers 202 with an img_ id; fetch the images with GET /v1/images/generations/{id} and delete them with DELETE. A background generation stores the prompt until it ends and the images under your organization’s retention, encrypted. See Images.
  • Image follow-ups. GET /v1/images/generations lists your background generations, newest first, with limit, after and status. "zabio": {"expires_after": {"anchor": "created_at", "seconds": N}} on a background generation deletes its images sooner than your organization’s retention. Credit is checked again when a background generation runs: one whose balance reached zero while it waited ends failed with insufficient_credits and bills nothing. POST /v1/images/edits also takes OpenAI’s JSON shape, the images as base64 data URLs or file ids; web addresses are not fetched. The CLI gains zabio image edit, zabio image --background, zabio image status and zabio image get. See Images.
  • Image generation. POST /v1/images/generations takes a prompt in OpenAI’s shape and returns the images as base64: OpenAI’s gpt-image-2.5-flare, gpt-image-2.5-sunburst and gpt-image-2 (any size within OpenAI’s limits, quality low to max, png, jpeg or webp, transparent backgrounds, partial images with stream), gpt-image-1-mini by name until its shutdown on 2026-12-01, and Google’s gemini-nano-banana-2.1, gemini-3.1-flash-lite-image and gemini-3-pro-image (jpeg, up to 4K, 14 aspect ratios). model: auto picks the cheapest model that produces the asked size, quality and format. Every call is billed on the tokens the vendor reports; the images are never stored. A long generation is kept alive on the connection (the 200 is committed after 90 seconds and a space written every 15 until the body arrives). The CLI gains zabio image PROMPT -o FILE. Edits and background generation are not available yet. See Images.
  • Audio in chat. user messages on POST /v1/chat/completions and POST /v1/responses take input_audio parts in OpenAI’s shape: the clip as base64, wav or mp3, up to 20 MB per request. A request with audio is routed only to models that take it (zabio.capabilities.audio_input in GET /v1/models): Google’s gemini-3.8-flash, gemini-3.6-flash, gemini-3.5-flash-lite, gemini-3-flash-preview and gemini-3.1-flash-lite, and OpenAI’s gpt-audio-1.5, gpt-audio and gpt-audio-mini by name (the last two until their shutdown on 2027-01-20). auto picks among the Gemini models; a named model that takes no audio answers 400 before anything is sent. The clip’s audio tokens are billed at the model’s audio price and reported in usage.prompt_tokens_details.audio_tokens. On chat completions the clip is never stored; a stored Responses turn keeps it under the conversation’s retention. See Chat completions.
  • Text to speech. POST /v1/audio/speech takes text in OpenAI’s shape and returns the audio, streamed as the model produces it, as mp3, opus, aac, flac, wav or pcm. model: auto picks the cheapest speech model that produces the format you ask for and has the voice you name; the models are Hexgrad’s Kokoro (served by Together AI), Deepgram’s Aura-2, ElevenLabs’ v4 and v4 Turbo, Google’s Gemini 3.8 Flash TTS and Flash-Lite TTS, and OpenAI’s gpt-4o-mini-tts, tts-1 and tts-1-hd by name until their shutdown on 2027-01-06. Nothing is transcoded: each model returns the formats its vendor produces, listed with its voices in GET /v1/models. Billed per character of input text, or in text and audio tokens where the vendor bills tokens; the call id, the model, the vendor, the cost and the units travel as x-zabio-* response headers, since the body is audio. The audio is never stored. The CLI gains zabio speak -o FILE TEXT. See Audio.
  • Audio transcription. POST /v1/audio/transcriptions takes an audio file in OpenAI’s shape and returns its text, as json, text, verbose_json, srt or vtt, streamed or whole. model: auto picks the cheapest transcription model able to do what the request asks; the models are OpenAI’s gpt-transcribe (and its retiring whisper-1 and gpt-4o-transcribe family until 2027-02-26), Google’s gemini-3.5-transcribe, ElevenLabs’ Scribe v2, Deepgram’s Nova-3, Whisper Large v3 on Together AI and Mistral’s Voxtral Mini and Small on DeepInfra, billed per minute of audio (or in audio tokens where the vendor bills tokens) at the provider’s list price plus the plan fee. ElevenLabs and Deepgram are new vendors; see the privacy page. The audio is never stored. Usage, the CSV export and the model list show minutes of audio and each model’s unit. See Audio.
  • A model a vendor stops serving is retired on its own. When a vendor’s model list has not shown a model for a day and a test call to it is refused, Zabio sets its retires_at seven days ahead: auto stops choosing it at once, requests that name it keep working until the date and carry the new x-zabio-retires-at header, and after the date they get a 404 model_not_found that says when the model was retired. If the vendor lists the model again before the date, the date is removed. Models retired earlier by hand follow the same rule from their date. See Models and the API reference.
  • More from ask on the MCP server, two new tools and three prompts. ask takes instructions (a system message), schema (a JSON answer in that shape), images (up to 4 https URLs), deadline (from 15 minutes the call may use the vendor’s flex tier, in apps that show progress) and thread (a stored conversation, started with new and continued with its conv_ id). Its closing line names the model, the vendor and the cost. estimate prices a task without running it, as POST /v1/public/analyze does, and budget shows what remains of the connection’s daily limit and the organization’s balance. The prompts ask, estimate and explain appear as slash commands in the apps that show them. See Connect an AI app
  • Parasail is a host. Sixteen open-weight models Zabio already serves have a copy on Parasail with an id ending in @parasail (DeepSeek V4 and V4.1, Qwen 3, 3.5, 3.6 and 3.8, Llama 3.3 70B, GLM 5.3 and MiniMax M3), each at the precision Parasail states (fp8, or the maker’s full precision for qwen3.5-9b@parasail) and billed at Parasail’s list price plus 5 percent, with cache hits at Parasail’s cached rate where it has one. The copies of models auto already uses (qwen3.5-9b, llama-3.3-70b-instruct-turbo, deepseek-flash, deepseek-v4-pro) join auto with their originals, so a call goes to Parasail when it is the cheapest host or when another host is down. The vendor is on by default and can be turned off like any other; you can bring your own Parasail key. See Models and the privacy page.
  • DeepInfra is a host. Sixteen open-weight models Zabio already serves have a copy on DeepInfra with an id ending in @deepinfra (DeepSeek V4 and V4.1, Qwen 3, 3.5 and 3.6, Llama 3.3 70B, gpt-oss-120b and MiniMax M3), each at the precision DeepInfra states (fp8, or the maker’s full precision for qwen3.5-9b@deepinfra and gpt-oss-120b@deepinfra) and billed at DeepInfra’s list price plus 5 percent, with cache hits at DeepInfra’s cached rate where it has one. The copies of models auto already uses (llama-3.3-70b-instruct-turbo, qwen3.5-9b, deepseek-flash, gpt-oss-120b, deepseek-v4-pro) join auto with their originals, so a call goes to DeepInfra when it is the cheapest host; the rest are called by name. The vendor is on by default and can be turned off on the Models page, and you can bring your own DeepInfra token. See Models and Privacy.
  • Moonshot AI is a vendor. Kimi K3.1 Pro (kimi-k3.1-pro), its high-speed variant (kimi-k3.1-pro-highspeed) and a second host of Kimi K3 (kimi-k3@moonshot) are served through Moonshot AI’s international API at Moonshot’s list price plus 5 percent, with cache hits billed at Moonshot’s cached rate. The two K3.1 Pro models are in auto/best. The vendor is on by default and can be turned off like any other; you can bring your own Moonshot key. See Models and the privacy page.
  • Model labels name the maker and the processing region. The eleven models whose label named the host, for example “DeepSeek (North America, via Together AI)” and “Qwen3.7 Max (US, via Alibaba Cloud)”, now show the maker’s model name and where the call is processed: “DeepSeek V4.1 Flash, processed in North America”, “Qwen3.7 Max, processed in the US”. The label is zabio.label in GET /v1/models and the line under the id in the model tables. Model ids, prices, zabio.vendor and the x-zabio-vendor header do not change; the privacy policy and the data processing addendum quote the new labels.
  • Models show their maker. owned_by in GET /v1/models, the new maker field of GET /v1/public/prices and the Maker column of the model tables name the company that made each model, for example moonshot (Moonshot AI) for kimi-k3 and zai (Z.ai) for the GLM models. Model ids and prices do not change.
  • Off-peak prices on calls that cross the boundary. A call on a model with an off-peak price takes it only when both its arrival and its completion fall inside the off-peak hours; a call that crosses the boundary in either direction is billed at the list price. Where a vendor states which instant it charges by, Zabio follows it, and GET /v1/public/prices says so in off_peak.charged_by. See Billing.
  • Current prices as data. GET /v1/public/prices lists the provider list price of every model, with as_of, when the prices last changed, and where each price was read. It needs no API key and answers 304 when the list has not changed since your copy. See Billing.
  • Workload report. Each month your organization shares calls, an independent model judges a sample of the answers blind (it sees the de-identified request and the answer, never which model produced it) and, for one in ten, asks a cheaper and a stronger model for their own answer to the same request. On the 2nd the owners receive a report by email, also under Usage, Workload report: the share of correct answers by type of work, what the cheaper replays said and the month’s sharing discount. A report needs at least 50 judged answers in the month. Nothing on your account changes from it.
  • Which vendors see a shared call. The privacy page now names them: a shared call is sent only to vendors whose published terms say they do not train on API inputs and outputs. DeepSeek receives none.
  • POST /v1/feedback is rate limited by the key’s own request limit, like every other call.
  • Tell Zabio when an answer was wrong. POST /v1/feedback takes {call_id, verdict, reason} under your API key, and the same object travels as zabio.feedback inside your next chat completions or Responses request. The Logs tab has thumbs up and down on every row. A bad verdict makes Zabio choose the model again on the next turn of that call’s session, as the MCP server’s feedback tool already did.
  • Lost calls are named. A call the server did not finish recording (it stopped between accepting the call and writing its row) now appears in Logs and the CSV as Lost, not billed, instead of disappearing.
  • Parallel tool calls on every model. An assistant message carrying several tool_calls, followed by one tool message per call (what an agent loop with parallel tool calls sends on its second turn), now reaches every provider as one assistant turn. Before, the OpenAI-compatible providers (DeepSeek, Together AI among them) rejected the second turn with “an assistant message with tool_calls must be followed by tool messages”. Stored conversations on the Responses API read back the same way.
  • Images on the Responses API. An input_image with a URL is fetched and inlined before anything is stored, as on chat completions, so every provider gets the same bytes under the same size limit, and a data URI is decoded rather than passed to the provider as a URL; later turns of the conversation never fetch the image again.
  • Running out of credit. When your balance reaches zero, requests now stop by default: a request is refused with 402 insufficient_credits until credit is added, and so is a request whose estimated cost does not fit the balance left after the calls in flight; the message names the estimate, the balance and the least amount to add. A balance below zero refuses every request until a top-up brings it above zero. Continuing on the free tier at zero is now a setting an owner turns on under Organization settings; organizations that were on the previous default have been moved and their owners emailed, and an organization whose owner had chosen to continue keeps the choice.
  • Provider errors in streams. A provider failure in the middle of a stream now ends it with the same error envelope a failed call gets (type, a code such as upstream_error, Zabio’s message), never the provider’s own text or Zabio’s internal error class. A provider’s refusal of the request itself (a 400) keeps its reason, with URLs, identifiers and addresses removed, and names the provider and the model in zabio.
  • Forced tool choice on Alibaba Cloud’s models. A request with tool_choice: required or a named function on a Qwen or GLM model with a reasoning level runs with thinking off, zabio.reasoning.applied says off and the level is listed in zabio.ignored_parameters with the reason tool_choice; Alibaba Cloud does not take a forced tool choice in thinking mode, and the call used to fail with a 400.
  • A disconnect is billed only for work the provider did. Leaving before the provider accepted the request (during Zabio’s own steps, or while a provider’s refusal was being retried) bills nothing; leaving while an answer was being written still bills the work done, as before.
  • Automatic model choice after a provider failure. When the provider chosen for an auto request fails, the next choice is the next candidate of the same selection; only when none is left does the earlier rule choose, and zabio.routing.fallback says so.
  • A reservation per call. Each paid call now sets aside its estimated cost (input, expected output and the fee, on the dearest model auto may pick) against your available balance when it is admitted, and settles to what it actually cost when it ends. A call that does not fit the balance left after the calls in flight is treated as a call at zero balance: served on the Free limits when your organization continues at zero, refused with 402 insufficient_credits when it is set to stop, the message naming the estimate and the balance in USD. Before, every concurrent call was admitted while the balance read above zero and the overshoot could exceed it. The response header x-zabio-balance-micros keeps reading the balance net of the calls in flight.
  • The Free plan follows the person. The Free allowance and its daily limits now count the organizations and the calls of the person who owns them (plus-tag and Gmail-dot aliases included), and an organization closed in the last 30 days still counts, so closing and recreating an organization does not start a new allowance.
  • Logs show the parameters a call ignored. A call’s detail on the Usage page’s Logs tab lists each parameter the request sent that Zabio accepted and could not apply, with the reason (not supported, the provider caches on its own, reasoning off with tools on this path, the model cannot reason).
  • The output estimate prices what is sent. On the Responses API, where OpenAI’s gpt-6-luna and gpt-6-sol keep their reasoning with tools, the cost estimate that routes and reserves now prices that reasoning instead of the chat completions rule’s off.
  • Shared calls. Your organization now shares a percentage of its calls with Zabio, de-identified, to improve how Zabio routes requests: 100 percent by default during the beta, set by the owner from 0 to 100 in Organization settings, with a share of its own per project and per API key. A call sent with store: false, the header X-Zabio-Share: no or zabio.share: false is never shared. The owner can delete what the organization shared from Settings. What is kept, for how long and who reads it is on the privacy page.
  • Sharing discount. Sharing earns up to 5 percent off Zabio’s fee: 0.05 percent of the fee per 1 percent shared, on the service fee and on the per call fee. The response’s zabio.cost gains share_discount_micros, usage totals and the CSV a Sharing discount line, and a call’s detail in Logs says whether it counted as shared. See Billing.
  • Reasoning with tools on OpenAI’s gpt-6-luna and gpt-6-sol. On the Responses API, where Zabio sends OpenAI calls, these two models now keep their reasoning when the request carries tools; before, Zabio turned it off on every call with tools, a rule OpenAI states for chat completions only. On the calls that stay on chat completions (a stop, your own OpenAI key, max_tokens under 16) reasoning is still off with tools, zabio.reasoning.applied says off and a level you sent is listed in zabio.ignored_parameters.
  • Thinking with tools on Together AI and Alibaba Cloud. A reasoning_effort you send to Together AI’s thinking models (qwen3.5-9b, kimi-k3, glm-5.2, the DeepSeek copies) and to Alibaba Cloud’s Qwen and GLM models is now kept when the request carries tools; Zabio used to turn thinking off there on every call with tools. Measured on each model with a live tool round trip. DeepSeek’s own API still runs with thinking off on tool requests.
  • The response says who cached. zabio.cache.applied on every chat completion and Responses answer, streamed or not: none (nothing was cached), zabio (Zabio’s cache markers on Anthropic wrote or read the cache) or vendor_automatic (the provider cached a repeated prefix on its own). OpenAI, Google and the other providers that cache automatically offer no way to decline it, so a request that sets zabio.cache.ttl or cache.breakpoints to none on them is accepted and the field is listed in zabio.ignored_parameters; on Anthropic none sends no marker, as before. The docs carry one table per provider saying who decides and what none does.
  • OpenAI’s models answer through OpenAI’s newer API. A long think on an OpenAI model is now watched like any other provider’s, and a call that is not streamed gets its 200 as soon as OpenAI accepts it. A call that sets stop, or runs on your own OpenAI key, is sent as before. Calls are billed at what OpenAI meters on that API, which differs slightly from before: about one more output token per call, and fewer input tokens on a call with tools or a JSON schema.
  • A call that fails is never billed. A streamed call that failed after it started was billed for its input; now no failed call is billed, also when the model provider had already started on it. Zabio pays the provider for that work. A call you cancel by closing the connection is billed for the work done, without the per-call fee. See Billing.
  • Long answers no longer time out. Zabio now receives every answer from the model provider as it is written and keeps your connection open until it is done, so a long answer is no longer cut at a fixed time: an attempt ends only when the provider stops sending data for 2 minutes, or after 30 minutes in all. A call that is not streamed gets its 200 status as soon as a provider accepts it (after 90 seconds at the latest, and at 90 seconds on OpenAI’s models that reason), then a space every 15 seconds until the answer; if it fails after that, the body is only an error object with the request id, so check for error before you read choices. A stream sends : ZABIO PROCESSING comments while nothing else is sent, and a stream that fails ends with the error event and a last chunk whose finish_reason is "error". Until any part of the answer has reached you, a failed attempt is tried again on another model. If you disconnect, the call is cancelled and you pay for the work done: the input and the output generated before you disconnected, streamed or not. See Errors.
  • Streamed calls bill exactly what unstreamed calls bill. On some models a streamed call counted cached tokens differently from the same call unstreamed; both now count the same tokens and bill the same amount.
  • No default output limit. A request that sets no max_tokens (or no max_output_tokens on the Responses API) now gets no output limit from Zabio, as on OpenAI. Before, chat completions and batch lines were limited to 16,000 tokens and Responses turns to 4,096. Providers that need a value (Anthropic, DeepSeek, Together AI) are sent the model’s maximum output. Set max_tokens whenever you want a limit. A Responses object now shows max_output_tokens as null when the turn set none. The balance checks before a call, the hold of a batch and the per-minute token limit count the output expected for the job, so no call that worked before is refused. When your balance cannot pay for the model’s full output, the call is sent the limit your balance can still pay for instead of being refused. The Free plan keeps its 2,000-token output limit, now also on Free batch lines. See Billing.
  • Plain auto means Zabio picks for you. The pages now describe plain auto (or a request with no model) as what it is: Zabio picks the cheapest model able to do the job, moving up a line when the job needs it. Routing is unchanged. On a plain auto call, Logs, the calls API (line, also in the CSV export), the MCP tools recent_calls and get_call, and escalated_from now show the line Zabio actually applied (cheap, balanced or best) instead of always balanced, past calls included; zabio.routing.line still shows auto. In GET /v1/models, plain auto lists with tier auto instead of balanced.
  • The analyzer states costs only. In POST /v1/public/analyze, each line (lines[]) and each combination of the switches (scenarios) gives its cost only, beside today; neither carries saving_usd any more. The playground on zabio.ai shows each model’s cost and, at the top, the request as sent and on auto side by side.
  • The analyzer’s tips show two costs, not a saving. In POST /v1/public/analyze, a tip no longer carries saving_usd: its params give the call as it is and with the tip (the line tip adds reference_model and reference_usd, the cache tip call_usd and cached_call_usd), so the difference shows without a saving claim. The playground’s tips say the same.
  • Savings are now computed when you look at them. What routing saved is worked out when you open the Dashboard or Usage, run zabio usage or use the MCP usage tool, and is never stored per call. For each key, Zabio takes its baseline model by the 95% rule over the key’s last 30 days at that moment, prices each auto answer of the period on that model at the price it had when the call ran, and subtracts what you paid, fees included. A key never shows less than zero, and a period with no saving shows zero. The new route GET /admin/v1/orgs/{org}/usage/savings returns the saving of a period, key by key. See Billing.
  • Calls show their cost only. zabio.cost no longer carries direct_micros, saved_micros, baseline or saving_note on chat completions, responses and batch lines, and the batch object no longer carries direct_micros or saved_micros. Logs, the call detail, the CSV export and the MCP tools recent_calls, get_call and ask show what each call cost, without a saving. The usage totals and the CSV export drop direct_cost_micros, direct_cost_usd, saved_micros and saved_usd, the CSV also drops baseline_pending, and a key no longer has baseline_model. Per call, the prompt cache shows as cached tokens and a split as the size of its excerpt: zabio.split no longer carries saved_micros or estimate, and the MCP tools no longer carry saved_usd or difference_usd. What caching saved over a period stays in Usage. See API reference.
  • High-stakes requests on plain auto get the strongest third. When a plain auto request is high stakes, Zabio now picks among the models in the top third of skill for the work, as auto/best does, instead of the top two thirds.
  • Hard and high-stakes requests keep reasoning after an empty answer. When an auto answer comes back empty because the model spent its whole output budget reasoning, a hard, frontier or high-stakes request no longer tries again with reasoning off. It tries again with a thinking budget that leaves room for the answer, on a stronger model if the first cannot take one, or at the lowest setting that still reasons. Easy and medium requests work as before, and the empty answer is still not billed. See Models.
  • Changes to the terms take effect when published. The terms no longer promise 30 days’ notice before a change: Zabio depends on providers, models and routing that change at any time. Zabio still tells organization owners about changes that matter, by email or in the web app.
  • Notice of new sub-processors is now by subscription. Under the data processing agreement, an organization owner who writes to privacy@zabio.ai gets 30 days’ notice by email before a new sub-processor handles your content. Model vendors are listed when added, without that notice; to keep a new vendor from receiving your content, allow only the providers you accept in Model access, and a vendor added later stays off until you allow it.
  • Prices and fees can change at any time. Zabio’s prices follow the providers’, and providers change theirs without notice, so a change to Zabio’s fees no longer waits 30 days. The pricing page always shows the current fees, and a change applies only to calls and top-ups made after it. See Billing.
  • Small auto requests pay less for automatic model choice. The price of automatic model choice now follows the request: an auto request pays the same as the model’s own cost for it, at least 0.00015 USD and at most 0.00025 USD, once per request. Before, every auto request paid 0.00025 USD. A request whose model costs 0.00004 USD now pays 0.00015 USD instead of 0.00025 USD; a request whose model costs more than 0.00025 USD pays the same as before. The same range applies to auto requests on your own provider keys. zabio.cost.call_fee_micros shows what each request paid. See Billing.
  • When to name a model. The playground’s tips now say when a small request pays as much for automatic model choice as for the model itself, and what naming the model auto picks would cost instead: the service fee only. The public analyzer returns it as the name_model tip. See Billing.
  • Empty answers: the limits and the marks. An empty answer on a named model is free only when max_tokens was at least 1,000 tokens (or the model’s smallest thinking budget, when higher), and each organization’s free empty answers are capped per UTC day at the larger of 0.50 USD and 2% of what it was billed that day; past either limit the empty answer is billed and zabio.no_answer says so. Logs, the CSV export and the MCP tools recent_calls and get_call now show each call’s role (an empty answer, a failed check) and reasoning setting, and mark the attempts Zabio did not bill; Errors only keeps the empty and failed-check attempts too. ask says when it answered on a second attempt. See Checked answers.
  • An empty answer is never billed. When a model spends its whole output budget reasoning and writes nothing, that call is not billed, whoever set the reasoning. On auto Zabio tries once more with less reasoning: the same model at its lowest setting, or a model whose reasoning can be held within the budget; a reasoning_effort of low on a model that can only reason fully or not at all no longer runs at full reasoning on that retry. With nothing to change, the empty answer arrives with zabio.no_answer saying why, and the MCP tool ask says “No answer” in words. With a max_tokens of 4,000 or less, auto prefers models whose reasoning can be limited. See Checked answers.
  • No negative savings. A call that cost more than its baseline now shows a saving of 0 and “no saving on this call” (saving_note), and every total counts it as 0. direct_micros against cost_micros still shows the difference, and the MCP tool get_call shows it as difference_usd.
  • Connected apps can be allowed to ask later. Zabio’s sign-in now names read and ask, so Claude, OpenCode, VS Code and Cursor offer both on Zabio’s page (you can still untick ask). A connection whose app was never offered ask lists it, and using it makes the app ask you to allow it. Approving Claude.ai, VS Code or Cursor again for the same project replaces the earlier connection. See Allowing ask later.
  • Connect Codex, Cursor, VS Code and OpenCode. Connect an AI app now has steps for each, next to Claude and Claude Code, and Zabio is listed in the official MCP Registry as ai.zabio/zabio.
  • Totals count no saving for calls made while a baseline was pending. A call on a key that had not reached its 20 calls is compared with itself, so it shows no saving. The dashboard, Usage, the CSV export, the CLI, the MCP usage tool and batch totals now count none for it either, even after the key’s baseline is set, so every total adds up the calls as each one showed. Before, those calls’ fees counted as a negative saving. The CSV export gains saved_micros, saved_usd and baseline_pending columns at the end. See API reference.
  • What a split saved, shown on its own. On a big request Zabio answered from an excerpt, zabio.split.saved_micros (and get_call’s split.saved_usd) states what the split saved against the request sent whole, at list price. It is separate from cost.saved_micros and shows while your key’s baseline is pending too. See Split and route.
  • get_call states the same saving as recent_calls. On a key that has its baseline, get_call showed a saving of 0 for a call recent_calls showed with its real saving. Both tools now read the saving and baseline_pending the same way.
  • No saving shown while a key’s baseline is pending. An auto call on a key that has not reached the 20 calls its baseline needs no longer carries saved_micros in zabio.cost; cost.baseline still shows pending and the progress. Before, the field showed the fees as a negative saving. Code that reads saved_micros should treat a missing value as no saving. See API reference.
  • An empty answer is retried one step up, never straight to the strongest model. When a model on auto spends its whole output budget thinking and answers nothing, Zabio now sends the request to the next cheapest model at the same skill level first, and on any line but best never to the strongest model for that kind of work. See Checked answers.
  • More from connected apps. get_call and ask show how Zabio split a big request, and reconnecting an app to the same project keeps its key’s savings baseline. See Connect an AI app.
  • Connected apps keep working across Zabio updates. A field Zabio adds to an MCP tool’s answer no longer makes an app that connected earlier refuse the answer. ask now also returns cached_tokens, and its input_tokens counts only the input not read from the cache, as get_call and Logs do. Connected apps shows when an app was last used within a minute. If Claude Code refused ask answers with “does not match the tool’s output schema” on 2026-10-14, reconnect it once with /mcp. See Connect an AI app.
  • Connect Claude and other AI apps to Zabio. Zabio now runs an MCP server. Add https://mcp.zabio.ai/mcp to Claude.ai, Claude Code or another MCP app and it can read Zabio’s docs, models and prices, your project’s usage and recent calls, explain why a call went to its model, and ask Zabio’s models questions within a daily limit you set. Each connection has its own API key, and Settings, Connected apps disconnects it. https://mcp.zabio.ai/public offers the docs, models and status with no sign-in. See Connect an AI app.
  • One id for every call. Every response’s zabio object now carries call_id, the id Logs shows for the call (call_ and the same hex as a chat response’s chatcmpl- id), on chat completions streamed or not, the Responses API, embeddings and batch results. See API reference.
  • A clear answer when a model cannot take tools. A call that names gpt-6-astra and sends tools now gets unsupported_parameter with a sentence naming the limit, instead of the provider’s refusal. The same call without tools works as before, and auto never sends a call with tools there. See Errors.
  • Price cuts reach you within a day. Zabio now reads every provider’s prices every day and applies a lower price as soon as it reads it. When a provider raises a price, Zabio keeps billing the old one until it confirms the change from what the provider charges.
  • Prices checked against real charges. Every day Zabio now compares the provider cost it records for each model with what OpenAI and Anthropic actually charged it, so a price that drifts from the provider’s is found within a day.
  • A clear answer when a provider is out of service. A call that names a model on a provider Zabio cannot use right now gets the new error vendor_unavailable, naming the provider, instead of a provider error. auto never sends a call there, and when another provider serves the same model, auto moves to it first. See Errors.
  • Usage replaces Logs. A new Usage page has a Stats tab and a Logs tab under one filter bar (range, project, key, provider, model), in your time zone and kept in the page address. Stats charts spend, what Zabio saved, requests, tokens and caching, hourly or daily, stacked by model or by provider, with Top keys and Top models; Logs lists the calls with an errors-only filter, and a call opens with its details. The dashboard’s tiles link into it. See The web app.
  • Calls keep their project. A call and its spend stay in the project they happened in, even when the key moves to another project later. Usage, the call list and the CSV export follow that.
  • What caching saved, on every call. Usage shows how much of your input came from the prompt cache and what that saved at the providers’ prices.
  • The web app on phones. The menu folds behind a Menu button on small screens, so every page uses the whole width.
  • On the API: the usage routes take provider and model filters, the call list takes from, to and errors, a new series route draws the charts, and calls carry their project_id. The usage summary no longer returns openrouter_equiv_micros and openrouter_equiv_usd.
  • One match per account, from your signup. The launch match is now your account’s, once, in your first organization, and its 30 days start when you sign up instead of at your first top-up. The dashboard counts the days down. Other organizations you create, and people you invite, have no match of their own. See Credits and promotions.
  • Keys that expire. A key can now expire: never, in 7, 30 or 90 days, or on a date you choose. An expired key is refused with the new error key_expired, shows as Expired, and whoever created it gets an email 7 days before. A new expiration brings it back. See Errors.
  • Spend limits that reset. A key’s spend limit now resets every day, week or month at 00:00 UTC (weeks from Monday), or covers the key’s whole life. When it is reached, the error says when it resets, and the key list shows each key’s spend against its limit. Existing monthly caps became monthly limits. On the API, the key’s spend_cap_micros_month is now spend_limit_micros with spend_reset.
  • A simpler way to create keys. Creating a key asks for its name, expiration and spend limit; everything else starts at its default. The Edit panel holds every setting, the key’s models included, and a key’s scopes can now be changed after it is created.
  • Prompt caching on every call. Zabio now caches every call automatically, whether it names auto or a model, and a call turns it off with cache.ttl none. The key’s default prompt cache setting is gone: on the API, cache_ttl_default was removed. See cache.ttl.
  • Notice before your own plan costs more. When Zabio moves an organization that has its own pricing plan to a plan that raises any of its fees, the change takes effect at least 30 days later and the organization’s owner gets an email listing each fee that changes; a change that only lowers fees applies at once, and a cancelled change is announced too.
  • A clearer error when no model is left. When the model access of your account, organization, project or key leaves auto no model to pick, the call now answers 400 no_model_allowed and names the level to check, and the model picker warns before a change that would leave a project or key without a model for auto, naming them.
  • Model access at four levels. Your account, each organization, each project and each key now choose which models they allow, and a call may use only what every level allows; auto picks only among those. One picker at every level: grouped by provider, with filters, a switch per provider that includes its future models, a checkbox per model and the auto lines at the top, which you can now turn off one by one. On the API, the key’s allowed_models is now model_access, and the organization’s vendor switch is a provider entry. See Model access.
  • One account above your organizations. Signing up now creates your account and its first organization, Default Organization, which you can rename in Settings. Only the account’s owner creates organizations and projects. People you invite belong only to the organizations you invite them to; someone you invited who wants their own account signs up with another email, and someone you invited who later belongs to no organization gets their own account at their next sign-in.
  • Members limited to projects. When you invite a member you can choose the projects they work in, and change them later from Members. They then see only those projects’ keys, calls, usage and conversations; owners and admins see everything. A project menu next to the organization menu shows one project or all of them, and a new key is created in the project you are in.
  • Keys in a project see their own project. On the API, a key that belongs to a project reads and changes only the conversations, responses, files and batches made with that project’s keys; a key at the organization’s level sees the whole organization. See Organizations, members and keys.
  • Updates no longer interrupt your calls. When Zabio releases an update, calls already running, streams included, finish on the version that started them, and new calls go to the new version without a pause.
  • Automatic recharge on the call that crosses your threshold. A recharge could wait for your next call or for the next check a few minutes later; it now starts on the call that takes your balance below the threshold.
  • Ten models retired on Alibaba Cloud. Alibaba Cloud no longer serves qwen3.6-max-preview, deepseek-v4-flash, qwen3-next-80b-a3b-thinking, qwen3-next-80b-a3b-instruct, qwen3-30b-a3b-thinking-2507, qwen3-30b-a3b-instruct-2507, qwen3-235b-a22b-thinking-2507, qwen3-235b-a22b-instruct-2507, qwen3-vl-235b-a22b-thinking and qwen3-vl-235b-a22b-instruct on its international service. The Models page marks them retired; choose another model there.
  • Prices now match each provider’s own price list. Six OpenAI models (gpt-6.1-sol, gpt-6-sol, gpt-6-luna, gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna) bill a cache write at 1.25 times the input price, as OpenAI does; cached input in an OpenAI batch is billed at the batch input price on gpt-4.1, gpt-4.1-mini, gpt-4.1-nano, gpt-4o, gpt-4o-mini, o1, o3, o3-mini and o4-mini; cached input on gemini-3.1-pro-preview costs the standard cache price in a batch and at flex; qwen3.5-plus-2026-04-20, qwen3.6-plus and qwen3.6-flash@alibaba-us, which Alibaba Cloud gives no cache discount, bill cached input at the input price; and gemini-3.5-flash, which Google now answers with Gemini 3.6 Flash, costs Gemini 3.6 Flash’s lower price. The Models page shows every price in force.
  • New models arrive on their own. When a provider adds a model, Zabio reads its price from the provider’s own price list, checks that the model answers, and makes it callable by name within the hour. Where the provider’s own sources show two prices, you pay the lower one. The Models page lists every model you can call.
  • Lower cache prices on Claude Sonnet 5.5. Cached input on claude-sonnet-5-5 is now billed at Anthropic’s own price, 0.10 USD per million tokens (0.05 in a batch), half of what it was.
  • Provider sales reach your price. When a provider sells a model below its list price to everyone, Zabio charges the sale price.
  • Three more models you can name, each at its provider’s own list price. The Models page lists them.
  • auto picks by skill for each kind of work. Every auto request now goes to the cheapest model skilled enough for its kind of work; auto/balanced and auto/best ask for more skill, and a hard or high-stakes request moves one step up. An answer that fails Zabio’s check is sent once more to a stronger model at Zabio’s cost. The dashboard, the Logs page and zabio.routing.task_type show the kind of work. See Models and routing.
  • Routing details show the tier you asked for and the skill floor used. zabio.routing.line now reads auto, cheap, balanced or best as you sent it, and the new zabio.routing.floor says which skill floor was applied. zabio.auto_eligible, the Models page and the new GET /v1/public/auto-models show the models auto can pick today, and the request analyzer shows auto’s own picks.
  • Requests no longer fail on a parameter the model does not take. When the model that answers does not take temperature or top_p, Zabio leaves it out and names it in zabio.dropped_parameters and the x-zabio-dropped-params header instead of returning an error. stop is always honored: where a model does not take it, Zabio cuts the answer itself. The Models page lists each model.
  • The whole saving is yours, with simpler fees. Zabio no longer takes a share of what routing saves you. Every call pays the service fee on the provider price, and every auto request pays a small fixed fee for automatic model choice, once, shown as its own line (call_fee_micros). This replaces 1.2.6’s “never for choosing the model”: choosing the model now has its own small price. The fees in force are on the pricing page.
  • Calculations always use reasoning. On math, auto turns a model’s reasoning on and picks a model that can reason, so sums, averages and other calculations are worked through before the answer.
  • Reasoning levels follow each model’s own levels. Where a model names its thinking levels, low, medium and high now ask for those levels instead of a fixed number of thinking tokens, and zabio.reasoning.applied reads the level. Two models that cannot stop thinking are now treated that way: with no level they think at their own default, and none gives their lowest level. The Models page says which, so you can leave room in max_tokens.
  • Thinking shows in your usage on more models. usage.completion_tokens_details.reasoning_tokens and the dashboard now count the thinking of models that report it in their own format, so you can see which part of your output was thinking.
  • Calculations reason on more models. Some models reason only when asked, and Zabio now records which ones: auto turns their reasoning on for calculations too. When you name one of them without a level, zabio.reasoning.applied now reads off, which is how they already answered.
  • You’re never charged for an empty answer caused by a setting Zabio chose. When auto turns reasoning on, it makes sure your max_tokens leaves room for the answer. If the model still spends the whole budget thinking and writes nothing, that call is free, and a model that reasons by default answers instead, or the same model again without reasoning when there is none. With a reasoning level you set yourself, calls are billed as before.
  • Simpler routing details. zabio.routing no longer includes subject and language, and the job that POST /v1/public/analyze returns no longer includes language. Routing never used them.
  • You pay for the answer, never for choosing the model. The small routing step that reads each auto request is now part of Zabio’s fee, so classifier_micros reads 0 and the saving on every auto call is larger.
  • Lower prompt cache costs on long prompts you send once. A long prompt is now written to the provider’s cache when it comes back, not the first time, so a document you send once no longer pays the cache write. Conversations and sessions are cached from the first turn, as before.
  • Two new providers, Alibaba Cloud (US) and Alibaba Cloud (International). 34 more models you can name, the Qwen family among them, with the processing location of each shown on the Models page. Automatic routing keeps choosing from the same models as before.
  • More models to choose from. The catalog carries 43 more models you can name, from every provider Zabio serves, each at the provider’s own list price. Automatic routing keeps choosing from the same models as before.
  • A fresh look for every email. Sign-in links, receipts, balance notices and invitations now share one clear layout, addressed to you by name, with the Zabio logo, in English or Spanish, and replies reach our support team.
  • Sign in with GitHub, next to Google, a password and the email link.
  • Automatic top-ups. Credit is added automatically when your balance runs low, up to a monthly cap you choose.
  • The Free plan. Start without a card: 100 requests a day on a set of fast models.
  • Move credit between your organizations. Owners can transfer paid credit from one organization to another from the Billing page.
  • Better observability. The dashboard shows your cache hit rate, your cost per million tokens and your spend by kind of work, and a new Logs page shows how long each call took.
  • Security improvements across sign-in, accounts and payments, including better fraud protection for sign-ups and payments.