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

Models and routing

Every model Zabio serves has a maker, the company that made it (owned_by in GET /v1/models and the Maker column of the tables below: moonshot for Kimi, zai for GLM, qwen for Qwen, meta for Llama, and Anthropic, OpenAI, Google, DeepSeek, xAI, Mistral AI and TypeSafe AI for their own models), a vendor (anthropic, openai, google, deepseek, xai, mistral, together, typesafe, alibaba-us, alibaba-intl or moonshot), a class (fast, balanced, powerful or frontier, a label for the pages, not a routing rule), a set of capabilities, a processing location and a dated price. You can ask for a model by name or let Zabio choose with one of the auto models.

For the exact shape of every request and response, with curl, Python and Node examples, see the API reference.

Vendor Models Processed in
Anthropic, OpenAI, Google their own models United States
xAI Grok, through xAI’s US endpoint United States
DeepSeek DeepSeek through DeepSeek’s own API, labelled “DeepSeek (China)” China
Together AI open-weight models (Llama, Qwen, gpt-oss) and copies of the DeepSeek models with ids ending in @together, labelled “processed in North America” North America
Mistral AI Mistral models, through Mistral’s global API not guaranteed: Mistral does not commit to an inference location there
TypeSafe AI Jev (jev-1.14.0 and jev-1.13.0), a decision model: typed answers with calibrated probabilities, called by name or chosen by auto for a classification job, see Decision models United States
Alibaba Cloud (US) Qwen and GLM models through Alibaba Cloud Model Studio’s Virginia workspace, with ids ending in @alibaba-us US (Virginia): data stored and processed only in the United States
Alibaba Cloud (International) Qwen models and DeepSeek V4 Flash through Alibaba Cloud Model Studio’s Singapore workspace, and copies of DeepSeek V4.1 Flash, GLM and Qwen 3.8 2.4T with ids ending in @alibaba-intl Outside mainland China: data stored in Singapore and processed anywhere outside mainland China
Moonshot AI Kimi K3.1 Pro (kimi-k3.1-pro, kimi-k3.1-pro-highspeed) and a copy of Kimi K3 (kimi-k3@moonshot), through Moonshot AI’s international API not guaranteed: Moonshot AI commits to no inference location; its privacy policy places its servers in Singapore
DeepInfra copies of open-weight models (DeepSeek, Qwen, Llama, gpt-oss, MiniMax) with ids ending in @deepinfra, served at the precision DeepInfra states for each (fp8, or the maker’s full precision for qwen3.5-9b@deepinfra and gpt-oss-120b@deepinfra) not guaranteed: DeepInfra commits to no inference location; its privacy policy says data is processed in the United States
Parasail copies of open-weight models (DeepSeek, Qwen, Llama, GLM, MiniMax) with ids ending in @parasail, served at the precision Parasail states for each (fp8, or the maker’s full precision for qwen3.5-9b@parasail) not guaranteed: Parasail commits to no inference location; its privacy policy says data may be processed in the United States, the European Union or elsewhere

Every vendor is on by default. Owners and admins can turn a whole vendor off, or single models, on the Models page of the web app; a vendor that is off never receives your requests. The DeepSeek models exist twice so you can choose: turn DeepSeek off and keep Together AI on to use the DeepSeek models processed in North America only.

auto chooses among the models with a measured skill for the request’s kind of work (see How routing picks a model), from the vendors and models your organization and key allow. Every model in the catalog can also be named. The decision models (jev-1.14.0, jev-1.13.0) are never candidates, but auto may answer a classification job with Jev (see Classification jobs on auto). When you want a narrower field, name a line (auto/best, auto/balanced), disable vendors or models for your organization, or set an allow list on the key.

When a vendor announces that a model will be retired, the model carries retires_at with that date. auto stops choosing it 10 days before the date; requests that name it keep working until the date, and every one of them carries the x-zabio-retires-at header with it. On and after the date a request that names the model gets a 410 model_retired that says when it was retired, before any vendor work, on every endpoint and in a batch. When the vendor names a model that replaces it, the model also carries successor with that model’s id, and the 410 names it too, so you can move to it before the date. When a vendor stops serving a model without announcing it (its model list has not shown the model for a day and a test call is refused), Zabio sets retires_at seven days ahead on its own and the same applies; if the vendor lists the model again before that date, the date is removed.

A line says how skilled the model must be. The classifier says what the request is: its kind of work, how hard it is and what is at stake. Every candidate has a skill score for that kind of work, and the line sets the floor; the cheapest model at or above the floor answers.

Model id What you get
auto/best the cheapest model whose skill for this kind of work is in the top third of the candidates
auto/balanced the cheapest model whose skill for this kind of work is in the middle third of the candidates or above
auto/cheap the cheapest model skilled enough for the job: any candidate with a skill score for this kind of work

Plain auto (or no model) is not one of the three lines. Zabio picks for you: the cheapest model able to do the job, moving up a line when the job needs it. On a kind of work with skill scores, plain auto has no floor of its own: any candidate with a skill score for the work can answer, and the cheapest does. Name a line when you want a stronger model on every request.

A hard request, or one with high stakes, moves the floor one step up: auto/cheap to the middle third, auto/balanced to the top third; auto/best is already at the top. Plain auto moves to the middle third on a hard job, and to the top third, as auto/best does, on a high-stakes one. When the cheapest model already meets the higher floor, nothing changes. zabio.routing.line shows the tier as you asked for it (auto for plain auto), and zabio.routing.floor the floor applied after any step up (none, middle or top). The call’s record shows the line Zabio applied (cheap, balanced or best): the MCP get_call tool and escalated_from name it. You can also name any model directly.

  1. The job. A small, fast model reads the request (the first part of your system prompt, the last user message, the names of your tools, whether images are attached) and names the job: its kind of work (zabio.routing.task_type, one of the tags in Kinds of work), a difficulty from 1 to 5, whether the stakes are high, with a confidence and a one line reason. You see all of it in zabio.routing. It never sees a price or a model name. When it is unsure, it rounds the difficulty up.
  2. Constraints. zabio.vendors limits the candidates to the vendors you list. Models switched off for your organization or missing from the key’s allow list are out.
  3. Capabilities. A request with images needs a vision model, a request with audio needs a model that takes audio input, a request for a spoken answer needs a model that answers in audio in the asked format and voice (zabio.capabilities.audio_output; auto does not answer in audio yet, see Audio answers), a request with video needs a model that takes video input (zabio.capabilities.video_input) whose vendor takes that video (YouTube URLs and WebM go to Gemini models only), tools need a tool capable model, response_format: json_schema needs structured output support, and the prompt plus max_tokens must fit the context window; with no max_tokens, auto keeps at least 4,096 tokens of the window free for the answer. Models that fail a check are skipped.
  4. Skill and price. Every candidate has a skill score for the kind of work, measured for that kind of work and ranked against the other current models. Only models with a score are candidates. The line sets a floor on that score (none, the middle third or the top third of the candidates), and the cheapest candidate at or above it answers. The price is this request’s: the size of your prompt, the output the model is expected to write with its reasoning, and a session’s warm prompt cache. Equal prices go to the model with the earlier id.
  5. Difficulty and stakes. A hard request, or one with high stakes, moves the floor one step up. A job is high stakes when acting on a wrong answer could cost more than asking again (harm to health or safety, money, a legal position or security, an action that cannot be undone, or a change to a live production system). Urgency, strong feelings, a support setting or an upset customer do not make a job high stakes by themselves. Before a job is marked high stakes, a second, narrower question asks what a mistake would actually cost. Its answer is used only on those jobs. On streamed calls and calls with tools it is asked together with the first question, so the answer does not wait for it; when it is not needed, Zabio pays for it.
  6. Attempts. If the chosen provider fails or does not answer in time, Zabio moves to the next source of the same model, or to the next model by the same rule, up to 3 attempts in total. zabio.attempts tells you how many were needed.

When no skill is measured. A request of kind other, or of a kind of work with no skill scores yet, is routed by Zabio’s earlier rule. For the job and its difficulty, Zabio knows how well each model does and how sure that knowledge is; sorted by the price of this request, the candidates form a curve, and a model that a cheaper one matches is never chosen. auto/best stops climbing when nothing pricier is probably better, auto/cheap stops at the first model that is not clearly worse than the top, auto/balanced sits between them. Plain auto takes the cheapest model on an easy job and the best rule on a high-stakes one; on other medium and harder jobs it picks as auto/balanced does, because under this rule Zabio cannot confirm that a cheap model is able to do the job. Under this rule a high-stakes job runs cheap as balanced and balanced as best, and frontier models are candidates on auto/best only. zabio.routing.cell names the knowledge it used. Where that knowledge comes from is described on the benchmarks page.

The account, the organization, a project and the key each choose which models they allow, and a call may use only what every level allows; auto only picks from those models. See Model access.

reasoning_effort on chat completions and reasoning.effort on the Responses API take one scale for every vendor: none, minimal, low, medium, high and xhigh, OpenAI’s names, so OpenAI clients work unchanged. Each model has a row in Zabio’s reasoning table, shown in the Reasoning column of the tables below and in zabio.reasoning on GET /v1/models: named levels (OpenAI, Anthropic, DeepSeek), Google’s thinking levels (Gemini: low, medium and high), on or off (Together AI’s thinking models), always on (grok-4.7), or no reasoning at all. A level the model does not have maps to the closest one it has, the lower of two equally close, and a request never fails because of the level. A level sent to a model without reasoning is ignored, listed in zabio.ignored_parameters, and zabio.reasoning.applied says not supported. DeepSeek’s own API runs its models with reasoning off when the request carries tools (DeepSeek requires the model’s earlier reasoning to be sent back on tool turns, which Zabio does not keep); Together AI’s and Alibaba Cloud’s thinking models keep the setting you chose with tools, with one exception on Alibaba Cloud: its function calling page states that a model in thinking mode does not take a forced tool choice, so a request with tool_choice: required or a named function on a Qwen or GLM model runs with thinking off, zabio.reasoning.applied says off and a level you sent is listed in zabio.ignored_parameters with the reason tool_choice; with tool_choice: auto the model keeps your level. OpenAI’s gpt-6-luna and gpt-6-sol take tools with reasoning on the Responses API, where Zabio sends OpenAI calls; on the calls that stay on chat completions (a request with stop, a call on your own OpenAI key, or max_tokens under 16) they run without reasoning when tools are present, zabio.reasoning.applied says off and a level you sent is listed in zabio.ignored_parameters. Every response reports zabio.reasoning: requested (what you sent) and applied (default for the vendor’s own default, off, the vendor’s level name, budget N, on, always on or not supported).

When you set no level on auto, Zabio chooses the reasoning setting together with the model. On easy and medium jobs the cheap and balanced lines use the model’s lowest setting: off where the model can stop reasoning, its lowest level otherwise. Zabio’s own tests found that on those jobs more thinking did not change the answers, only the cost and the wait, and reasoning tokens are often most of what a reasoning model bills. The best line, high-stakes jobs, and hard and frontier jobs keep the model’s default. Calculations always reason: on a math job, a model whose reasoning is off by default runs at its lowest setting that reasons, and a model that cannot reason answers only when no candidate can. Reasoning that Zabio turns on needs room: when max_tokens is too small for it on every candidate, the math job runs at the models’ default settings instead. If an answer comes back empty because the reasoning spent the whole budget, that call is not billed, whoever set the reasoning. Two limits apply. On a named model it is free only when max_tokens was at least 1,000 tokens, or the model’s smallest thinking budget when that is higher; under that you chose both the model and a budget too small for its reasoning, and the empty answer is billed. 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 that, empty answers are billed. zabio.no_answer says which applied. On auto the request then runs once more with less reasoning: the same model at its lowest setting where it can reason less (a model with levels drops to its lowest, an on/off model turns reasoning off, a thinking budget is held to half of max_tokens), or else the cheapest model at the same skill floor whose reasoning can be held within max_tokens, at its lowest setting; models that always reason are left out of that move. Hard and frontier jobs and high-stakes jobs never retry with reasoning off: the same model runs again with a thinking budget of half of max_tokens when the model takes a budget; otherwise the request moves to a stronger model that takes one, with the same budget; otherwise the same model runs again at its lowest setting that still reasons. A stream or a batch request retries on the same model only. When that answer is empty too, or there is nothing to change, you receive the empty answer with zabio.no_answer saying why, and it is not billed. A named model is never retried. With a max_tokens of 4,000 or less, auto prefers models whose reasoning can be limited over models that always reason. zabio.routing.effort says what was chosen and zabio.routing.effort_reason why. A level you set holds on the first attempt. Thinking tokens count inside max_tokens, so leave room above the thinking: set max_tokens to the answer you need plus the thinking you allow.

Send zabio.session_id (any string up to 128 characters; the header x-zabio-session-id also works) on calls that belong together. Within a session the job is classified once per user message and reused across the tool-result turns that follow, so those calls do not pay for a new classification. Zabio classifies again on a new user message and when a signal fires:

Trigger (as reported in zabio.routing.trigger) Meaning
no_prior_decision first call of the session
cache_hit the previous decision was reused
low_confidence_prior the earlier decision had confidence below 0.7
long_idle more than 30 minutes since the last call
tool_failure_streak three tool results in a row looked like errors
customer_feedback the user wrote something like “that is wrong” or “try again”
new_human_message a new user message in the session; the job is classified again
explicit a model was named; no routing
escalation the model asked for a stronger model (see below)

Session state is metadata only (line, model, confidence, a few counters, the size of the last prompt and the cache lifetime applied to it). It lives for one hour after the last call and at most 24 hours. Prompts are never stored for it.

Within a session, routing counts the prompt cache. After a turn, the model that served it holds your context in its vendor’s cache for a while (Anthropic: the 5 minute or 1 hour lifetime that was applied; DeepSeek, xAI and Mistral: about five minutes, since they cache without markers; OpenAI: 30 minutes by default on GPT-5.6 and later and a few minutes on its older models, and routing counts five minutes for it, the shortest figure OpenAI states). On the next turn Zabio prices that model with the cached part of the prompt at its cache read rate, and every other candidate with the same part at the fresh rate plus the cache write it would pay, and the line applies its usual rule to those prices. So a conversation stays on its model when moving would cost more than it saves, and still moves when another model is cheaper by more than the cost of warming it, when the job changes on a new message, or when the model asks to escalate. There is no pin: the cheapest honest price wins. Google and Together AI models get neither the discount nor the write, because Zabio has no cache to count there. When the cache decided the pick, zabio.routing.sticky names the model kept and the one cold pricing would have chosen; otherwise the field is absent.

A request under auto that carries a large source (a contract, a log, a support thread, a set of reviews or records) together with a question that needs only part of it can be answered in two steps when that is cheaper than sending the whole source to the model routing chose. Zabio calls this split and route. Version 1 does one kind of split, distill then answer:

  1. Who qualifies. Only a virtual model (auto, auto/best, auto/balanced, auto/cheap, or no model); a named model is never split. The last user message must be plain text of at least about 4,000 tokens on Zabio’s estimate, with no image parts and no tool results, the request must carry no tools, and Zabio’s free estimate of the unsplit call must be at or above 0.010 USD. Cheaper calls run exactly as before.
  2. The questions. A second call to Jev (TypeSafe’s decision model, the one that classifies the job, on Zabio’s key and at Zabio’s cost) reads the same excerpt the classifier reads and says whether the request is a large source plus a separate question that needs at most about half of it.
  3. The price check. Zabio prices both paths: the whole request on the routed model, and the extract on the cheap model plus the check plus the answer on the excerpt. The split goes ahead only when it is estimated at 70 percent of the unsplit call or less.
  4. The extract. A cheap model (gpt-5.4-nano, or gemini-3.5-flash-lite when the first declines or returns nothing) copies the lines of the source the question needs, word for word and in their order. Its output is streamed and cut off past 40 percent of the source. Every line that does not occur in the source is dropped, so nothing the extractor rewrote or invented reaches the answering model. An extract with fewer than half of its lines found in the source, an empty one, or one above 40 percent of the source stops the split.
  5. The check. Jev reads the beginning and the end of your message and the excerpt and says whether the excerpt holds what the question names; below 0.5 the split stops.
  6. The answer. The routed model, at the setting routing chose, receives your message with its first 2,000 and last 2,000 characters kept verbatim (on line boundaries, where the instructions usually live) and the source between them replaced by the excerpt under this line: [Zabio excerpt: the material between these markers was shortened by a preprocessing step to the lines relevant to the request (K of N lines kept, in their original order). If you need more of the original material to answer well, reply with exactly the line NEED_MORE_SOURCE and nothing else.], followed by the excerpt and [end of excerpt]. Everything else about the call (the system prompt, earlier messages, max_tokens, response_format) is unchanged.
  7. The way back. When the model answers with the line NEED_MORE_SOURCE, Zabio sends the full request to the same model and returns that answer. On a stream the reserved line is read on the first frames, before anything reaches you, and the full-source stream replaces it, so the first tokens may take a moment longer.

Every guard that fails sends the request whole to the routed model, as if nothing had happened, and zabio.split.fallback says which one: extractor_declined (the extractor said the task needs the whole material), empty_extract, extract_too_big, low_verbatim, check_rejected, need_more_source, extract_failed, check_failed. A split that fell back never costs you more than the unsplit call; see Billing.

What you see: zabio.split on every response while the feature is on, with applied, the reason when the request was not split (named_model, opt_out, not_eligible, below_threshold, jev_no, jev_failed, price_check), every call of the split (steps: the questions, the extract, the check and the answer, each with its vendor, model, tokens, list price and whether it was billed), the excerpt’s size and how much of the extractor’s output was found verbatim in your source, and the check’s probability. zabio.model_served and zabio.routing keep naming the model that answered and how it was chosen. To send a request whole, set zabio.split: "off" or the header x-zabio-split: off; there is no per-key default, the choice lives in the request.

Not covered in version 1: only the last user message is split. A large source that lives in earlier turns or in a stored conversation’s history is not distilled (compaction handles size there), nor are several large messages, images, tool results, or a previous_response_id chain whose source arrived turns ago. Within a session, the turn after a split prices no warm cache, because what the vendor cached was the rewritten message.

OpenAI and Google serve some models on a flex tier: the same model at about half the price, slower (their own latency target is up to 15 minutes) and with best-effort capacity. A request that can wait says so with zabio.deadline (or the header x-zabio-deadline), a duration of 15 minutes or more; under 15 minutes, or with no deadline, nothing changes. The models with a flex price are OpenAI’s GPT-5, GPT-5.x and GPT-6 chat models together with o3 and o4-mini, and every Google Gemini model; the GPT-4 family, o1, o3-mini, chat-latest and every other model is served and priced at standard, deadline or not.

With a deadline, auto prices every candidate that has a flex tier at its flex price and the rest at standard, then runs the line rule as usual, so a model can win because it is cheaper in flex. A named model with a flex price is served at flex. When the vendor refuses a flex request for capacity, Zabio retries the same model at the standard tier at once and bills the price actually used. The pre-steps of a split (the extract and the checks) run at the standard tier; only the answer can be flex. zabio.tier on the response says what was requested, what served the call and whether capacity forced the standard tier. See Chat completions for the field and Billing for the fee.

On auto turns of the Responses API, Zabio may give the model a hidden tool it can call when a task is beyond it. When that happens, Zabio re-runs the same turn on a stronger model, with the full history, and returns that answer: the cheapest model one step up the skill floor for the kind of work, or, for a request routed by the earlier rule, the next model up the curve that its knowledge says is probably better. The response carries zabio.escalated_from with the line and model that gave up (for example cheap/gpt-5.4-nano). The escalation itself never appears in your output. Escalation is not offered on /v1/chat/completions.

On a request that is not streamed, Zabio checks the answer against what the request asked for before it reaches you:

  • with response_format json_schema: the answer is JSON, and nothing else, that matches your schema (types, required fields, enums, lengths, numeric bounds, patterns);
  • with response_format json_object: the answer is one JSON object;
  • with tools: every tool call names one of your tools and its arguments are JSON that matches that tool’s parameters, and a tool_choice of required or a tool’s name gets a call;
  • on any request: the answer is not empty because the model spent its whole output budget thinking (finish_reason length with no content).

When the answer to an auto request fails and auto chose the model by its skill score, it sends the request once more to the cheapest model one step up the skill floor, the failed model left out, and you receive that answer. An answer left empty because the model spent its whole output budget thinking is never billed, and it follows the reasoning rule instead: the same model once more with less reasoning, or the cheapest model at the same skill floor whose reasoning can be bounded, never the strongest model auto can pick on a line other than best (see Reasoning levels and effort inside auto). You pay only for the answer you receive: the failed attempt is Zabio’s cost and is not in zabio.cost. When another model answers, zabio.escalated_from names the one that failed; zabio.attempts counts both attempts. A request is retried once at most; when the second answer fails too, you receive it as it is. Named models are checked and never retried: you chose the model; an empty answer on one is still not billed. Streamed requests are not checked, because the answer reaches you while it is being written; an empty streamed answer is answered again with less reasoning before you see it where the model can reason less, and otherwise reaches you unbilled with zabio.no_answer. A schema is read with references inside the schema only; Zabio never fetches a referenced file or URL.

Name any model id from the table below, for example claude-sonnet-5 or gpt-5.4. The request goes to that model. If the model is unhealthy or unavailable, or does not answer within Zabio’s deadline (3 minutes for fast and balanced class models, 10 minutes for powerful and frontier; a stream must send its first data within 60 seconds on fast and balanced, 10 minutes on powerful and frontier, and Zabio keeps the connection alive meanwhile, see Long calls), zabio.fallback decides what happens:

zabio.fallback Behaviour
same_vendor (default) another model of the same vendor and class
cross_vendor any model of the same class
none fail instead of falling back

The default comes from the key setting fallback_policy. The response always tells you what was served in zabio.model_served and zabio.vendor.

A max_tokens too large for the model you named does not move you to another model: Zabio lowers it to what the model can produce beside your input, never below 4,096 tokens, and keeps the model; on Anthropic models it is sent as you set it, since Anthropic stops at the window. The call falls back only when the input itself does not fit the model’s context window, or leaves less room than that.

OpenAI takes function tools for gpt-6-astra on its Responses API only, so Zabio sends every gpt-6-astra request that carries tools there; requests without tools are served as before. You call it the same way on /v1/chat/completions and /v1/responses, streamed or not, and tool calls come back in the usual shape. Some requests with tools cannot go there, and Zabio answers them with a 400 unsupported_parameter whose message names the limit, before any vendor work and without moving you to another model:

  • stop set. OpenAI’s Responses API has no stop sequences, and Zabio never bills you for text past your own stop. Send the request without stop, or without tools.
  • max_tokens from 1 to 15. OpenAI’s Responses API takes 16 at least, and sending 16 would bill past your limit. Use 16 or more, or leave max_tokens out. If you leave it out and your balance pays for fewer than 16 output tokens of this model, Zabio lowers the limit to what your balance pays for, so as not to bill past it, and OpenAI then refuses the call with its own error (not billed); add credit and send it again.
  • Your organization’s own OpenAI key. Zabio does not control that key’s access to the Responses API, so tools on gpt-6-astra run only on Zabio’s key. Remove the key under Billing, vendor keys, or send the request without tools.
  • Audio. gpt-6-astra takes no audio input, so a request with an audio clip is refused for that reason, tools or not.

Under auto, a request with tools that hits one of these limits is never sent to gpt-6-astra: another model with tools serves it.

In a batch or a background response, gpt-6-astra takes tools too: Zabio sends its lines with tools to OpenAI’s batch for the Responses API, at the batch price, and the limits above apply to each line as to a normal call.

Jev is TypeSafe AI’s “System One” model, in two versions you can name: jev-1.14.0, the newest, and jev-1.13.0. Both answer the questions below at the same price and limits, and each answers as the version you named. It does not write text. It answers typed questions about a conversation and returns, with every answer, the calibrated probability of each option. That makes it a fit for the jobs where you want a label, a yes or no, or a score and nothing else: tagging, routing, moderation, sentiment, language detection, urgency. It answers in well under a second and costs 0.042 USD per million input tokens, with output free; Zabio adds the service fee of (see the pricing page), as on any call (see Billing).

You call it through /v1/chat/completions like any other model, with response_format set to json_schema. Each property of the schema’s object becomes one question:

Property in your schema Jev question What you get back
a string with an enum of 2 to 255 values a choice one of the values, with the probability of every value
a boolean a yes or no true or false, with the probability of yes
an integer with minimum and maximum 2 to 10 levels apart a score the weighted level rounded to the nearest whole number, with the probability of every level
a number with minimum and maximum 2 to 10 levels apart a score the weighted level itself, for example 0.27 on a 0 to 1 scale

usage.completion_tokens is the count TypeSafe reports for its answers; those tokens are free. Give every property a description: Jev reads it as the question. The schema’s own description is read as the shared instructions. The answer keeps your property order. What Jev reads is the whole conversation: the system prompt and every user and assistant message.

What Jev cannot answer is refused with 400 unsupported_for_model before anything is sent, and param names the field: a request without json_schema, a free text string, a nested object, an array, a number without bounds or with more than 10 levels, a nullable or anyOf property, tools, images, tool results, and stream: true (Jev answers in one piece).

A worked example: a support ticket into one of five categories, with a boolean for urgency and a 1 to 5 priority.

Terminal window
curl https://api.zabio.ai/v1/chat/completions \
-H "Authorization: Bearer $ZABIO_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "jev-1.13.0",
"messages": [
{"role": "system", "content": "You triage support tickets for a hosting company."},
{"role": "user", "content": "My card was charged twice this month. Please fix it today, I am about to travel."}
],
"response_format": {"type": "json_schema", "json_schema": {"name": "ticket", "strict": true, "schema": {
"type": "object",
"description": "Classify the support ticket in the conversation.",
"properties": {
"category": {"type": "string", "enum": ["billing", "bug", "feature_request", "account", "other"], "description": "The kind of request."},
"urgent": {"type": "boolean", "description": "The customer needs an answer today."},
"priority": {"type": "integer", "minimum": 1, "maximum": 5, "description": "How important the ticket is, 5 the most."}
},
"required": ["category", "urgent", "priority"],
"additionalProperties": false
}}}
}'

The answer is the assistant message’s content, valid against your schema, and zabio.decision carries the probabilities behind it:

{
"id": "chatcmpl-01a0e9d47c4d75d385ba714a2736ca34",
"object": "chat.completion",
"model": "jev-1.13.0",
"choices": [{"index": 0, "message": {"role": "assistant", "content": "{\"category\":\"billing\",\"urgent\":true,\"priority\":4}"}, "finish_reason": "stop", "logprobs": null}],
"usage": {"prompt_tokens": 484, "completion_tokens": 83, "total_tokens": 567},
"zabio": {
"vendor": "typesafe",
"model_served": "jev-1.13.0",
"routing": {"reason": "explicit model", "trigger": "explicit"},
"cost_micros": 21,
"cost": {"vendor_micros": 20, "service_fee_micros": 1, "call_fee_micros": 0, "savings_fee_micros": 0, "classifier_micros": 0},
"decision": {
"category": {"type": "choice", "probabilities": {"billing": 1, "bug": 0, "feature_request": 0, "account": 0, "other": 0}, "confidence": 1},
"urgent": {"type": "noul", "probability": 0.84},
"priority": {"type": "score", "value": 4.01, "probabilities": {"1": 0, "2": 0.02, "3": 0.11, "4": 0.71, "5": 0.16}, "confidence": 0.75}
},
"cache": {"read_tokens": 0, "write_tokens": 0, "ttl_applied": "none", "applied": "none"}
}
}

Read zabio.decision when you want to act on confidence: send the ticket to a person when the top category is under 0.7, or pick your own threshold for urgent instead of the 0.5 that decides the boolean.

Health, fallback and cost work as for any named model. Neither Jev version is a fallback for the other, so under the default same_vendor policy a TypeSafe outage fails the call with 503 no_provider_available; under cross_vendor a fast text model with structured output answers the same schema instead, without zabio.decision, and zabio.model_served says so. Jev is never a fallback for a text model. A request to Jev is sent whole to TypeSafe AI in the United States, like a request to any vendor you name; see Privacy.

Under auto (and auto/best, auto/balanced, auto/cheap), Zabio may answer a classification job with Jev (jev-1.14.0) instead of the text model routing chose. A request qualifies when its response_format is a json_schema whose every property is one Jev can answer (a string with an enum, a boolean, an integer or number with minimum and maximum 2 to 10 levels apart), with no tools, no images and no stream. Everything else routes exactly as always.

Which of the two answers depends on what Zabio has measured. On an easy classification job (a support ticket, an email’s intent, a review’s sentiment, a moderation check, a language), our tests of November 2026 compared Jev with the text models routing picks for such jobs, on 180 requests: Jev got every field right on 85 percent of them, gpt-6-luna on 88 percent, which our tests count as level with Jev, and claude-haiku-5-5 on 81 percent. There the line’s rule chooses between Jev and the routed model on those measurements, as on medium jobs below. Zabio’s estimate of each side’s price counts what that side is billed for: for the routed model, the tokens your JSON schema adds to the input and the thinking of a model whose setting reasons; for Jev, TypeSafe’s own count. On our test requests about 6 percent of easy jobs went to gpt-6-luna, each one billed at or below Jev’s price, and Jev answered the rest. A routed model our tests have not measured on easy jobs does not answer while Jev is measured: Jev answers, which on a short request can cost slightly more than that model would have, at most Jev’s own price for the call, which stayed under 0.00004 USD on our test requests under 200 input tokens. In September 2026 Jev got every field right on 78 percent of such requests, and the cheapest text model of the time on 51 percent.

On a classification job rated medium (a long ticket thread, a long email or lead, a detailed review), our tests of November 2026 compared Jev with the text models routing picks for such jobs, on the same requests: Jev got every field right on 82 percent of them, gpt-6-luna on 80 percent, which our tests count as level with Jev, and claude-haiku-5-5 on 70 percent. There the line’s rule chooses between Jev and the routed model on those measurements, as it chooses between any two models: of two models measured level the one Zabio estimates cheaper answers, and a model measured worse than Jev does not answer on the best and balanced lines. Jev was the cheaper one on every medium request we tested, so it answers these jobs as before; a routed model our tests have not measured on medium jobs does not answer while Jev is measured: Jev answers, which can cost slightly more than that model would have, at most Jev’s own price for the call.

On any other qualifying job (a classification job rated hard, or one the classifier reads as another kind of work), where we have not measured Jev against the routed models, Zabio compares its estimate of Jev’s price with the routed model’s for the same request and asks Jev only when Jev is cheaper.

A safety net: Jev returns the probability of every answer, and when any field of its answer is under 0.5 (the option it chose is less likely than not), Zabio discards the answer and sends the request to the text model as always. That Jev call is Zabio’s cost, never yours. On the first study set behind this feature (September 2026) Jev was right on 92 percent of fields against 75 percent for the text models then routed, and its answers came back in about 0.15 seconds.

How you can tell: zabio.model_served is jev-1.14.0, zabio.vendor is typesafe, and zabio.decision carries the probabilities, as when you name Jev. The answer keeps your schema’s shape and property order. zabio.routing still describes the job the classifier saw. Billing is as for any auto call: Jev’s price plus the service fee and the fee for automatic model choice (see Billing). When Jev’s answer is discarded, you pay exactly what the text model’s call costs. A kept answer is billed at Jev’s price, which on a short, easy classification job can be slightly above the text model’s, as described above.

Naming a model keeps full control: a request to gpt-5.4-nano or any other named model is never answered by Jev, and a request that names jev-1.14.0 or jev-1.13.0 is always answered by that version of Jev. A request to Jev on auto is sent whole to TypeSafe AI in the United States, like a request that names it; see Privacy.

GET /v1/models returns the auto models followed by the catalog, priced for your organization:

{
"object": "list",
"data": [
{"id": "auto", "object": "model", "created": 1757808000, "owned_by": "zabio",
"zabio": {"tier": "auto", "auto_eligible": true, "virtual": true}},
{"id": "claude-sonnet-5", "object": "model", "created": 1757808000, "owned_by": "anthropic",
"zabio": {"tier": "balanced", "auto_eligible": true,
"capabilities": {"vision": true, "tools": true, "json_schema": true, "context_window": 1000000, "max_output": 128000},
"pricing": {"input": 2.10, "output": 10.50, "cache_read": 0.21, "cache_write": 2.63, "currency": "USD"},
"residency": "global"}}
]
}

GET /v1/models/{id} returns one model; auto/cheap style ids work (percent encode the slash if your client needs it). Keys with an allow list only see their models. Unknown ids return 404 model_not_found.

pricing is the provider list price plus your plan’s service fee, rounded up to the cent, in USD per million tokens. cache_write is the 5 minute cache write rate. See Billing for how fees work.

zabio.free_tier is true on the models of the free model list, the ones a Free account can name; GET /v1/public/free-tier serves that list with no key. While your organization runs on the Free limits, every model also carries zabio.availability: available, or locked for a model off the list, which a request may not name until a top-up (403 free_tier_model). The auto ids are always available and choose among the free models. See Billing, “The Free plan”.

The table below is generated from the catalog in the code. Prices are the provider list prices in USD per million tokens. Zabio bills this price plus the service fee of (see the pricing page), and on every auto request the price of automatic model choice, the model’s own cost for the request, at least (see the pricing page) and at most (see the pricing page)a fixed (see the pricing page); Zabio takes no cut of what routing saves (see Billing). Cache write rates: Anthropic charges 1.25 times the input price for a 5 minute cache and 2 times for a 1 hour cache; OpenAI charges 1.25 times the input price on GPT-5.6 and later (the cache write column) and nothing extra on its earlier models; Google charges nothing extra to write. On GPT-5.6 and later OpenAI also sells a 12 hour cache lifetime at 2.5 times the input price; Zabio never asks for it, so every cache write there is the default 30 minute one, billed at the cache write column, and a lifetime field you send (prompt_cache_options, prompt_cache_retention) is not passed on. Google’s 3.x Flash promotional prices end on 2026-12-31. On the models with a flex tier (OpenAI’s GPT-5, GPT-5.x and GPT-6 chat models, o3 and o4-mini, and every Google Gemini model) a call with a zabio.deadline of 15 minutes or more is billed at half the listed input, output and cache read prices (the cached input of gpt-5.4 at 0.13 USD per million and of o4-mini at 0.138, as OpenAI lists them); see Flex tier.

The Auto column says whether auto can pick the model today: it has a skill score for at least one kind of work, or the earlier rule can choose it for a request of kind other (see How routing picks a model). It follows the daily skill data, so the page reads it live when it loads; zabio.auto_eligible in GET /v1/models and GET /v1/public/auto-models carry the same answer. Every model can be named. Most of the models marked no were added so that every model you can reach through other routers is also reachable here, at the vendor’s own list price.

Prices as of November 20, 2026, in USD per million tokens, the latest change in Zabio’s catalog.

Model Maker Processed in Auto Context window Max output Input Output Cache read Reasoning Off-peak input / output
claude-haiku-4-5 Anthropic United States yes 200K 64K 1.00 5.00 0.10 no -
gpt-5.4-nano OpenAI United States yes 400K 128K 0.20 1.25 0.02 levels, none to xhigh, default off -
gpt-5-mini OpenAI United States yes 400K 128K 0.25 2.00 0.025 levels, minimal to high, default the vendor’s, always reasons -
gpt-4.1-mini OpenAI United States no 1M 32768 0.40 1.60 0.10 no -
gpt-4o-mini OpenAI United States no 128K 16384 0.15 0.60 0.075 no -
gemini-3.5-flash-lite Google United States yes 1M 65536 0.30 2.50 0.03 levels, none to high, default the vendor’s -
mistral-small-2603 Mistral AI Not guaranteed yes 256K 256K 0.15 0.60 0.015 no -
qwen3.5-9b Qwen North America yes 262144 262144 0.17 0.25 0.00 on or off, default off -
gpt-6-luna OpenAI United States no 400K 128K 0.10 0.50 0.01 levels, none to xhigh, default the vendor’s -
gpt-5.6-luna OpenAI United States no 400K 128K 0.20 1.20 0.02 levels, none to xhigh, default the vendor’s -
gpt-5-nano OpenAI United States no 400K 128K 0.05 0.40 0.005 levels, minimal to high, default the vendor’s, always reasons -
gpt-4.1-nano OpenAI United States no 1M 32768 0.10 0.40 0.025 no -
gemini-3.1-flash-lite Google United States no 1M 65536 0.25 1.50 0.025 levels, none to high, default the vendor’s -
ministral-3b-2512 Mistral AI Not guaranteed no 131072 131072 0.10 0.10 0.01 no -
ministral-8b-2512 Mistral AI Not guaranteed no 262144 262144 0.15 0.15 0.015 no -
ministral-14b-2512 Mistral AI Not guaranteed no 262144 262144 0.20 0.20 0.02 no -
glm-5.3-flash Z.ai North America no 1048575 1048575 0.15 0.50 0.03 always on -
muse-glimmer-30b Meta North America no 131072 131072 0.35 1.50 0.04 always on -
deepseek-v4-flash-0731 DeepSeek North America no 1048576 1048576 0.14 0.28 0.03 on or off, default off -
minimax-m3 MiniMax North America no 524288 524288 0.30 1.20 0.06 always on -
claude-haiku-5-5 Anthropic United States no 1M 128K 0.10 0.50 0.01 levels, none to xhigh, default the vendor’s -
qwen3.8-omni-flash Qwen Outside mainland China no 1M 131072 0.15 0.47 0.016 budgets, none to high, default off -
qwen3.8-flash Qwen Outside mainland China no 1M 131072 0.15 0.47 0.03 budgets, none to high, default off -
qwen3.7-plus Qwen Outside mainland China no 1M 131072 0.32 1.28 0.064 budgets, none to high, default off -
qwen3.5-plus-2026-04-20 Qwen Outside mainland China no 1M 65536 0.40 2.40 0.00 budgets, none to high, default off -
qwen3.5-plus-2026-02-15 Qwen Outside mainland China no 1M 65536 0.40 2.40 0.00 budgets, none to high, default off -
qwen3.5-flash-2026-02-23 Qwen Outside mainland China no 1M 65536 0.10 0.40 0.00 budgets, none to high, default off -
qwen3.6-flash Qwen Outside mainland China no 1M 65536 0.25 1.50 0.05 budgets, none to high, default off -
qwen3.6-35b-a3b Qwen Outside mainland China no 262144 65536 0.375 2.25 0.00 budgets, none to high, default off -
qwen3.5-35b-a3b Qwen Outside mainland China no 262144 65536 0.25 2.00 0.00 budgets, none to high, default off -
qwen3.5-27b Qwen Outside mainland China no 262144 65536 0.30 2.40 0.00 budgets, none to high, default off -
qwen3.5-122b-a10b Qwen Outside mainland China no 262144 65536 0.40 3.20 0.00 budgets, none to high, default off -
qwen3-vl-235b-a22b-thinking Qwen Outside mainland China no 131072 32768 0.40 4.00 0.00 always on -
qwen3-vl-235b-a22b-instruct Qwen Outside mainland China no 262144 32768 0.40 1.60 0.00 no -
qwen3-next-80b-a3b-thinking Qwen Outside mainland China no 262144 32768 0.15 1.20 0.00 always on -
qwen3-next-80b-a3b-instruct Qwen Outside mainland China no 262144 32768 0.15 1.20 0.00 no -
qwen3-30b-a3b-thinking-2507 Qwen Outside mainland China no 262144 32768 0.20 2.40 0.00 always on -
qwen3-30b-a3b-instruct-2507 Qwen Outside mainland China no 262144 32768 0.20 0.80 0.00 no -
qwen3-235b-a22b-thinking-2507 Qwen Outside mainland China no 262144 131072 0.23 2.30 0.00 always on -
qwen3-235b-a22b-instruct-2507 Qwen Outside mainland China no 262144 131072 0.23 0.92 0.00 no -
deepseek-v4-flash
DeepSeek V4 Flash, processed outside mainland China
DeepSeek Outside mainland China no 1M 393216 0.20 0.40 0.04 on or off, default off -
qwen3.7-plus@alibaba-us
Qwen3.7 Plus, processed in the US
Qwen US (Virginia) no 1M 131072 0.32 1.28 0.064 budgets, none to high, default off -
qwen3.6-flash@alibaba-us
Qwen3.6 Flash, processed in the US
Qwen US (Virginia) no 1M 65536 0.25 1.50 0.00 budgets, none to high, default off -
deepseek-v4-flash-0731@alibaba-intl DeepSeek Outside mainland China no 131072 32768 0.44 1.32 0.088 no 0.22 / 0.66
beam-1.0-501b-a23b Reflection AI North America no 131072 131072 0.40 1.20 0.04 no -
deepseek-v4.1-flash-1017 DeepSeek North America no 131072 131072 0.30 1.20 0.006 no -
minimax-m3.1-flash MiniMax North America no 1048576 131072 0.15 0.60 0.015 no -
qwen3.5-9b@deepinfra Qwen Not guaranteed yes 262144 81920 0.10 0.15 0.00 levels, none to high, default off -
deepseek-v4.1-flash-1017@deepinfra DeepSeek Not guaranteed no 1048576 131072 0.098 0.294 0.003 levels, none to high, default off -
qwen3.6-35b-a3b@deepinfra Qwen Not guaranteed no 262144 65536 0.10 0.95 0.10 levels, none to high, default off -
deepseek-v4-flash@deepinfra DeepSeek Not guaranteed no 1048576 65536 0.09 0.18 0.018 levels, none to high, default off -
qwen3-235b-a22b-instruct-2507@deepinfra Qwen Not guaranteed no 262144 65536 0.09 0.55 0.00 no -
deepseek-v4-flash-0731@deepinfra DeepSeek Not guaranteed no 1048576 384K 0.06 0.18 0.015 levels, none to high, default off -
qwen3.5-35b-a3b@deepinfra Qwen Not guaranteed no 262144 81920 0.14 1.00 0.05 levels, none to high, default off -
qwen3-vl-235b-a22b-instruct@deepinfra Qwen Not guaranteed no 262144 65536 0.20 0.88 0.11 no -
qwen3-next-80b-a3b-instruct@deepinfra Qwen Not guaranteed no 262144 65536 0.09 1.10 0.00 no -
minimax-m3@deepinfra MiniMax Not guaranteed no 524288 131072 0.28 1.10 0.056 always on -
deepseek-v3.2 DeepSeek Not guaranteed no 131072 16384 0.26 0.38 0.13 no -
deepseek-v4-flash-vision-exp DeepSeek Not guaranteed yes 131072 16384 0.216 0.647 0.007 no -
gemma-3-12b-it Google Not guaranteed yes 131072 16384 0.05 0.15 0.00 no -
gemma-3-27b-it Google Not guaranteed yes 131072 16384 0.08 0.16 0.00 no -
gemma-3-4b-it Google Not guaranteed yes 131072 16384 0.05 0.10 0.00 no -
gemma-4-26b-a4b-it Google Not guaranteed yes 131072 16384 0.07 0.34 0.00 no -
gemma-4-31b-it Google Not guaranteed yes 131072 16384 0.20 0.40 0.00 no -
granite-4.2-8b IBM Not guaranteed no 131072 16384 0.06 0.25 0.015 no -
ling-3.0-flash inclusionAI Not guaranteed no 131072 16384 0.06 0.18 0.012 no -
ling-3.0-flash-vl inclusionAI Not guaranteed no 131072 16384 0.06 0.18 0.012 no -
ling-3.1-flash inclusionAI Not guaranteed no 131072 16384 0.045 0.135 0.009 no -
llama-guard-4-12b Meta Not guaranteed no 131072 16384 0.18 0.18 0.00 no -
muse-glimmer-30b@deepinfra Meta Not guaranteed no 131072 16384 0.30 1.20 0.04 no -
phi-4 Microsoft Not guaranteed no 131072 16384 0.07 0.14 0.00 no -
mistral-small-24b-instruct-2501 Mistral AI Not guaranteed yes 131072 16384 0.05 0.08 0.00 no -
mistral-small-3.2-24b-instruct-2506 Mistral AI Not guaranteed no 131072 16384 0.075 0.20 0.00 no -
gpt-oss-20b OpenAI Not guaranteed yes 131072 16384 0.03 0.14 0.015 no -
qwen3-14b Qwen Not guaranteed yes 131072 16384 0.12 0.24 0.00 no -
qwen3-30b-a3b Qwen Not guaranteed yes 131072 16384 0.12 0.50 0.00 no -
qwen3-32b Qwen Not guaranteed yes 131072 16384 0.08 0.28 0.00 no -
qwen3.5-27b@deepinfra Qwen Not guaranteed no 131072 16384 0.26 2.60 0.00 no -
qwen3-vl-30b-a3b-instruct Qwen Not guaranteed yes 131072 16384 0.15 0.60 0.00 no -
inkling-small Thinking Machines Not guaranteed yes 131072 16384 0.45 1.20 0.10 no -
mimo-v2.6-flash Xiaomi Not guaranteed no 131072 16384 0.14 0.28 0.003 no -
mimo-v2.6-pro Xiaomi Not guaranteed no 131072 16384 0.387 0.783 0.003 no -
glm-4.7 Z.ai Not guaranteed no 131072 16384 0.40 1.75 0.08 no -
glm-5.3-flash@deepinfra Z.ai Not guaranteed no 131072 16384 0.15 0.50 0.03 no -
llama-4-scout-17b-16e-instruct Meta Not guaranteed yes 131072 16384 0.10 0.30 0.00 no -
qwen3.5-9b@parasail Qwen Not guaranteed yes 262144 262144 0.10 0.25 0.00 on or off, default off -
qwen3.6-35b-a3b@parasail Qwen Not guaranteed no 262144 262144 0.10 0.90 0.025 on or off, default off -
qwen3.5-35b-a3b@parasail Qwen Not guaranteed no 262144 262144 0.15 1.00 0.05 on or off, default off -
qwen3-235b-a22b-instruct-2507@parasail Qwen Not guaranteed no 131072 131072 0.14 0.80 0.05 no -
qwen3-next-80b-a3b-instruct@parasail Qwen Not guaranteed no 262144 262144 0.10 1.10 0.07 no -
qwen3-vl-235b-a22b-instruct@parasail Qwen Not guaranteed no 131072 32768 0.21 1.90 0.10 no -
deepseek-v4-flash@parasail DeepSeek Not guaranteed no 1048576 1048576 0.14 0.28 0.07 on or off, default off -
deepseek-v4-flash-0731@parasail DeepSeek Not guaranteed no 1048576 1048576 0.14 0.28 0.05 on or off, default off -
deepseek-v4.1-flash-1017@parasail DeepSeek Not guaranteed no 1048576 1048576 0.30 1.20 0.006 on or off, default off -
minimax-m3@parasail MiniMax Not guaranteed no 1048576 524288 0.30 1.20 0.06 always on -
gpt-audio-mini OpenAI United States no 128K 16384 0.60 2.40 0.00 no -
ripple-65b-preview Ricursive Intelligence North America no 131072 131072 0.20 0.60 0.02 no -
qwen-plus Qwen Outside mainland China yes 131072 32768 0.40 1.20 0.00 no -
qwen-plus-2025-01-25 Qwen Outside mainland China yes 131072 32768 0.40 1.20 0.00 no -
minimax-m3.1-flash@parasail MiniMax Not guaranteed no 1048576 524288 0.15 0.60 0.015 always on -
minimax-m3-fast@parasail MiniMax Not guaranteed no 1048576 524288 0.45 1.80 0.09 always on -
inkling-2-small Thinking Machines North America yes 131072 131072 0.30 1.00 0.03 no -
inkling-2-small@deepinfra Thinking Machines Not guaranteed no 131072 16384 0.285 0.95 0.029 no -
qwen3.5-omni-flash Qwen Outside mainland China no 196608 65536 0.40 2.20 0.00 no -
qwen3.5-omni-flash-2026-03-15 Qwen Outside mainland China no 196608 65536 0.40 2.20 0.00 no -
deepseek-v4.1-flash-1017@alibaba-intl
DeepSeek V4.1 Flash 1017, processed outside mainland China
DeepSeek Outside mainland China no 1M 393216 0.30 1.20 0.006 on or off, default off 0.15 / 0.60
Model Maker Processed in Auto Context window Max output Input Output Cache read Reasoning Off-peak input / output
claude-sonnet-5 Anthropic United States yes 1M 128K 2.00 10.00 0.20 levels, none to xhigh, default the vendor’s -
claude-sonnet-4-6 Anthropic United States no 1M 128K 3.00 15.00 0.30 levels, none to xhigh, default off -
gpt-5.4-mini OpenAI United States yes 400K 128K 0.75 4.50 0.075 levels, none to xhigh, default off -
gpt-5.4 OpenAI United States yes 400K 128K 2.50 15.00 0.25 levels, none to xhigh, default off -
gpt-4.1 OpenAI United States no 1M 32768 2.00 8.00 0.50 no -
gemini-3.8-flash Google United States yes 1M 65536 0.75 3.75 0.075 levels, low to high, default the vendor’s, always reasons -
gemini-3.5-flash Google United States no 1M 65536 0.75 3.75 0.075 levels, none to high, default off -
deepseek-flash
DeepSeek (China)
DeepSeek China yes 1048576 393216 0.30 1.20 0.006 levels, none to xhigh, default off 0.15 / 0.60
mistral-large-2512 Mistral AI Not guaranteed yes 256K 256K 0.50 1.50 0.05 no -
codestral-2508 Mistral AI Not guaranteed yes 128K 128K 0.30 0.90 0.03 no -
gpt-oss-120b OpenAI North America yes 131072 131072 0.15 0.60 0.00 always on -
deepseek-flash@together
DeepSeek V4.1 Flash, processed in North America
DeepSeek North America yes 1M 1M 0.30 1.20 0.006 on or off, default off -
gpt-5.2 OpenAI United States no 400K 128K 1.75 14.00 0.175 levels, none to xhigh, default off -
gpt-5.1 OpenAI United States no 400K 128K 1.25 10.00 0.125 levels, none to high, default off -
gpt-5 OpenAI United States no 400K 128K 1.25 10.00 0.125 levels, minimal to high, default the vendor’s, always reasons -
o4-mini OpenAI United States no 200K 100K 1.10 4.40 0.275 levels, low to high, default the vendor’s, always reasons -
o3-mini OpenAI United States no 200K 100K 1.10 4.40 0.55 levels, low to high, default the vendor’s, always reasons -
gpt-3.5-turbo OpenAI United States no 16385 4096 0.50 1.50 0.00 no -
claude-sonnet-5-5 Anthropic United States no 1M 128K 2.00 10.00 0.10 levels, low to xhigh, default the vendor’s, always reasons -
claude-sonnet-4-5 Anthropic United States no 200K 64K 3.00 15.00 0.30 no -
gemini-3.7-flash Google United States no 1M 65536 0.75 3.75 0.075 levels, low to high, default the vendor’s, always reasons -
gemini-3.6-flash Google United States no 1M 65536 0.75 3.75 0.075 levels, none to high, default off -
gemini-3-flash-preview Google United States no 1M 65536 0.50 3.00 0.05 levels, none to high, default off -
glm-5.3 Z.ai North America no 1048575 1048575 1.40 4.40 0.26 always on -
inkling Thinking Machines North America no 524288 524288 1.00 4.05 0.17 always on -
glm-5.2 Z.ai North America no 1048575 1048575 1.40 4.40 0.26 on or off, default off -
mistral-large-4-0 Mistral AI Not guaranteed no 524288 524288 0.70 2.10 0.07 on or off, default off -
qwen3.8-27b Qwen Outside mainland China no 1M 131072 0.50 3.00 0.10 budgets, none to high, default off -
qwen3.6-max-preview Qwen Outside mainland China no 262144 65536 1.30 7.80 0.26 budgets, none to high, default off -
qwen3.6-27b Qwen Outside mainland China no 262144 65536 0.60 3.60 0.00 budgets, none to high, default off -
qwen3.6-plus Qwen Outside mainland China no 1M 65536 0.50 3.00 0.00 budgets, none to high, default off -
qwen3.5-397b-a17b Qwen Outside mainland China no 262144 65536 0.60 3.60 0.00 budgets, none to high, default off -
deepseek-flash@alibaba-intl
DeepSeek V4.1 Flash, processed outside mainland China
DeepSeek Outside mainland China no 1M 393216 0.30 1.20 0.06 on or off, default off 0.15 / 0.60
glm-5.3@alibaba-intl
GLM 5.3, processed outside mainland China
Z.ai Outside mainland China no 202752 131072 1.40 4.40 0.00 always on -
glm-5.2@alibaba-intl
GLM 5.2, processed outside mainland China
Z.ai Outside mainland China no 202752 131072 1.40 4.40 0.00 budgets, none to high, default off -
glm-5.2@alibaba-us
GLM 5.2, processed in the US
Z.ai US (Virginia) no 202752 131072 1.40 4.40 0.00 budgets, none to high, default off -
deepseek-v4-pro-0813 DeepSeek Outside mainland China no 131072 32768 1.32 3.96 0.264 no 0.66 / 1.98
gpt-6.1-terra OpenAI United States yes 8192 4096 1.00 5.00 0.05 no -
llama-3.3-70b-instruct-turbo@deepinfra Meta Not guaranteed yes 131072 16384 0.10 0.32 0.00 no -
deepseek-flash@deepinfra DeepSeek Not guaranteed yes 1048576 131072 0.14 0.42 0.004 levels, none to high, default off -
gpt-oss-120b@deepinfra OpenAI Not guaranteed yes 131072 131072 0.037 0.17 0.018 always on -
qwen3.6-27b@deepinfra Qwen Not guaranteed no 262144 81920 0.32 3.20 0.00 levels, none to high, default off -
qwen3.5-397b-a17b@deepinfra Qwen Not guaranteed no 262144 81920 0.45 3.00 0.22 levels, none to high, default off -
deepseek-r1-0528 DeepSeek Not guaranteed no 131072 16384 0.50 2.15 0.35 no -
deepseek-v4-pro-0813@deepinfra DeepSeek Not guaranteed no 131072 16384 1.30 2.60 0.10 no -
kimi-k2.6 Moonshot AI Not guaranteed no 131072 16384 0.75 3.50 0.15 no -
qwen3.8-27b@deepinfra Qwen Not guaranteed no 131072 16384 0.15 1.875 0.037 no -
hy4-preview Tencent Not guaranteed no 131072 16384 0.834 2.501 0.042 no -
glm-4.6 Z.ai Not guaranteed no 131072 16384 0.50 2.00 0.10 no -
glm-5.2@deepinfra Z.ai Not guaranteed no 131072 16384 0.562 1.80 0.105 no -
glm-5.3@deepinfra Z.ai Not guaranteed no 131072 16384 0.562 2.50 0.125 no -
llama-3.3-70b-instruct-turbo@parasail Meta Not guaranteed yes 131072 16384 0.22 0.50 0.11 no -
deepseek-flash@parasail DeepSeek Not guaranteed yes 1048576 1048576 0.30 1.20 0.006 on or off, default off -
qwen3.8-27b@parasail Qwen Not guaranteed no 262144 262144 0.24 2.20 0.05 on or off, default off -
qwen3.5-397b-a17b@parasail Qwen Not guaranteed no 262144 262144 0.50 3.60 0.30 on or off, default off -
glm-5.3@parasail Z.ai Not guaranteed no 1048576 1048576 1.40 4.40 0.26 always on -
gpt-audio-1.5 OpenAI United States no 128K 16384 2.50 10.00 0.00 no -
gpt-audio OpenAI United States no 128K 16384 2.50 10.00 0.00 no -
glm-5.1 Z.ai Outside mainland China yes 131072 32768 1.40 4.40 0.00 no -
kimi-k2.7-code Moonshot AI Outside mainland China yes 131072 32768 0.95 4.00 0.00 no -
deepseek-flash-fast@parasail DeepSeek Not guaranteed no 1048576 1048576 0.45 1.80 0.009 on or off, default off -
glm-5.3-fast@parasail Z.ai Not guaranteed no 1048576 1048576 2.10 6.60 0.26 always on -
gemini-4-hydrogen Google United States yes 200K 65536 1.50 7.50 0.00 no -
qwen3.5-omni-plus Qwen Outside mainland China no 196608 65536 1.40 8.30 0.00 no -
qwen3.5-omni-plus-2026-03-15 Qwen Outside mainland China no 196608 65536 1.40 8.30 0.00 no -
hy4 Tencent North America no 131072 131072 0.834 2.501 0.042 no -
Model Maker Processed in Auto Context window Max output Input Output Cache read Reasoning Off-peak input / output
claude-opus-5-5 Anthropic United States yes 1M 128K 4.00 20.00 0.20 levels, low to xhigh, default the vendor’s, always reasons -
claude-opus-5 Anthropic United States yes 1M 128K 5.00 25.00 0.50 levels, low to xhigh, default the vendor’s, always reasons -
claude-opus-4-8 Anthropic United States no 1M 128K 5.00 25.00 0.50 levels, none to xhigh, default off -
gpt-5.5 OpenAI United States yes 400K 128K 5.00 30.00 0.50 levels, none to xhigh, default the vendor’s -
gemini-3.1-pro-preview Google United States yes 1048576 65536 2.00 12.00 0.20 levels, low to high, default the vendor’s, always reasons -
deepseek-v4-pro
DeepSeek (China)
DeepSeek China yes 1M 393216 1.32 3.96 0.044 levels, none to xhigh, default off 0.66 / 1.98
grok-4.7 xAI United States yes 500K 500K 2.20 6.60 0.55 always on -
mistral-medium-3-5 Mistral AI Not guaranteed yes 256K 256K 1.50 7.50 0.15 no -
deepseek-v4-pro@together
DeepSeek V4 Pro, processed in North America
DeepSeek North America yes 1048576 1048576 1.32 3.96 0.13 on or off, default off -
gpt-6.1-sol OpenAI United States no 400K 128K 2.00 10.00 0.10 levels, low to xhigh, default the vendor’s, always reasons -
gpt-6-sol OpenAI United States no 400K 128K 2.00 10.00 0.20 levels, none to xhigh, default the vendor’s -
gpt-5.6-sol OpenAI United States no 400K 128K 4.00 20.00 0.40 levels, none to xhigh, default the vendor’s -
gpt-5.6-terra OpenAI United States no 400K 128K 2.00 12.00 0.20 levels, none to xhigh, default the vendor’s -
o3 OpenAI United States no 200K 100K 2.00 8.00 0.50 levels, low to high, default the vendor’s, always reasons -
gpt-4o OpenAI United States no 128K 16384 2.50 10.00 1.25 no -
chat-latest OpenAI United States no 400K 128K 5.00 30.00 0.50 no -
claude-opus-4-7 Anthropic United States no 1M 128K 5.00 25.00 0.50 levels, none to xhigh, default off -
claude-opus-4-6 Anthropic United States no 1M 128K 5.00 25.00 0.50 levels, none to xhigh, default off -
claude-opus-4-5 Anthropic United States no 200K 64K 5.00 25.00 0.50 no -
grok-4.6 xAI United States no 500K 500K 2.20 6.60 0.55 always on -
qwen3.8-2.4t-a95b Qwen North America no 1010K 1010K 2.00 6.00 0.25 always on -
kimi-k3 Moonshot AI North America no 1048576 1048576 3.00 15.00 0.30 on or off, default off -
qwen3.8-max-0902 Qwen Outside mainland China no 1M 131072 2.00 6.00 0.40 budgets, none to high, default off -
qwen3.8-2.4t-a95b@alibaba-intl
Qwen3.8 2.4T, processed outside mainland China
Qwen Outside mainland China no 1M 131072 2.00 6.00 0.40 always on -
qwen3.7-max Qwen Outside mainland China no 1M 131072 2.50 7.50 0.50 budgets, none to high, default off -
qwen3.7-max@alibaba-us
Qwen3.7 Max, processed in the US
Qwen US (Virginia) no 1M 131072 2.50 7.50 0.50 budgets, none to high, default off -
kimi-k3.1-pro Moonshot AI Not guaranteed yes 1048576 1048576 2.00 10.00 0.20 levels, low to xhigh, default the vendor’s, always reasons -
kimi-k3.1-pro-highspeed Moonshot AI Not guaranteed yes 1048576 1048576 4.00 20.00 0.20 levels, low to xhigh, default the vendor’s, always reasons -
kimi-k3@moonshot Moonshot AI Not guaranteed no 1048576 1048576 3.00 15.00 0.30 levels, low to xhigh, default the vendor’s, always reasons -
deepseek-v4-pro@deepinfra DeepSeek Not guaranteed yes 1048576 65536 1.30 2.60 0.10 levels, none to high, default off -
kimi-k3@deepinfra Moonshot AI Not guaranteed no 131072 16384 2.85 14.25 0.285 no -
deepseek-v4-pro@parasail DeepSeek Not guaranteed yes 1048576 1048576 1.32 3.96 0.044 on or off, default off -
kimi-k3.1-pro@together Moonshot AI North America yes 131072 131072 2.00 10.00 0.20 no -
kimi-k3.1-pro@deepinfra Moonshot AI Not guaranteed yes 131072 16384 1.90 9.50 0.19 no -
glm-5.3-prime Z.ai Outside mainland China no 131072 32768 2.80 8.80 0.00 no -
claude-fable-5-2 Anthropic United States yes 1M 128K 8.00 40.00 0.20 levels, low to xhigh, default the vendor’s, always reasons -
deepseek-v4-pro-fast@parasail DeepSeek Not guaranteed no 1048576 1048576 1.98 5.94 0.066 on or off, default off -
gemini-4-carbon Google United States yes 200K 65536 2.00 10.00 0.00 no -
gemini-4-argon Google United States yes 200K 65536 2.00 10.00 0.00 no -
gpt-6.2-astra OpenAI United States no 8192 4096 8.00 40.00 0.40 no -
gpt-6.2-sol OpenAI United States no 8192 4096 2.00 10.00 0.10 no -
gemini-4-neon Google United States no 200K 65536 3.00 15.00 0.00 no -
qwen4-max Qwen Outside mainland China no 1M 131072 1.50 4.50 0.075 budgets, none to high, default off -
kimi-k3.1-pro@alibaba-intl
Kimi K3.1 Pro, processed outside mainland China
Moonshot AI Outside mainland China no 1048576 1048576 2.00 10.00 0.20 on or off, default the vendor’s -
Model Maker Processed in Auto Context window Max output Input Output Cache read Reasoning
claude-fable-5-1 Anthropic United States yes 1M 128K 10.00 50.00 0.25 levels, low to xhigh, default the vendor’s, always reasons
gpt-6-astra OpenAI United States yes 400K 128K 10.00 50.00 1.00 levels, low to xhigh, default the vendor’s, always reasons
o1 OpenAI United States no 200K 100K 15.00 60.00 7.50 levels, low to high, default the vendor’s, always reasons
gpt-4-turbo OpenAI United States no 128K 4096 10.00 30.00 0.00 no
gpt-4 OpenAI United States no 8192 8192 30.00 60.00 0.00 no
claude-fable-5 Anthropic United States no 1M 128K 10.00 50.00 1.00 levels, low to xhigh, default the vendor’s, always reasons
Model Maker Processed in Max input tokens Input Output
jev-1.13.0 TypeSafe AI United States 32K 0.042 0.00
jev-1.14.0 TypeSafe AI United States 32K 0.042 0.00

Decision models answer typed questions (fixed choices, yes or no, scores) with calibrated probabilities instead of writing text. They are never chosen by auto; call them by name with response_format: json_schema. Output tokens are free. See Decision models.

Model Maker Processed in Price per minute Formats Longest input
gpt-transcribe OpenAI United States 0.0045 json, text, stream -
whisper-1 OpenAI United States 0.006 json, text, verbose_json, srt, vtt, segment, word -
gpt-4o-transcribe OpenAI United States 2.50 / 1M audio tokens (0.006*) json, text, stream -
gpt-4o-mini-transcribe OpenAI United States 1.25 / 1M audio tokens (0.003*) json, text, stream -
whisper-large-v3 OpenAI North America 0.0015 json, text, verbose_json, srt, vtt, segment, word 240 min
voxtral-mini-3b-2507 Mistral AI Not guaranteed 0.001 json, text, verbose_json 30 min
voxtral-small-24b-2507 Mistral AI Not guaranteed 0.003 json, text, verbose_json 30 min
gemini-3.5-transcribe Google United States 2.00 / 1M audio tokens (0.003*) json, text, verbose_json, srt, vtt, segment, word 30 min
scribe-v2 ElevenLabs United States 0.0037 json, text, verbose_json, srt, vtt, segment, word 600 min
nova-3 Deepgram Not guaranteed 0.0043 json, text, verbose_json, srt, vtt, segment, word -
nova-3-multilingual Deepgram Not guaranteed 0.0052 json, text, verbose_json, srt, vtt, segment, word -
scribe-v2-medical ElevenLabs United States 0.0037 json, text, verbose_json, srt, vtt, segment, word 600 min
whisper-large Deepgram Not guaranteed 0.0048 json, text, verbose_json, srt, vtt, segment, word -

Transcription models are called through POST /v1/audio/transcriptions, by name or as auto, which picks the cheapest model able to do what the request asks. The price is per minute of audio where the vendor bills minutes; a model billed in audio tokens shows its token price and, marked with an asterisk, the vendor’s own per-minute estimate at its stated tokens per second, which is never what a call is billed. See Audio.

Model Maker Processed in Audio in Audio out Formats Voices
gpt-audio-1.5 OpenAI United States 32.00 / 1M 64.00 / 1M wav, mp3, flac, opus, pcm16, aac alloy, ash, ballad, coral, echo, sage, shimmer, marin, cedar, fable, nova, onyx, verse
gpt-audio OpenAI United States 32.00 / 1M 64.00 / 1M wav, mp3, flac, opus, pcm16, aac alloy, ash, ballad, coral, echo, sage, shimmer, marin, cedar, fable, nova, onyx, verse
gpt-audio-mini OpenAI United States 10.00 / 1M 20.00 / 1M wav, mp3, flac, opus, pcm16, aac alloy, ash, ballad, coral, echo, sage, shimmer, marin, cedar, fable, nova, onyx, verse
qwen3.5-omni-plus Qwen Outside mainland China - 44.00 / 1M, text free pcm16, streamed only Tina, Cindy, Liora Mira, Raymond, Katerina, Ryan, Mia, Theo Calm, Serena, Maia, Evan, Qiao, Momo, Wil, Angel, Li Cassian, Joyner, Gold, Jennifer, Aiden, Mione, Sunny, Dylan, Eric, Peter, Joseph Chen, Marcus, Li, Rocky, Kiki, Sohee, Eliška, Alek, Arda, Dolce, Lenn, Ono Anna, Sonrisa, Bodega, Andre, Radio Gol, Rizky, Roya, Hana, Jakub, Griet, Marina, Siiri, Ingrid, Sigga, Bea, Chloe, Emilien, Sunnybobi, Ethan, Harvey
qwen3.5-omni-plus-2026-03-15 Qwen Outside mainland China - 44.00 / 1M, text free pcm16, streamed only Tina, Cindy, Liora Mira, Raymond, Katerina, Ryan, Mia, Theo Calm, Serena, Maia, Evan, Qiao, Momo, Wil, Angel, Li Cassian, Joyner, Gold, Jennifer, Aiden, Mione, Sunny, Dylan, Eric, Peter, Joseph Chen, Marcus, Li, Rocky, Kiki, Sohee, Eliška, Alek, Arda, Dolce, Lenn, Ono Anna, Sonrisa, Bodega, Andre, Radio Gol, Rizky, Roya, Hana, Jakub, Griet, Marina, Siiri, Ingrid, Sigga, Bea, Chloe, Emilien, Sunnybobi, Ethan, Harvey
qwen3.5-omni-flash Qwen Outside mainland China - 11.90 / 1M, text free pcm16, streamed only Tina, Cindy, Liora Mira, Raymond, Katerina, Ryan, Mia, Theo Calm, Serena, Maia, Evan, Qiao, Momo, Wil, Angel, Li Cassian, Joyner, Gold, Jennifer, Aiden, Mione, Sunny, Dylan, Eric, Peter, Joseph Chen, Marcus, Li, Rocky, Kiki, Sohee, Eliška, Alek, Arda, Dolce, Lenn, Ono Anna, Sonrisa, Bodega, Andre, Radio Gol, Rizky, Roya, Hana, Jakub, Griet, Marina, Siiri, Ingrid, Sigga, Bea, Chloe, Emilien, Sunnybobi, Ethan, Harvey

These chat models answer in audio when a request to POST /v1/chat/completions sends modalities: ["text", "audio"] and audio: {"voice", "format"}; call them by name, auto does not answer in audio. They stream an answer in pcm16 only; the other formats come in one piece. A model marked “streamed only” answers in audio only when the request sends stream: true. The text of the answer is billed at the model’s output price in the tables above and the audio at the price per million audio tokens here; on a model marked “text free”, an answer that carries audio bills its audio alone. Voices are AI-generated. See Chat completions.

Model Maker Processed in Price Formats Voices Longest input
eleven-v4 ElevenLabs United States 80.00 / 1M chars mp3, opus, wav, pcm 19 10000 chars
eleven-v4-turbo ElevenLabs United States 40.00 / 1M chars mp3, opus, wav, pcm 19 -
gemini-3.8-flash-tts Google United States 0.50 / 1M text tokens in, 9.00 / 1M audio tokens out (0.0022 / 10 s*) wav, pcm 30 -
gemini-3.8-flash-lite-tts Google United States 0.50 / 1M text tokens in, 6.00 / 1M audio tokens out (0.0015 / 10 s*) wav, pcm 30 -
aura-2 Deepgram Not guaranteed 30.00 / 1M chars mp3, opus, flac, wav, pcm 90 2000 chars
kokoro-82m Hexgrad North America 4.00 / 1M chars mp3, wav, pcm 54 -
gpt-4o-mini-tts OpenAI United States 0.60 / 1M text tokens in, 12.00 / 1M audio tokens out mp3, opus, aac, flac, wav, pcm 13 4096 chars
tts-1 OpenAI United States 15.00 / 1M chars mp3, opus, aac, flac, wav, pcm 9 4096 chars
tts-1-hd OpenAI United States 30.00 / 1M chars mp3, opus, aac, flac, wav, pcm 9 4096 chars
aura-1 Deepgram Not guaranteed 15.00 / 1M chars mp3, opus, flac, wav, pcm 12 2000 chars
eleven-flash-v2 ElevenLabs United States 40.00 / 1M chars mp3, opus, wav, pcm 19 30000 chars
eleven-flash-v2.5 ElevenLabs United States 40.00 / 1M chars mp3, opus, wav, pcm 19 40000 chars
eleven-multilingual-v2 ElevenLabs United States 80.00 / 1M chars mp3, opus, wav, pcm 19 10000 chars
eleven-v3 ElevenLabs United States 80.00 / 1M chars mp3, opus, wav, pcm 19 5000 chars
eleven-v3-conversational ElevenLabs United States 40.00 / 1M chars mp3, opus, wav, pcm 19 -
flux-tts Deepgram Not guaranteed 45.00 / 1M chars mp3, opus, flac, wav, pcm 266 -

Speech models are called through POST /v1/audio/speech, by name or as auto, which picks the cheapest model that offers the asked format and voice. The price is per million characters of input text where the vendor bills characters; a model billed in audio tokens shows its token prices and, marked with an asterisk, the vendor’s own price per 10 seconds of audio at its stated tokens per second, which is never what a call is billed. Each model returns the formats its vendor produces; nothing is transcoded. Voices are AI-generated. See Audio.

Speaks English only: aura-1, eleven-flash-v2. A speech request does not say its language, so auto never picks these models; call them by name.

Model Maker Processed in Price Sizes Qualities Formats
gpt-image-2.5-flare OpenAI United States 5.00 / 1M text tokens in, 8.00 / 1M image tokens in, 30.00 / 1M image tokens out (0.211 / 1024x1024*) 1024x1024, 1536x1024, 1024x1536; any WIDTHxHEIGHT within the vendor’s limits low, medium, high, xhigh, max png, jpeg, webp
gpt-image-2.5-sunburst OpenAI United States 5.00 / 1M text tokens in, 8.00 / 1M image tokens in, 30.00 / 1M image tokens out (0.211 / 1024x1024*) 1024x1024, 1536x1024, 1024x1536; any WIDTHxHEIGHT within the vendor’s limits low, medium, high, xhigh, max png, jpeg, webp
gpt-image-2 OpenAI United States 5.00 / 1M text tokens in, 8.00 / 1M image tokens in, 30.00 / 1M image tokens out (0.211 / 1024x1024*) 1024x1024, 1536x1024, 1024x1536; any WIDTHxHEIGHT within the vendor’s limits low, medium, high png, jpeg, webp
gpt-image-1-mini OpenAI United States 2.00 / 1M text tokens in, 2.50 / 1M image tokens in, 8.00 / 1M image tokens out (0.036 / 1024x1024*) 1024x1024, 1536x1024, 1024x1536 low, medium, high png, jpeg, webp
gemini-nano-banana-2.1 Google United States 1.50 / 1M text tokens in, 1.50 / 1M image tokens in, 30.00 / 1M image tokens out (0.0336 / 1024x1024*) 42 sizes from 1024x1024 to 6336x2688 - jpeg
gemini-3.1-flash-lite-image Google United States 0.25 / 1M text tokens in, 0.25 / 1M image tokens in, 30.00 / 1M image tokens out (0.0336 / 1024x1024*) 14 sizes from 1024x1024 to 1584x672 - jpeg
gemini-3-pro-image Google United States 2.00 / 1M text tokens in, 2.00 / 1M image tokens in, 120.00 / 1M image tokens out (0.1344 / 1024x1024*) 30 sizes from 1024x1024 to 6336x2688 - jpeg
flux-2-dev@deepinfra Black Forest Labs Not guaranteed 0.0179 / 1024x1024 1024x1024, 1536x1024, 1024x1536 - jpeg
grok-imagine-image xAI Not guaranteed 0.02 / 1024x1024, 0.002 / input image on an edit 1024x1024, 1248x832, 832x1248 - jpeg
grok-imagine-image-2.0 xAI Not guaranteed 0.04 / 1024x1024, 0.06 / 1024x1024 medium, 0.01 / input image on an edit 1024x1024, 1248x832, 832x1248 low, medium jpeg

Image models are called through POST /v1/images/generations and POST /v1/images/edits, by name or as auto, which picks the cheapest model that produces the asked size, quality and format by the vendors’ own price figures. A model billed in tokens shows its text and image token prices and, marked with an asterisk, the vendor’s own price for one 1024x1024 image (at high quality where the model has quality levels), which is never what a call is billed: a call is billed on the tokens the vendor reports. A model billed per image shows its price for one image of the size named, which is what a call is billed. Google’s images carry a SynthID watermark. See Images.

Model Maker Processed in Price Lengths Resolutions Starts from Sound
grok-imagine-video-1.5 xAI Not guaranteed 0.08 / s, 0.01 / image 1 to 15 s 480p, 720p, 1080p a prompt, an image on, can be turned off
grok-imagine-video-1.5-lite xAI Not guaranteed 0.02 / s, 0.01 / image 1 to 15 s 480p, 720p, 1080p a prompt, an image on, can be turned off
wan2.6-t2v Alibaba US (Virginia) 0.1 / s at 720p, 0.15 / s at 1080p 2 to 15 s 720p, 1080p a prompt always
wan2.6-i2v Alibaba US (Virginia) 0.1 / s at 720p, 0.15 / s at 1080p 2 to 15 s 720p, 1080p an image always
gemini-omni-1.1-flash Google United States 17.50 / 1M video tokens (0.1014 / s at 720p*) 3 to 10 s 360p, 720p, 1080p, 4k a prompt, an image always

Video models are called through POST /v1/videos, by name or as auto, which picks the cheapest model that makes the asked length and resolution from the asked input by the vendors’ own prices. A video is made as a job: the request answers at once with a video id, and GET /v1/videos/{id} says when it is ready. A video is billed per second of the video made, at the price of the resolution made; a model billed in video tokens (Gemini Omni) shows, marked with an asterisk, the vendor’s own price of a second at 720p, an estimate: its call is billed on the tokens the vendor reports. See Videos.

Model Maker Processed in Max input tokens Input
text-embedding-3-small OpenAI United States 8192 0.02
text-embedding-3-large OpenAI United States 8192 0.13
gemini-embedding-001 Google United States 8192 0.15
mistral-embed-2312 Mistral AI Not guaranteed 8K 0.10

Off-peak: DeepSeek bills half its list price outside peak hours. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, except Chinese public holidays; weekends and those holidays are off-peak all day. A call takes the off-peak price only when both its arrival and its completion fall inside the off-peak hours; otherwise it pays the list price.

Off-peak: Alibaba Cloud (International) bills half its list price outside peak hours. Peak hours are 00:00 to 14:00 UTC, Monday to Friday, except Chinese public holidays; weekends and those holidays are off-peak all day. A call takes the off-peak price only when both its arrival and its completion fall inside the off-peak hours; otherwise it pays the list price.

Long context: when the prompt of a gpt-5.4 request exceeds 272K tokens, every token of that request is billed at 5.00 input, 22.50 output and 0.50 cache read.

Long context: when the prompt of a gpt-5.5 request exceeds 272K tokens, every token of that request is billed at 10.00 input, 45.00 output and 1.00 cache read.

Long context: when the prompt of a gpt-6-astra request exceeds 272K tokens, every token of that request is billed at 20.00 input, 75.00 output and 2.00 cache read.

Long context: when the prompt of a gemini-3.1-pro-preview request exceeds 200K tokens, every token of that request is billed at 4.00 input, 18.00 output and 0.40 cache read.

Long context: when the prompt of a grok-4.7 request reaches 200K tokens, every token of that request is billed at 4.40 input, 13.20 output and 1.10 cache read.

Long context: when the prompt of a gpt-6.1-sol request exceeds 272K tokens, every token of that request is billed at 4.00 input, 15.00 output and 0.20 cache read.

Long context: when the prompt of a gpt-6-luna request exceeds 272K tokens, every token of that request is billed at 0.20 input, 0.75 output and 0.02 cache read.

Long context: when the prompt of a gpt-6-sol request exceeds 272K tokens, every token of that request is billed at 4.00 input, 15.00 output and 0.40 cache read.

Long context: when the prompt of a gpt-5.6-sol request exceeds 272K tokens, every token of that request is billed at 8.00 input, 30.00 output and 0.80 cache read.

Long context: when the prompt of a gpt-5.6-terra request exceeds 272K tokens, every token of that request is billed at 4.00 input, 18.00 output and 0.40 cache read.

Long context: when the prompt of a gpt-5.6-luna request exceeds 272K tokens, every token of that request is billed at 0.40 input, 1.80 output and 0.04 cache read.

Long context: when the prompt of a grok-4.6 request reaches 200K tokens, every token of that request is billed at 4.40 input, 13.20 output and 1.10 cache read.

Long context: when the prompt of a claude-haiku-5-5 request exceeds 100K tokens, every token of that request is billed at 0.50 input, 2.50 output and 0.05 cache read.

Long context: when the prompt of a qwen3.7-plus request exceeds 256K tokens, every token of that request is billed at 0.96 input, 3.84 output and 0.192 cache read.

Long context: when the prompt of a qwen3.5-plus-2026-04-20 request exceeds 256K tokens, every token of that request is billed at 0.50 input, 3.00 output and 0.00 cache read.

Long context: when the prompt of a qwen3.5-plus-2026-02-15 request exceeds 256K tokens, every token of that request is billed at 0.50 input, 3.00 output and 0.00 cache read.

Long context: when the prompt of a qwen3.6-flash request exceeds 256K tokens, every token of that request is billed at 1.00 input, 4.00 output and 0.20 cache read.

Long context: when the prompt of a qwen3.6-max-preview request exceeds 128K tokens, every token of that request is billed at 2.00 input, 12.00 output and 0.40 cache read.

Long context: when the prompt of a qwen3.6-plus request exceeds 256K tokens, every token of that request is billed at 2.00 input, 6.00 output and 0.00 cache read.

Long context: when the prompt of a qwen3.7-plus@alibaba-us request exceeds 256K tokens, every token of that request is billed at 0.96 input, 3.84 output and 0.192 cache read.

Long context: when the prompt of a qwen3.6-flash@alibaba-us request exceeds 256K tokens, every token of that request is billed at 1.00 input, 4.00 output and 0.00 cache read.

Long context: when the prompt of a gpt-6.1-terra request exceeds 272K tokens, every token of that request is billed at 2.00 input, 7.50 output and 0.10 cache read.

Long context: when the prompt of a qwen-plus request exceeds 256K tokens, every token of that request is billed at 1.20 input, 3.60 output and 0.00 cache read.

Long context: when the prompt of a gpt-6.2-astra request exceeds 272K tokens, every token of that request is billed at 16.00 input, 60.00 output and 0.80 cache read.

Long context: when the prompt of a gpt-6.2-sol request exceeds 272K tokens, every token of that request is billed at 4.00 input, 15.00 output and 0.20 cache read.

whisper-1 retires on 2027-02-26; auto stops choosing it on 2027-02-16. Requests that name it keep working until the vendor’s date.

gpt-4o-transcribe retires on 2027-02-26; auto stops choosing it on 2027-02-16. Requests that name it keep working until the vendor’s date.

gpt-4o-mini-transcribe retires on 2027-02-26; auto stops choosing it on 2027-02-16. Requests that name it keep working until the vendor’s date.

gpt-4o-mini-tts retires on 2027-01-06; auto stops choosing it on 2026-12-27. Requests that name it keep working until the vendor’s date.

tts-1 retires on 2027-01-06; auto stops choosing it on 2026-12-27. Requests that name it keep working until the vendor’s date.

tts-1-hd retires on 2027-01-06; auto stops choosing it on 2026-12-27. Requests that name it keep working until the vendor’s date.

gpt-audio retires on 2027-01-20; auto stops choosing it on 2027-01-10. Requests that name it keep working until the vendor’s date.

gpt-audio-mini retires on 2027-01-20; auto stops choosing it on 2027-01-10. Requests that name it keep working until the vendor’s date.

gpt-image-1-mini retires on 2026-12-01; auto stops choosing it on 2026-11-21. Requests that name it keep working until the vendor’s date.

Retired on 2026-11-14, no longer served by the vendor: o4-mini, o3-mini, o1, gpt-4.1-nano, gpt-4-turbo, gpt-4, gpt-3.5-turbo.

Retired on 2026-10-12, no longer served by the vendor: claude-sonnet-4-5.

Retired on 2026-10-11, no longer served by the vendor: qwen3.6-max-preview, qwen3-vl-235b-a22b-thinking, qwen3-vl-235b-a22b-instruct, 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, deepseek-v4-flash.

Reasoning: the column says how reasoning_effort maps onto each model (levels the vendor names, a thinking budget per level, on or off, always on, or no reasoning) and what Zabio applies when a request sets nothing; auto chooses the setting with the model. See Models.

Generated on 2026-11-20 from the launch catalog; do not edit.

Some vendors price by time or by request size:

  • DeepSeek off-peak. DeepSeek charges half its list price outside peak hours. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday; weekends and Chinese public holidays are off-peak all day. The table shows both prices, and zabio.pricing.off_peak in GET /v1/models carries the off-peak price and the peak hours. A call takes the off-peak price only when both its arrival and its completion fall inside the off-peak hours; otherwise it pays the list price. auto prefers DeepSeek more when it is off-peak. The copies served by Together AI have no off-peak price.
  • OpenAI and Google long context. gpt-5.4, gpt-5.5 and gpt-6-astra bill a request whose prompt exceeds 272K input tokens at twice the input and cache rates and 1.5 times the output rate, for the whole request. gemini-3.1-pro-preview bills a request whose prompt exceeds 200K tokens at 4.00 input, 18.00 output and 0.40 cache read. A prompt’s size is every input token of the request, cached or not, audio included: the sound of a video counts toward the 200K and is billed at the input rate, 2.00 up to the line and 4.00 above it. The tiers are shown under the table and in zabio.pricing.long_context, and auto prices every candidate at your prompt’s size, so a long prompt is steered to a model without a surcharge when one is as capable.
  • Sales. When a provider sells a model below its list price to everyone, the table shows the sale price and Zabio bills it. When the sale ends, the price returns to the list price.
  • Anthropic long context. claude-haiku-5-5 is priced by prompt length: a request whose prompt exceeds 100K tokens (fresh plus cached input) is billed at 0.50 input, 2.50 output and that band’s cache rates (0.05 to read, 0.625 for a 5 minute write, 1.00 for a 1 hour write), for the whole request. The tier is shown under the table and in zabio.pricing.long_context.
  • xAI long context. When the prompt of a grok-4.7 request (fresh plus cached input) reaches 200K tokens, every token of that request is billed at the long-context rate, shown under the table and in zabio.pricing.long_context. The prices are those of xAI’s US endpoint, 1.1 times xAI’s list price.

The same data is available as JSON at generated/models.json in the documentation source. Every row carries the model’s maker (maker, and maker_name as the table prints it), the list prices above and two objects, flex and batch, one per vendor processing tier. When the model has the tier, the object is {"applies": true, "input_per_mtok": 1.25, "output_per_mtok": 7.5, "cache_read_per_mtok": 0.13} with the prices Zabio bills on it, in USD per million tokens before the fee; when it does not, the object is {"applies": false} and carries no prices. Flex is the slower tier a zabio.deadline of 15 minutes or more asks for (see Flex tier); batch is the price of a line in a batch.

  • DeepSeek thinking. DeepSeek models think before answering unless told not to. When you send no reasoning_effort, Zabio turns thinking off; send a level to turn it on. DeepSeek knows none, low, high and max: none is off, xhigh is sent as max, and medium and minimal, which it does not have, become low and none. Requests that declare tools, or carry tool calls and results, always run with thinking off, because DeepSeek then requires the model’s earlier reasoning to be sent back, which Zabio does not keep. deepseek-flash accepts images; deepseek-v4-pro does not. Neither supports response_format: json_schema; json_object works.
  • grok-4.7 always reasons and bills its reasoning tokens as output tokens, which Zabio includes in usage.completion_tokens. xAI does not take stop on it, so Zabio applies your stop sequences itself; presence_penalty and frequency_penalty are never sent to any model. It refuses images smaller than 8 by 8 pixels or 512 pixels in total.
  • Mistral AI. Mistral’s chat models do not reason, except mistral-large-4-0: a reasoning_effort you send to the others is ignored, listed in zabio.ignored_parameters, and zabio.reasoning.applied says not supported. mistral-large-4-0 thinks before answering when asked. Zabio turns its thinking off when you send no reasoning_effort; it knows only off and on, so none turns thinking off and any other level turns it on (zabio.reasoning.applied says on). A conversation can move to a Mistral model mid-way: tool call ids from other vendors are accepted as they are.
  • Together AI. Some Together AI models think by default (qwen3.5-9b and the DeepSeek copies). Zabio turns thinking off unless you send reasoning_effort; a level you send is kept when the request carries tools (unlike DeepSeek’s own API, Together completes a tool round trip with thinking on); these models only know on and off, so none turns thinking off and any other level turns it on (zabio.reasoning.applied says on). gpt-oss-120b always reasons, whatever the level, so leave room in max_tokens for it.
  • DeepInfra. The models with ids ending in @deepinfra are the same models as their originals, served by DeepInfra at the precision shown on the models page and billed at DeepInfra’s list price plus 5 percent, with cache hits at DeepInfra’s cached rate where it has one; auto may pick a copy when it is the cheapest host of a model it already uses. The DeepSeek and Qwen copies think only when you send a reasoning_effort (none, low, medium or high; minimal becomes none and xhigh becomes high), and keep the level on a tool round trip; gpt-oss-120b@deepinfra and minimax-m3@deepinfra always reason. Reasoning tokens are inside usage.completion_tokens; the DeepSeek copies report the count and the others do not, so zabio.reasoning.tokens is an estimate on them. temperature, top_p and stop reach the model.
  • Parasail. The models with ids ending in @parasail are the same models as their originals, served by Parasail at the precision shown on the models page and billed at Parasail’s list price plus 5 percent, with cache hits at Parasail’s cached rate where it has one; auto may pick a copy when it is the cheapest host of a model it already uses, and a copy stands in when another host of the model is down. The DeepSeek and Qwen copies think by default on Parasail; Zabio turns thinking off unless you send a reasoning_effort, and these copies know only on and off, so none turns thinking off and any other level turns it on (zabio.reasoning.applied says on); a level you send is kept when the request carries tools. glm-5.3@parasail and minimax-m3@parasail always reason. Reasoning tokens are inside usage.completion_tokens; the Qwen copies and deepseek-flash@parasail report the count, the others do not, so zabio.reasoning.tokens is an estimate on them. temperature, top_p and stop reach the model.
  • Moonshot AI. Kimi K3.1 Pro and Kimi K3 always reason and report their reasoning tokens inside usage.completion_tokens. Moonshot takes reasoning_effort as low, high or max, with max as its default when you send nothing: xhigh is sent as max, medium and minimal become low, and none, which these models do not have, becomes low as well (zabio.reasoning.applied says what was sent). temperature and top_p are fixed by the vendor and never sent (see the table below). Moonshot caches repeated prompt prefixes on its own for 5 minutes and bills the written prefix at the input price, so a cache hit costs the cached rate and a miss costs nothing extra. On a tool round trip Moonshot asks for the model’s earlier reasoning to be sent back, which Zabio does not keep; the model answers without it.

Gemini models can spend output tokens on internal thinking before they answer. low, medium and high are Google’s own thinking levels, and xhigh maps to high. When you send no reasoning_effort or send none, Zabio turns thinking off where the model allows it (gemini-3.5-flash, gemini-3.6-flash, gemini-3-flash-preview) so your max_tokens goes to the answer; the flash-lite models do not think unless asked. Three models always think, and you should budget for it:

  • gemini-3.8-flash and gemini-3.7-flash think at Google’s medium when you set nothing, and none gives their lowest level, low.
  • gemini-3.1-pro-preview thinks at Google’s high when you set nothing and needs a few hundred output tokens even for a short answer.

A response with empty content and finish_reason: "length" means the budget was too small.

temperature, top_p and stop reach the model when it takes them. A model that does not take one never makes the request fail: Zabio leaves the parameter out and names it in zabio.dropped_parameters and the x-zabio-dropped-params header. stop is always honored: where the provider does not take it, Zabio cuts the answer at the first sequence itself, streamed or not. Every model was checked with each parameter on 2026-10-07, and each model added since on the day it was added.

Models temperature top_p stop
OpenAI gpt-5, gpt-5-mini, gpt-5-nano, gpt-5.5, gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-6-astra, gpt-6-luna, gpt-6-sol, gpt-6.1-sol, o3, o4-mini, chat-latest dropped dropped applied by Zabio
OpenAI o1, o3-mini dropped dropped yes
OpenAI gpt-5.4, gpt-5.4-mini, gpt-5.4-nano, gpt-5.2, gpt-5.1 dropped while reasoning dropped while reasoning applied by Zabio
Anthropic Claude 5 models, claude-opus-4-8, claude-opus-4-7 dropped dropped yes
Anthropic claude-haiku-4-5, claude-opus-4-5, claude-sonnet-4-5 yes dropped when temperature is also sent yes
Anthropic claude-opus-4-6, claude-sonnet-4-6 dropped while thinking dropped when temperature is also sent, and below 0.95 while thinking yes
Google Gemini models dropped (no effect on these models) dropped (no effect on these models) yes
xAI grok-4.7 and grok-4.6; Together AI gpt-oss-120b yes yes applied by Zabio
Together AI kimi-k3 yes dropped below 0.95 yes
Moonshot AI kimi-k3.1-pro, kimi-k3.1-pro-highspeed and kimi-k3@moonshot dropped (fixed at 1.0 by the vendor) dropped (fixed at 0.95 by the vendor) yes
Mistral mistral-large-4-0 yes yes applied by Zabio while thinking
Every other model yes yes yes

“While reasoning” and “while thinking” mean the setting sent reasons: a reasoning_effort above none, or the setting auto chose for the job.

claude-fable-5-2, claude-fable-5-1, claude-opus-5-5 and claude-sonnet-5-5 do not take a forced tool choice, that is tool_choice set to required or to a tool’s name (Anthropic’s Fable 5.2 page: “Forced tool use returns a 400 error”). Zabio cannot leave a forced choice out without changing the answer, since it promises a tool call. A request that names one of these models with a forced choice is refused with 400 unsupported_parameter before it reaches Anthropic: send tool_choice: auto, or name another model. On auto these models are not candidates for such a request, and another model answers. auto and none work on every model.