Billing
Zabio is prepaid. You add credits in US dollars, every call is debited from the balance, and requests stop when the balance runs out. There are no subscriptions and no fees on top-ups. The Free plan, the top-up match and referral rewards, and how each adds to your balance, are in Credits, match and referrals.
What a call costs
Section titled “What a call costs”Every call is billed at the provider’s list price plus Zabio’s fees, each shown as its own line:
- a service fee of (see the pricing page) on the provider price of every call;
- on every
autorequest, the price of automatic model choice: the same as the model’s own cost for the request, at least (see the pricing page) and at most (see the pricing page), so a small request pays the least and a large one pays the mosta fixed (see the pricing page). It is paid once per request, whatever happens inside it: a classification reused from the same session, Smart Splitting’s two calls, a check that moved the answer up a band,zabio_escalate, or an answer by Jev. Everyautoline of a batch and everyautobackground response is one request.
A call to a model you name also pays (see the pricing page). A call to a model you name pays the service fee only.
We take no cut of what routing saves you: every cent of the saving stays with you. Prices and fees can change, for example when providers change theirs, and the pricing page always shows the current ones.
Every response shows what the call cost in zabio.cost: vendor_micros, service_fee_micros, call_fee_micros (the fee of an auto request, the price of automatic model choice) and share_discount_micros (what the sharing discount took off the fees, already out of them). It shows the cost only: what routing saved is computed for a period, never per call (see What routing saved). An answer left empty because the model spent its whole output budget reasoning is never billed (see Checked answers). savings_fee_micros and classifier_micros always read 0 (see The classifier); they stay so code that reads them keeps working. Amounts are in micros: 1,000,000 micros is one US dollar.
When to name a model. On a small request, the price of automatic model choice can be as much as the model itself. If your requests are small and alike, and auto keeps choosing the same model for them (Logs and zabio.routing show which), name that model in model: those calls pay the service fee only. The playground’s tips say when a request is one of these. Keep auto for requests that vary, where the choice pays for itself.
Worked example
Section titled “Worked example”You send a short request with model: "auto". The classifier finds an easy job and Zabio picks the cheapest model able to do it. The fee lines follow the plan in force.
| Line | Amount |
|---|---|
Provider list price of the call (vendor_micros) |
0.0100 USD |
Service fee on the provider price (service_fee_micros) |
(see the pricing page) |
Automatic model choice (call_fee_micros) |
(see the pricing page) |
Billed (cost_micros) |
(see the pricing page) |
The same request sent with model: "gpt-5.5" directly would bill (see the pricing page).
A split request
Section titled “A split request”Under auto, a big request whose question needs only part of its source can be answered in two steps, an extract on a cheap model and an answer on the routed model (see Split and route). The two calls are billed at their providers’ list prices as two calls of one request: the service fee on both, and the request’s one fee for automatic model choice. The classifier, the questions and the check Zabio ran to decide the split are Zabio’s own cost and never billed.
Illustrative prices. You send a 30,000-token contract with a question about one clause; routing picks a model at 4 USD per million input tokens; the extract runs on a model at 0.20 USD per million; the answer is the same length on both paths, say 0.0100 USD of output.
| Line | Amount |
|---|---|
| Provider list price of the extract: 30,000 tokens in at 0.20 USD per million, plus a short output | 0.0070 USD |
| Provider list price of the answer on the excerpt: about 2,000 tokens in at 4 USD per million, plus the output | 0.0180 USD |
Provider price of the request, both calls (vendor_micros) |
0.0250 USD |
Billed (cost_micros) |
(see the pricing page) |
The same request sent whole would bill (see the pricing page). zabio.cost on a split request reports these totals, and zabio.split.steps lists each call with its price; the questions and the check appear there with billed: false.
When a split falls back to the full source, whatever the reason, you pay exactly what the unsplit call costs: the extract and the check Zabio ran are Zabio’s cost, listed in zabio.split.steps with billed: false, and zabio.cost is the ordinary breakdown of the one call that answered. With your own vendor key, the extract runs on it when the key covers the extractor’s vendor, like any call.
A classification job answered by Jev
Section titled “A classification job answered by Jev”Under auto, a classification job with a strict JSON schema may be answered by Jev (jev-1.14.0) instead of the routed text model: on an easy or medium classification job by the line’s rule on our measurements of Jev against the routed models; on other jobs when Zabio’s estimate says Jev is cheaper (see Classification jobs on auto). On a short job whose routed model our tests have not measured, Jev 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. The call is billed as any auto call: Jev’s list price (0.042 USD per million input tokens, output free), the service fee on it, and the request’s one fee for automatic model choice.
When Jev’s answer is discarded (a field under the threshold, or a failed call), you pay exactly what the text model’s call costs, as if Jev had never been asked; the Jev call is Zabio’s cost and does not appear in zabio.cost.
A flex call
Section titled “A flex call”A request with a zabio.deadline of 15 minutes or more can be served on the vendor’s flex tier where the model has one (see Flex tier): the same model at about half the list price. The call is billed at the flex price the vendor charged, plus the usual fees. In what routing saved, a flex call is compared with your key’s baseline model at its flex price, so the vendor’s own discount is never counted as a saving.
When the vendor refuses a flex request for capacity, Zabio serves the same model at the standard tier and bills the standard price; zabio.tier.fallback says capacity.
An off-peak call
Section titled “An off-peak call”Some vendors charge less outside their peak hours (see the DeepSeek note under Prices; GET /v1/public/prices lists every model’s off_peak price and hours). A call takes the off-peak price only when both its arrival and its completion fall inside the off-peak hours; a call that crosses the boundary in either direction is billed at the list price. Where a vendor states that it charges by the request’s start or by its completion, Zabio follows that instead, and the charged_by field of the price says which.
A batched line
Section titled “A batched line”A line of a batch is billed like a call served at the vendor’s batch price, the rate the vendor’s batch API charges, about half of the list price, plus the usual fees: every auto line is one request and pays the fee for automatic model choice once. In what routing saved, a batched line is compared with your key’s baseline model at its batch price.
When the batch is accepted, an estimate of its cost, fees included, is set aside on your balance, the hold, and released as lines are billed; the hold is only the estimate, and each line is billed for the tokens it really used. If the estimate exceeds your available balance, the batch is refused at creation with insufficient_credits and nothing is billed. A running batch is never stopped for money.
A call with web search
Section titled “A call with web search”Web search costs 0.0105 USD a search: the provider’s price of 10 USD per 1,000 searches plus Zabio’s 5 percent. The results the model reads are billed as input tokens of the call. See Web search.
An AI Firewall check
Section titled “An AI Firewall check”A check on POST /v1/guard/check or POST /v1/moderations with the patterns only is free, up to 60 checks a minute per API key. With the model check it costs 0.50 USD per million tokens checked, as the AI Firewall’s paid tier does on a call, on its own line (“AI firewall”), with no fee on top: checking 2,000 tokens of fetched pages costs 0.0010 USD. cost_micros on the answer says what each check cost. The model check needs credit your organization added, such as a top-up; the trial credit alone does not pay for it. See AI Firewall check.
Current prices
Section titled “Current prices”GET https://api.zabio.ai/v1/public/prices lists the provider list price of every model Zabio serves, in US dollars per million tokens, before Zabio’s fees (the fees are the lines above, and GET /v1/public/plan lists them). It needs no API key, so a script or a spreadsheet can read it.
as_ofsays when the prices last changed: the last price change Zabio applied, or the moment a price announced with a date came into force. It is not the time of your request.- Each model has
maker, the company that made it (asowned_byinGET /v1/models),input,output,cache_read,cache_write_5mandcache_write_1h,price_from(when its current price started) andsource(where the price was read: the vendor’s price API, its price page, its billing report, OpenRouter’s list, a test call Zabio made or Zabio’s operators). long_context,off_peak,flexandbatchcarry the rates of those tiers, ornullwhen the model has none. A cache rate of 0 means the vendor has no separate rate for that part: a cache read or write is then billed at the input rate.retires_atis the vendor’s announced retirement date, ornull.
The list is cached for five minutes. Send its ETag back in If-None-Match and an unchanged list answers 304 with no body. Every field is described in openapi.yaml.
Sharing discount
Section titled “Sharing discount”Your organization shares a percentage of its calls with Zabio, de-identified, to improve how Zabio routes requests (100 percent by default during the beta; see Privacy). In return the fee is discounted: 0.05 percent of the fee for every 1 percent shared, up to 5 percent off at 100 percent, on the service fee and on the per call fee alike (on your own vendor keys, which pay no service fee, on the per call fee). The owner sets the share from 0 to 100 in Organization settings; a project or an API key can set its own, and the share in force for a call is the key’s, else the project’s, else the organization’s, else the plan’s default.
A call is never shared, and earns no discount, when the request says so: store: false, the header X-Zabio-Share: no, or zabio.share: false in the body. A key, project or organization at 0 shares nothing and earns nothing.
Every response states what came off in zabio.cost.share_discount_micros. The usage totals and the CSV export carry a Sharing discount line, and each call’s detail in Logs says whether it counted as shared and at what share. The discount changes nothing about credits: it reduces what a call bills.
What routing saved
Section titled “What routing saved”No call carries a routing saving of its own: API responses, Logs, the CSV export and the MCP tools that show a call (recent_calls, get_call, ask) show what the call cost. The saving is computed only when you look at it, for the period you choose: on the Dashboard and in Usage (Saved by routing), in the CLI (zabio usage), in the MCP usage tool, and through the API (GET /admin/v1/orgs/{org}/usage/savings). It is computed for each API key, like this:
- The key’s baseline model. The model at the 95th percentile of list price per token among the key’s successful calls in the last 30 days, counted back from the moment you look: the calls are ordered by that price, the most expensive 5 percent are set aside so a burst does not count, and the next one is the baseline model. It follows what your hardest jobs needed. The rule needs 20 successful calls in those 30 days: under 20, the 5 percent set aside is zero calls, and one expensive call would set the baseline on its own. Until then the key shows no saving, only how many calls it has so far.
- Only
autocounts. Only requests toauto,auto/cheap,auto/balancedandauto/bestcan save. A model you named is your own choice, so it saves nothing. Calls on the Free tier are left out. - What the baseline model would have cost. Every answer the key’s
autorequests received in the period, priced on the key’s baseline model at the price that model had when the call ran: the same tokens, with the same cache reads and writes at that model’s own cache rates (uncached where its vendor has no prompt cache), and at its flex or batch price when the call ran at flex or batch (its standard price when it has none). When a price changed close to the time of a call, the lower price is used, so the saving is never overstated. - What you paid. Everything the key’s
autorequests in the period were billed, fees included, with the extract of a split request and any retry that was billed. On a call served with your own provider key, what the provider charged you, at its list price, is added, because you paid that too. - The saving. What the baseline model would have cost minus what you paid, never below zero. When the key’s
autocalls cost as much as the baseline model, or more (for example, every call went to that same model), the key shows a saving of zero and says so. The total is the sum of the keys’ savings: a key that saved nothing counts as zero and never reduces another key’s saving.
A cache discount is never counted as a saving, and neither is a vendor’s flex or batch discount: the baseline model is priced with the cache and the tier you would have had there too. Because the baseline model and its 30 days are taken at the moment you look, the same period can show a different saving later, as your keys’ traffic changes. A figure is kept for one minute, so reloading the page right away shows the same number.
Worked example: the saving of a month
Section titled “Worked example: the saving of a month”Illustrative numbers. In one month a key made 1,000 auto requests. When you open the Dashboard, the key’s baseline model over its last 30 days is GPT-5.5.
| Line | Amount |
|---|---|
| The same 1,000 answers on GPT-5.5: the same tokens and cache, at GPT-5.5’s price when each call ran | 50.00 USD |
What the key’s auto requests were billed, fees included |
12.00 USD |
| Saved by routing | 38.00 USD |
Ten turns into a conversation, a call reads 18,000 tokens of history from the provider’s prompt cache. In the 50.00 USD above, that call is priced on GPT-5.5 with the same 18,000 tokens read from GPT-5.5’s cache, at its cache rates, not as if they were new input: you would have had the cache on GPT-5.5 too, so the provider’s discount never shows as a saving.
A second key made 10 auto requests in its first days. It has 10 successful calls in its last 30 days, under the 20 the rule needs, so it shows “Shows after 20 calls on a key (10 so far)” and adds nothing to the total.
The classifier
Section titled “The classifier”When you send auto, a small, fast model reads the request and names the job: its task type, difficulty and stakes. The fee for automatic model choice is the price of that step and of picking the model; the classifier’s own call is never passed on to you. The response names the classifier that ran in zabio.routing.classifier. classifier_micros stays in zabio.cost and always reads 0, so code that reads it keeps working. Before version 1.2.6 the classifier’s cost was passed on to you at the provider’s price, and the dashboard still shows that line for those days.
Calls served with your own vendor keys pay no service fee; their auto requests pay the fee for automatic model choice. See Own keys.
Every fee and the minimum top-up are a versioned pricing plan. The plan in force is on the pricing page, on the Billing page and in zabio billing. A plan change never rewrites past calls or top-ups: each keeps the version it was billed under.
Answers Zabio checks
Section titled “Answers Zabio checks”On a request that is not streamed, Zabio checks the answer against what the request asked for: JSON that matches your response_format schema, a JSON object for json_object, tool calls that name one of your tools with arguments that match its parameters, and an answer that is not empty because the model used its whole output budget thinking. When auto sends a failed answer once more, to a stronger model or to a model that reasons by default, you pay only for the answer you receive: the failed attempt is Zabio’s cost and is not in zabio.cost. zabio.escalated_from then names the first model. Streamed requests are not checked, because the answer reaches you while it is being written. See Checked answers.
Calls that fail or that you cancel
Section titled “Calls that fail or that you cancel”A call that fails is never billed, also when the model provider had already started on it: Zabio pays the provider for that work. If you close the connection before the answer is complete, the call is cancelled and you pay for the work done, the input and the output generated before you disconnected, streamed or not, without the per-call fee.
The Free plan
Section titled “The Free plan”A new organization starts on the Free plan: no credit card, no trial credit, and no expiry. Free is one organization per person: the first organization you create has the Free allowance; a second one needs a top-up before its first call. A Free organization has one member and two API keys (the keys of connected AI apps do not count). Every feature paid accounts have is included: streaming, tools, JSON output, Smart Splitting, Jev, batches, flex and background responses.
The limits, per organization:
| Limit | Value |
|---|---|
| Requests a day | 100 successful requests, reset at 00:00 UTC. Failed calls do not count |
| Requests a minute | 20 |
| Request size | 16,000 input tokens and 2,000 output tokens. An unset max_tokens is capped at 2,000; a larger one is refused |
| Batches | each line of a batch counts as one request, with the same size limit |
Free accounts also have a daily fair-use limit on what their calls cost Zabio. Normal use never reaches it; when it is reached the API answers 429 free_tier_limit with param: daily_cost, says when it resets, and resumes at 00:00 UTC. Over any limit the response names the limit in param and in the x-zabio-free-limit header, carries Retry-After when the limit resets, and ends with the one thing that removes every limit: a top-up. See Errors.
The free models. Free calls run on the free model list: today gpt-oss-120b and DeepSeek Flash at Together AI (North America), mistral-small-2603, gpt-5.4-nano, qwen3.5-9b and gemini-3.5-flash-lite, served by the vendors themselves under their standard API terms, which do not train on your data. The list is live at GET https://api.zabio.ai/v1/public/free-tier, with no key, and GET /v1/models marks each model with zabio.free_tier. auto, auto/cheap, auto/balanced and auto/best keep working and choose among the free models; on auto, the response’s zabio.plan.paid_would_use says which model a paid account would have been given, when it differs.
Locked models. Every other model stays visible, in the dashboard, in the Playground and in GET /v1/models (zabio.availability: locked), and is locked: a request that names one answers 403 free_tier_model, “Top up your account to unlock this model”. Named models you serve with your own vendor key are not locked.
What every Free response says. x-zabio-plan: free, the day’s requests in x-ratelimit-limit-requests-day, x-ratelimit-remaining-requests-day and x-ratelimit-reset-requests-day, the minute bucket in x-ratelimit-*-requests, and a zabio.plan object (tier: free, reason: free_plan, paid_would_use on auto, topup_url). The dashboard shows today’s requests against the limit and the time to the reset.
Any credit removes every limit. From the first top-up on, or from any other credit in your balance (a referral reward, the match, credit we add to your account), every model is unlocked, there are no daily request or size limits, the member and key caps are lifted, and you pay as you go at the prices above. There is no plan to pick and nothing to cancel.
The trial of every model. The owner of a Free organization can try every model before paying: add a payment card on the dashboard or the Billing page and your organization gets 5.00 USD of calls on any model for 30 days. Stripe checks the card for 0 USD; nothing is charged, now or later. The card is saved for one-click top-ups and automatic recharge, and you can remove it any time from the Billing page. The trial is given once per person and once per card, in your first organization, and only while it has never topped up. While it runs, the dashboard says what is left (“Trial of every model: 3.20 USD left, 21 days”) and every model is open, at the prices above; the Free plan’s member and key caps still apply. When the 5.00 USD or the 30 days end, whichever comes first, whatever is left of the trial is taken back and your organization is back on the Free plan’s models and limits, unless you topped up. A top-up during the trial keeps every model after it. The trial is promotional credit with its own rules (it expires, it never transfers); see Credits and promotions.
When a paid account’s balance reaches zero, its requests stop until the next top-up: 402 insufficient_credits, the message naming where to add credits. If you would rather keep requests working on the free models within the Free limits when the balance runs out, an owner turns on “Continue on the free tier” in the organization’s settings (zero_balance: continue_free on the API; it is off by default): then zabio.plan.reason is zero_balance, the dashboard shows a banner, and the owners and admins get an email. Nothing is taken away either way: your keys and members stay. A balance below zero (the last call overshooting its estimate) refuses every request, the free tier included, until a top-up brings it above zero; the message names the least amount to add.
Adding credits
Section titled “Adding credits”Credits are bought through Stripe from the Billing page (the payment form opens inside the page) or with zabio billing topup --amount 50, which opens Stripe’s payment page. Rules:
- minimum (see the pricing page) per top-up, maximum 5,000.00 USD per payment;
- amounts in whole cents;
- no fee on top-ups: 50.00 USD paid is 50.00 USD of credit;the top-up fees on the pricing page are taken off the credit, and the payment form shows the fee and the credit before you pay;
- new accounts can add up to 50.00 USD per payment until two payments have cleared or the account is 30 days old; the form says so while it applies;
- the balance updates as soon as Stripe confirms the payment, usually within seconds. If your browser lost the confirmation page, the Billing page checks the payment when you return;
- the payer receives a receipt by email with the amount, the new balance and the billing details.
Billing details
Section titled “Billing details”Your first top-up asks for the billing name (a person or a company), the country, the address, the city, the state or region where the country has one, and the postal code, plus an optional tax ID (VAT, EIN) for your invoices. Stripe keeps them on your customer record and Zabio prints them on every receipt. You can see and edit them any time in Billing settings; nothing is asked at sign-up or on the Free plan.
Saved cards and one-click top-ups
Section titled “Saved cards and one-click top-ups”When you add credits you can tick “Save this card for one-click top-ups and automatic recharge”. A card confirmed when you sign up or start the trial of every model is saved the same way. Stripe keeps the card; Zabio stores only the brand, the last four digits and the expiry, and lists the card on the Billing page, where owners and admins can make it the default or remove it. With a saved card, Add credits offers “Pay with Visa 4242” and charges the card at once; if your bank asks you to confirm the payment (3D Secure), the confirmation opens in the page and the credits are added when it completes. A declined payment says why (insufficient funds, expired card, and so on) and nothing is charged.
Card numbers never touch Zabio. The payment form and the saved cards are Stripe’s; Zabio sees the result of each payment and your Stripe customer id.
Automatic recharge
Section titled “Automatic recharge”Off unless you turn it on. On the Billing page an owner or admin sets the balance below which Zabio recharges, the amount it adds each time (at least the minimum top-up) and a monthly cap (200.00 USD unless you change it), and picks the saved card. Automatic recharge can be turned on after your first top-up, with a saved card. It is checked after each billed call, so a recharge usually lands before the balance reaches zero; when it cannot, your organization runs on the Free limits until it does. Each recharge sends you a receipt by email. If your bank asks you to confirm a charge, you get a link to confirm it and nothing else is attempted until you do; if a charge fails, Zabio retries once an hour later, then pauses automatic recharge and emails you; reaching the monthly cap pauses it until the next month and emails you. A disputed payment switches it off. You can change the settings, pause or turn it off at any time; the Billing page shows what was added this month.
Disputes
Section titled “Disputes”If a card payment is disputed with the bank, the organization is suspended at once, its API keys stop working and the Zabio team reviews the dispute before anything else happens. If you believe a charge was wrong, write to support first: a refund through Stripe reduces your balance and nothing is suspended.
Moving credit between your organizations
Section titled “Moving credit between your organizations”Balances are per organization. If you own more than one, you can move paid credit from one to another on the Billing page of the sending organization (owners only, from the web app). The rules:
- only paid credit moves: credit you bought with a top-up, or credit moved to you. Trial, promotional, referral and match credit stays where it was granted;
- only paid credit at least 7 days old moves, and credit you receive waits its own 7 days before it can move on;
- the amount is at most what is left of that credit within your balance. When a balance mixes paid and promotional credit, your spending is taken to have used the promotional credit first, so what can move is never more than the paid credit you still hold;
- one transfer a day per sending organization, at least 1.00 USD;
- nothing moves while either organization is suspended or has a payment under review.
A transfer is not a payment: nothing is charged, the receiving organization’s “added in total” does not change, and it does not count as a first top-up for the limits on new accounts. The receiving organization runs on the paid plan while it holds the credit, with every model and no Free limits. Both ledgers show the movement as transfer_out and transfer_in, and the recent transfers are listed on the Billing page of each side.
The ledger
Section titled “The ledger”Every movement is a line in the ledger (Billing page, or zabio billing ledger): topup, trial, usage (the day’s calls, written daily), refund, adjustment, transfer_out, transfer_in, allowance (the trial of every model, shown as “Trial of every model”) and allowance_expiry (what was left of it at its end, “Trial ended”). Each line carries the signed amount, the balance after it, and a reference such as the Stripe session id. Call level usage can be exported as CSV from the dashboard.
Low balance warnings
Section titled “Low balance warnings”Owners and admins receive an email when the balance falls below 20 percent of the last top-up and again below 5 percent. At zero, an organization that turned the free tier on gets the free-tier email (what the limits are, how to top up) and another when a top-up takes it back; the rest get the exhausted notice. Each level mails once; any credit added resets the warnings.
What happens at zero
Section titled “What happens at zero”Three sentences say the rule. A request whose estimated cost (its input, the output the model is expected to write and the fee, on the dearest model auto may pick) does not fit the balance left after the calls still in flight is refused with 402 insufficient_credits, the message naming the estimate, the balance and the least amount to add. At zero a request is refused the same way, unless an owner turned on “Continue on the free tier”, in which case the organization runs on the Free limits described above (auto and the free models keep answering within the daily limits, every response saying so in zabio.plan). Below zero, which a call already admitted can leave by overshooting its estimate, every request is refused, the free tier included, until a top-up brings the balance above zero. Calls to a named model that your own vendor key serves keep working in every case, because Zabio has no provider cost on them.
A request with no max_tokens when the balance is low. A request that sets no max_tokens has no output limit. When your balance cannot pay for the model’s full output on that request, Zabio sends the model the output limit your balance can still pay for, fees included, instead of refusing the call; the answer may then stop early with finish_reason length. In a batch this is worked out for each line, counting the line’s share of the hold. With a normal balance it never applies, and it never applies when you set max_tokens yourself or on your own provider keys.
Requests are refused with 402 insufficient_credits in four cases: the call’s estimated cost does not fit the balance left after the calls still in flight, each of which has its own estimate set aside until it ends; the balance is at zero and the organization has not turned the free tier on; the balance is below zero; or it is a person’s second organization at zero, which has no Free allowance. The message names the estimate, the balance and the least amount to add, in USD. The estimate rule covers chat, Responses, embeddings, speech, transcriptions, image generations and edits, background image generations and videos; a background image generation and a video have their estimate set aside when they are queued, and credit is checked again when they run. In every case, named models on your own vendor keys still work; auto does not, because the classifier runs on Zabio’s key, and neither does a stored conversation that needs compaction, whatever model the turn names, because the summary runs on Zabio. Add credits, or move paid credit from another organization you own, and everything resumes immediately.
Refunds
Section titled “Refunds”Refunds are handled through Stripe. When a payment is refunded, the ledger receives a refund line for the amount and the balance goes down accordingly. Write to support with the Stripe session id or the ledger reference if you need a refund.
Spend controls
Section titled “Spend controls”Beyond the balance, you can cap spend where it happens:
- a spend limit per API key, reset every day, week or month, or set for the key’s whole life;
- a monthly quota per member, applied to every key that member created;
- requests per minute and tokens per minute limits per key.
When a cap is reached, calls on that key return 429 quota_exceeded until the next calendar month. See Teams and Errors.