Skip to main content
Pick a model with the provider/model convention and pass it as the model field. The live, authoritative list (with current pricing) is always at GET /v1/models.

​
Available models

​
Audio transcription

See Audio transcription for a copy-paste file upload and streaming example.

​
Image generation

openai/gpt-image-2 generates images via POST /v1/images/generations, billed per token at the rates in the table below (text input plus image output tokens). See Image generation for the request shape and options.

​
Token-priced models

Token prices are USD per 1M tokens. More providers and models are added over time. The table shows standard rates. google/gemini-3.8-flash, google/gemini-3.7-flash and google/gemini-3.6-flash are promo-priced at 0.75/0.75 / 3.75 through 2026-12-31. zai/glm-5.3-flash is promo-priced at 0.075/0.075 / 0.25 through 2026-09-09. GET /v1/models always returns the effective price you’re billed at.
meta/muse-spark-1.3-contributor is Meta’s contributor tier. It is the same model as meta/muse-spark-1.3 — same weights, same 1M-token context, same modalities — and it is cheaper because Meta uses the prompts and completions you send it to improve their products. The standard meta/muse-spark-1.3 is not used for training. Nothing else in the gateway routes between the two: the model id you send is the tier you get, and a request never falls back from one to the other. Send anything confidential to the standard id.
The table above is the whole catalog: GET /v1/models returns exactly the ids that route, and anything else is a 404 — except a bare id on an endpoint that natively speaks one provider (a bare Anthropic id on /v1/messages, a bare Google id on generateContent), which is prefixed with that provider before lookup.

​
Retirement

Models are occasionally withdrawn. When one is, we announce a date first and the model keeps serving until then — we will email you before it happens. While a retirement is announced, GET /v1/models carries a deprecation block for that model with the date and its successor, so you can check programmatically at any time. Once the model is withdrawn the id returns 404 and the error names its successor, so a client that missed the notice is still told where to go.

​
Credits

Usage is billed from a prepaid credit balance. Credits are denominated in micro-USD:
  • 1 credit = 0.000001∗∗,so∗∗0.000001**, so **1.00 = 1,000,000 credits
  • Buying 10increditsadds10,000,000credits.Thefixed510 in credits adds 10,000,000 credits. The fixed 5% platform fee is 0.50, making the pre-tax total $10.50
  • New accounts get a signup bonus so the first call works immediately

​
How a request is priced

where input_price / output_price are the per-1M-token prices above — the same rate regardless of which of a model’s backends actually serves the request — with no usage markup applied on top. Because a credit is one micro-USD, the “per 1M” division cancels: tokens × price is already micro-USD. The amount is rounded up, deducted when the request finishes, and recorded in the usage ledger (GET /v1/usage). Some models carry more than the two base rates, and the formula extends the same way — each token class times its own rate, summed, rounded up once. GET /v1/models reports every rate that applies to a model:
  • audio_input_per_mtok_usd — audio input tokens priced above the text rate (see Audio & video input)
  • cached_input_per_mtok_usd / cache_write_per_mtok_usd — prompt-cache reads and writes
  • context_price_tiers — long-context brackets: once the prompt crosses a tier’s threshold, its rates apply to the request
Audio transcription uses duration instead:
Billed on the audio’s duration, and only after the transcription completes successfully.
Example — gpt-4o-mini, 11 input + 3 output tokens: ⌈11 × 0.15 + 3 × 0.60⌉ = ⌈3.45⌉ = 4 credits ($0.000004).
The separate fixed 5% platform fee applies only when prepaid credits are purchased. If your balance is empty, model endpoints return 402 (see Errors). Check your balance with GET /v1/account.
Requests routed through your own provider credential (BYOK) skip this formula entirely — they’re billed 0 credits, since the provider charges you directly.

​
Topping up

Buy credits with a hosted Stripe Invoice:
Open invoice_url and pay. Credits are added automatically once the invoice is paid, and Stripe retries cannot double-credit them. The minimum top-up is **5∗∗,andfractionalamountsareroundeddowntoawholedollar.Afixed5feeisaddedtothenetcreditamount,so5**, and fractional amounts are rounded down to a whole dollar. A fixed 5% platform fee is added to the net credit amount, so 10 of credits has a 0.50feeandcosts0.50 fee and costs 10.50 before tax. The invoice has a single Impossible Credits line item covering the credits and fee. Stripe Tax adds VAT or sales tax on top when enabled. GET /v1/billing shows your balance, the minimum top-up, the fixed platform fee rate, and the provider-usage multiplier.