Pricing and credits

Rates

One rule for every model in the catalog: the Base tier pays 20% under the provider's list price and the Scale tier pays 25% under, per token, on every rate the provider publishes (input, cached input, output, image, audio).

As an example, gpt-5.6-sol per 1M tokens:

The full per-model table, including grok-4.3, both Kimi K2.7 deployments, and the Scale-tier image and realtime models, is on the models page.

Output tokens include reasoning tokens — see reasoning effort. For models whose vendor reports reasoning outside completion_tokens (Grok does), billed output is total_tokens − prompt_tokens, which matches what the upstream meters. There are no other charges: no subscription, no seat fees, no per-request fee.

Worked examples

Base tier.

A typical gpt-5.6-sol request — 10K input, 2K output:

10,000 ×  $4.00/M = $0.0400
 2,000 × $24.00/M = $0.0480
                    -------
                    $0.0880   (OpenAI list: $0.1100 — saves $0.022, 20%)

A cache-heavy gpt-5.6-sol request — 100K input of which 80K cached, 5K output:

20,000 ×  $4.00/M = $0.0800   (uncached input)
80,000 ×  $0.40/M = $0.0320   (cached input)
 5,000 × $24.00/M = $0.1200
                    -------
                    $0.2320   (same request with no cache hits: $0.5200)

The same request at the Scale tier: $0.075 + $0.030 + $0.1125 = $0.2175.

The typical request on kimi-k2.7-code instead — 10K input, 2K output: 10,000 × $0.76/M + 2,000 × $3.20/M = $0.0140, about six times cheaper than gpt-5.6-sol for token-for-token workloads that fit a smaller model.

Every charge is computed from the token counts in the response's usage block — the same counts committed to in the signed receipt, so you can audit any line item.

Credits

• 1 credit = $1 of inference on any model at your tier's rates. Buying credits does not lock in a rate; it funds a balance that is drawn down at whatever tier you hold when each request runs.
• Minimum purchase: $5. Maximum: $25,000 per transaction (the available maximum may be lower, subject to a global capacity limit).
• Purchase and view balance at /dashboard/billing. Credit validity is governed by the Terms of Service.

Balance edge case: large streamed responses

Admission is checked at request start, but a streamed response is metered as it generates. A large stream admitted against a small balance can therefore finish slightly negative. The stream is never cut off mid-response — you keep the tokens you were served — but the next request is blocked with 402 insufficient_credits until you top up, and the deficit is deducted from that purchase. Realtime sessions behave the same way per response: the response that exhausts the balance completes, then the session closes.

Cached input billing

Prompt caching is automatic — no configuration, no separate storage fee. When the upstream reuses a cached prompt prefix, those tokens appear in usage.prompt_tokens_details.cached_tokens (Chat Completions) or usage.input_tokens_details.cached_tokens (Responses) and bill at the model's cached rate with your tier's discount applied — on gpt-5.6-sol that is $0.40/M Base, $0.375/M Scale, a 10× discount against uncached input. Prompts structured with their stable prefix first (system prompt, tool definitions, few-shot examples, then the variable suffix) hit the cache most often.

gpt-5.4-pro is the one exception: the upstream offers no cached rate for it, so all input bills at the input rate.

Cached tokens are recorded separately in the receipt (cached_input_tokens), so the discount is verifiable per request.

Scale tier

When your lifetime purchases reach $5,000, your account moves to the Scale tier automatically:

• Rates drop to 25% under list on every model.
• gpt-image-2 (image generation) and gpt-realtime-2.1 (realtime voice) unlock.
• The change is permanent. There is no minimum spend to maintain, no annual reset, no negotiation.
• It applies to every request from the moment the qualifying purchase settles, and to your entire existing balance — credits are dollars, so a tier upgrade means each remaining dollar buys more tokens.

Progress toward the threshold is shown in /dashboard/billing.