Pricing and credits

Rates

Every model in the catalog costs 40% below provider list price on both Base and Scale. The discount applies to input, cached input, output, cache creation, image, and audio rates. Provider long-context premiums still apply, with the same 40% discount.

For example, gpt-5.6-sol per million tokens:

See the pricing page for every model and rate. Output includes billable reasoning tokens. There is no subscription, seat fee, or per-request fee on standard prepaid pricing.

Worked examples

GPT-5.6 Sol with 10,000 input and 2,000 output tokens:

10,000 ×  $2.40/M = $0.0240
 2,000 × $12.00/M = $0.0240
                    -------
                    $0.0480

Provider list cost is $0.0800. You save $0.0320, or 40%.

GPT-5.6 Sol with 100,000 input tokens, including 80,000 cached, and 5,000 output tokens:

20,000 ×  $2.40/M = $0.0480   (uncached input)
80,000 ×  $0.24/M = $0.0192   (cached input)
 5,000 × $12.00/M = $0.0600
                    -------
                    $0.1272

The same request costs $0.2120 at provider list rates. Base and Scale pay the same $0.1272.

Kimi K2.7 Code with 10,000 input and 2,000 output tokens: 10,000 × $0.57/M + 2,000 × $2.40/M = $0.0105.

Charges use the provider token counts recorded in the response and signed receipt. Past usage keeps its recorded charge and discount; a pricing change applies to new usage.

Credits

• 1 credit = $1 of inference at the current published rates. Credits fund a dollar balance; they do not lock a token price.
• Minimum purchase: $5. Maximum: $25,000 per transaction. The available maximum may be lower when the service approaches its capacity limit.
• Buy credits and view your balance at /dashboard/billing. Credit validity is governed by the Terms of Service.

Large streamed responses

A large response admitted against a small balance can finish slightly negative. Generation is not cut off mid-response. Your next request is blocked with 402 insufficient_credits until you top up; the deficit is deducted from that purchase. Realtime sessions follow the same rule per response.

Caching

Cache reads bill at the model's cached-input rate, with 40% off list. For GPT-5.6 Sol, that is $0.24 per million cached tokens. Prompt caching is automatic on supported Azure models; Claude also supports explicit cache_control blocks.

Cache creation is a separate rate where the provider reports it. Claude charges 1.25 times input for five-minute writes and twice input for one-hour writes, before the discount. Cache-read and cache-write tokens are included in total input and charged once at their respective rates. See Claude usage.

GPT-5.4 Pro has no separate discounted cache-read rate; its input still receives the standard 40% discount.

Scale access

Your account moves to Scale automatically at $5,000 in lifetime purchases. Scale unlocks image generation and realtime voice. Access is permanent, with no ongoing minimum spend. Both tiers pay the same 40% below list.

Enterprise deals

We also offer enterprise deals for teams with larger workloads or specific requirements. Visit Enterprise to tell us about the models you need, expected usage, and requirements.