Pay for the answer
Per-token pricing with no seats and no minimums. Everything the panel writes for you — the brief, every draft, the audit — streams into your reasoning stream, so the bill matches what you can read.
Launch draft prices, read live from GET /v1/pricing. They may change before general availability.
The rate card
| Loading prices… |
How a bill is made
Tokens in, tokens out
You pay for the tokens you send and the tokens you receive, at the rate of the model id you called. Nothing else.
You see every draft
The brief, each lensed draft and the audit stream as reasoning_content and bill as reasoning tokens at the output rate. Late drafts we cancel are free; an answer copied from a draft is billed once.
Estimated counts
Token counts are estimated at roughly four characters per token across the panel, so one request has one price whichever models ran.
Old turns cost a tenth
Two caches work for you automatically, and you can steer one of them with a header.
Conversation prefix
When you resend a conversation within an hour, the earlier turns bill at the cached-input rate. Only the new turn pays full input price. Cached tokens appear in usage.prompt_tokens_details.cached_tokens.
Exact repeats
An identical request at temperature 0 is served from the response cache (TTL 1 hour) at cached rates. Send x-tempest-cache: on to cache at other temperatures, or x-tempest-cache: off to bypass it.
A ceiling on every key
Give each key a monthly spend cap in the console. Once it is reached, requests return HTTP 402 insufficient_quota until the 1st of the month (UTC). Nothing runs on past the line.
Check where a key stands at any time with GET /v1/usage: month spend, cap, remaining, and a per-model breakdown.
Pricing FAQ
How do I get started?
Failure Model Labs is invite-only. Ask for an invite key on the contact page, create your Failure ID with it, then make API keys in the console.
Why is a fused answer more tokens than a single model?
Because you are reading four models think. On a hard prompt the brief and three or more drafts stream to you as reasoning before the answer, and each bills as reasoning tokens. Use Flash when one model is enough — it never drafts — or Tempest 1, which only fuses when the prompt is hard.
How do you count tokens?
We estimate at about four characters per token. The figures in each response's usage object are the ones you are billed on, and usage.cost gives the amount in USD.
What happens when I hit my spend cap?
That key returns HTTP 402 with type insufficient_quota until the 1st of the next month (UTC). Raise the cap in the console to continue sooner.
Can we be invoiced?
Invoiced billing is available by arrangement for enterprise customers. Write to sales@failure.fail.
Will prices change?
These are launch drafts and may change. Changes will be announced in the changelog, and the live figures always come from GET /v1/pricing.