Models
In short
- Seven Claude models, priced from Anthropic's published list.
- You pay list price plus a 5.5% fee, rounded up once.
- The hold is the most a call could cost; the rest comes back.
Models (cheaper inference) are Petrova's second product. Tools come first (tools).
Petrova is not live. Every model below is answered by a test fixture today. The prices are the real list prices Petrova would charge against, copied from Anthropic's published pricing on 22 October 2026 into one table: packages/catalog/src/rates.ts.
The models
Send the id in the model field. The id must match exactly (or be a dated snapshot of it). Any other name is refused with a 400, never priced by guess. GET /v1/models lists the same ids.
Prices are US dollars per million tokens.
| Model id | Status | Context | Input | Output | Cache read | Cache write 5 min | Cache write 1 h |
|---|---|---|---|---|---|---|---|
claude-opus-5-5 (default) | current | 1,000,000 | $4 | $20 | $0.20 | $5 | $8 |
claude-fable-5-1 | current | 1,000,000 | $10 | $50 | $0.25 | $12.50 | $20 |
claude-sonnet-5-5 | current | 1,000,000 | $2 | $10 | $0.10 | $2.50 | $4 |
claude-haiku-5-5, prompt up to 100,000 tokens | current | 1,000,000 | $0.10 | $0.50 | $0.01 | $0.125 | $0.20 |
claude-haiku-5-5, prompt over 100,000 tokens | current | 1,000,000 | $0.50 | $2.50 | $0.05 | $0.625 | $1 |
claude-opus-5 | legacy | 1,000,000 | $5 | $25 | $0.50 | $6.25 | $10 |
claude-sonnet-5 | legacy | 1,000,000 | $2 | $10 | $0.20 | $2.50 | $4 |
claude-haiku-4-5 | legacy | 200,000 | $1 | $5 | $0.10 | $1.25 | $2 |
Legacy models stay so old ids keep working.
Only Claude models are in the table. There are no OpenAI, Google or open-weight models yet. A request for gpt-4o (or any id not above) is refused as "not priced".
GET /v1/models returns the id, catalog id, status and context size. It does not return prices yet, and it is not in OpenAI's list shape (object: "list"). Prices are on this page and from the MCP catalog tool.
What you pay
A call is charged its supplier cost plus a 5.5% fee, rounded up once (feeNusd, packages/billing/src/settlement.ts). Today the supplier cost is the list price above, so:
| Model | You pay per million tokens (in / out) |
|---|---|
| Opus 5.5 | $4.22 / $21.10 |
| Sonnet 5.5 | $2.11 / $10.55 |
| Haiku 5.5 (short prompts) | $0.1055 / $0.5275 |
When a cheaper reseller serves a call, you pay its cost plus 25% of the saving, at most 6% of what you pay, never above list (billing). No reseller is signed, so today every call pays list + 5.5%.
Batch prices are half of every price in the table. They are used inside Petrova (for example by the repo memory worker) and by the MCP cost tools. There is no batch API for customers yet.
How the hold is sized
Before a call is sent, Petrova holds the most it could cost:
- input: body bytes ÷ 3, rounded up, at the highest input-side price the request could reach,
- plus
max_tokensat the output price, - plus the 5.5% fee.
When the call ends you pay what it really used and the rest comes back at once. Example: Opus 5.5, a 9,000-byte body and max_tokens 4,096 holds $0.099086. The call used 2,000 tokens in and 500 out: list price $0.018000 + fee $0.000990 = $0.018990 charged, $0.080096 back.
A large max_tokens makes a large hold. With a small balance, a call with max_tokens 128,000 on Opus 5.5 holds about $2.70, so a few parallel agents can tie up a small balance even though each call costs cents. Set max_tokens to what you need.
Formats
The same models work in three formats:
- Anthropic Messages:
POST /v1/messages(Claude Code, the Anthropic SDK). - OpenAI Chat Completions:
POST /v1/chat/completions(the OpenAI SDK). - OpenAI Responses:
POST /v1/responses(Codex).
Caps, holds, failover and loop pause work the same in all three. See how it works for what the OpenAI routes do not support.
Last checked against the code: 16 November 2026. Something wrong? The source of this page is in the repository, so a fix is one edit.