Documentation
Models
One key reaches every model in the catalogue. You choose which one answers per request, by name, so switching model is a one-word change rather than a new integration.
The catalogue
The table below is read from the live gateway, so it is what the API will actually route right now. Put the id in the model field of your request body.
| Model id | Made by | Context window |
|---|---|---|
claude-haiku-4-5 | Anthropic | 200,000 tokens |
claude-sonnet-4-6 | Anthropic | 1,000,000 tokens |
claude-opus-4-6-v1 | Anthropic | 1,000,000 tokens |
deepseek-v3.2 | DeepSeek | 164,000 tokens |
deepseek-r1 | DeepSeek | 128,000 tokens |
kimi-k3 | Moonshot AI | 1,000,000 tokens |
gpt-oss-120b | OpenAI | 128,000 tokens |
qwen3-next-80b | Qwen | 256,000 tokens |
qwen3-vl-235b | Qwen | 256,000 tokens |
qwen3-coder-next | Qwen | 256,000 tokens |
The context window is the largest input a model accepts, counted in tokens. A dash means the model does not publish one through the gateway.
How model ids work
A Kakehashi model id is a stable public name, not the upstream provider's internal identifier. claude-sonnet-4-6 is the name you send; which provider route, region and build sits behind it is ours to manage, and can change without your code changing.
- Ids are case-sensitive and must match exactly. An unknown id is rejected rather than guessed at.
- Every key reaches every model. Keys are not restricted to a subset of the catalogue.
- The catalogue is not frozen. Models are added as they become available and withdrawn when a provider retires them, so read the list rather than hardcoding it — the same way this page does.
How models are priced
You are charged per token, at a published price per million input tokens and per million output tokens, and the cost of each request is deducted from your prepaid balance in US dollars. Prices differ per model, which is the main reason to reach for a smaller one when the work does not need a larger one.
Per-model prices are shown on the portal's Models page once you are signed in. The figures there are the ones you pay: Kakehashi's service fee is already included, so there is nothing to add on afterwards.
How that balance is bought, and in which currency, is covered in Billing and credit.
Reasoning models need room to think
Some models in the catalogue spend output tokens on reasoning before producing any visible answer. If you cap max_tokens tightly, such a model can use the whole allowance thinking and return finish_reason: "length" with an empty content string.
max_tokens and ask again.Model parameters
Parameters in the request body are passed through to the model as they are, so anything your chosen model supports — temperature, top_p, max_tokens, stop and so on — behaves the way that model's own documentation describes. Support differs per model: a parameter one model honours may be ignored or rejected by another.