ASAI inference

An OpenAI-compatible endpoint. Point any OpenAI SDK at it, change the base URL and the model name, and it works — streaming and tool calls included.

text

Base URL   https://asai.aperturesyndicate.com/api/v1

Method	Endpoint
`POST`	`/chat/completions`	Chat completion, streamed or whole
`GET`	`/models`	The three model names

Keys

Every request carries Authorization: Bearer asai_…. There are two kinds of asai_ key, and they differ only in who pays:

Key	Where	Who pays
Dev Studio key	Dev Studio → ASAI keys	The wallet folder the key is attached to, at twice the model's price. No ASAI plan needed
Personal key	ASAI → Settings → API	Your ASAI energy, the same pool as the chat. Needs ASAI Pro or Etherane

A key is shown once, when it is created — we keep only its hash. Revoke a leaked key and issue a new one. Keys from the old developer portal list, the ones that do not start with asai_, do not open this API.

Models

Model	For
`asai-mini`	Fast everyday answers
`asai-arc`	Harder reasoning, longer answers
`asai-code`	Code — the default when the name matches nothing

The name is matched loosely: anything containing arc goes to asai-arc, anything containing mini or chat to asai-mini, everything else to asai-code. The model field of the response always carries one of the three public names.

Request

python

from openai import OpenAI

client = OpenAI(
    api_key="asai_…",
    base_url="https://asai.aperturesyndicate.com/api/v1",
)

stream = client.chat.completions.create(
    model="asai-code",
    messages=[{"role": "user", "content": "Explain this stack trace"}],
    stream=True,
)
for chunk in stream:
    print(chunk.choices[0].delta.content or "", end="")

Field
`model`	One of the names above
`messages`	Non-empty; roles `system`, `user`, `assistant`, `tool`. `content` is a string or a list of `text` and `image_url` parts
`stream`	`true` for server-sent events in the OpenAI chunk format
`temperature`	Optional
`max_tokens` / `max_completion_tokens`	Optional, capped at 32 768. Without it each model uses its own default
`tools`	Function tools; `parameters` must be a JSON schema with `type: "object"`
`tool_choice`	`auto`, `none`, `required`, or a named function

Other OpenAI fields are ignored. Responses follow the OpenAI format, including usage.

How a request is paid for

A request costs what its tokens cost on the model that answered: fresh input, cached input and output are priced separately.

Dev Studio key — twice that price is charged to the key's wallet folder after the answer, including a stream you cut off midway. When the folder reaches zero, requests get 402 until you add funds in Wallet.
Personal key — the price is taken from your ASAI energy, with the same weekly budget and daily brake as the chat. asai-arc is open on plans that include it.

Errors

Errors come in the OpenAI shape, so SDKs raise their usual exceptions:

json

{ "error": { "message": "Invalid API key.", "type": "invalid_request_error", "code": "invalid_api_key" } }

Status	`code`	Means
`400`	`invalid_request`	Body is not JSON, or `messages` is empty or malformed
`401`	`invalid_api_key`	Missing, wrong or revoked key
`402`	`budget_exhausted`	The wallet folder behind a Dev Studio key is empty
`403`	`model_not_available`	Personal key: the plan has no API access, or no `asai-arc`
`429`	`invalid_request`	60 requests per minute per key exceeded, or a personal key ran out of daily or weekly energy — the message says which
`503`	`billing_unavailable`	We could not check the wallet folder — retry
`5xx`		The model service is unavailable — retry with backoff

Rules

Do not resell access to the API.
Keep keys on your server. A key in a browser bundle or an app binary will be found.
Content rules are the same as in the chat: ASAI Acceptable Use.
Full terms: Developer & API Terms.

BackSign-in with Syndicate NextResonance catalogue