Anthropic SDK

Aile's /v1/messages is the Anthropic wire format exactly.

Pythonpython

1import os
2from anthropic import Anthropic
3
4client = Anthropic(
5  base_url="https://api.aile.sh",
6  api_key=os.environ["AILE_KEY"],
7)
8
9msg = client.messages.create(
10  model="cc/claude-sonnet-5",
11  max_tokens=256,
12  messages=[{"role": "user", "content": "Say hello."}],
13)
14print(msg.content[0].text)

The base URL has no /v1

The Anthropic SDK appends /v1/messages itself, so it takes the bare origin: https://api.aile.sh. This is the opposite of the OpenAI client, which wants the /v1 suffix. If you pass https://api.aile.sh/v1 anyway, Aile serves the doubled /v1/v1/messages as /v1/messages, so it still works.

Authentication

The Anthropic SDK sends x-api-key and only that. Aile accepts it, and accepts Authorization: Bearer from OpenAI-shaped clients. Both reach the same key.

Non-Claude models work here too

The endpoint fixes the wire dialect; the model you name still comes from the whole catalogue, and names its provider. codex/gpt-5.6-sol through the Anthropic SDK routes cleanly to a Codex account:

Pythonpython

1msg = client.messages.create(
2    model="codex/gpt-5.6-sol",
3    max_tokens=256,
4    messages=[{"role": "user", "content": "Say hello."}],
5)

Aile converts your Anthropic body into canonical OpenAI Chat first, then into the provider's own shape. See Model names.

Sending Aile headers

Pythonpython

1msg = client.messages.create(
2  model="cc/claude-sonnet-5",
3  max_tokens=256,
4  messages=[{"role": "user", "content": "Say hello."}],
5  extra_headers={"x-aile-max-price": "8"},
6)

Models that always think

Claude Opus 5.5, Sonnet 5.5, Fable 5.1 and Mythos 5.1 refuse a request that turns thinking off or forces a tool choice. Rather than forward that into a 400, Aile sends the nearest request the model accepts:

You send	What goes upstream
`thinking: {"type": "disabled"}`	Sonnet 5.5: `{"type": "between_tools"}`, its no-up-front-thinking setting. The others: thinking left at the model's default
`reasoning_effort: "none"` or `"minimal"` (OpenAI SDK)	the lowest thinking setting, at effort `low`
`thinking: {"type": "enabled", "budget_tokens": N}` on a model that only takes adaptive thinking	`{"type": "adaptive"}`, at the effort you named
`tool_choice` forcing `any` or one tool	`auto`
an effort the model does not take (`xhigh` or `max`)	the nearest one it does

No response header reports these rewrites. The same rules apply whether you call /v1/messages or /v1/chat/completions.

Counting tokens

Pythonpython

1client.messages.count_tokens(
2    model="cc/claude-sonnet-5",
3    messages=[{"role": "user", "content": "…"}],
4)

Advisory only: it probes no sharer, bills nothing, and estimates at roughly four characters per token.

Errors come back in Anthropic's shape

Because you called /v1/messages, errors are:

JSONjson

1{ "type": "error", "error": { "type": "invalid_request_error", "message": "…" } }

There is no code field in this shape: that is what the Anthropic SDK parses. The shape follows the endpoint you called, regardless of the model you asked for. See Error codes.