Capabilities

What actually works today beyond the base request shape — tool calling, vision, and the in-repo SDKs — each confirmed by live testing, not assumed.

Why this page exists

The base contract in the reference covers model, messages, max_tokens, temperature, top_p, stop, and stream. It says nothing about the request shapes below, so there was no way to tell whether they worked. Each one here was confirmed against a real deployed edge, not inferred from reading code — see Limits for what was tested and found broken instead.

Tool calling

Tool calling is a first-class, router-aware capability: the router only serves a tools-enabled request from a model that actually supports function calling, and this is covered by the router's own test suite. It is not a "probably passes through" extra.

r = client.chat.completions.create(
    model="auto",
    messages=[{"role": "user", "content": "What's the weather in Helena, Montana?"}],
    tools=[{
        "type": "function",
        "function": {
            "name": "get_weather",
            "parameters": {
                "type": "object",
                "properties": {"city": {"type": "string"}},
                "required": ["city"],
            },
        },
    }],
    tool_choice="auto",
)

tool_choice supports the standard values:

Value	Behavior
`"auto"`	Model decides whether to call a tool.
`"required"`	Forces a call.
`"none"`	Suppresses tool use — the model answers in text only.
`{"type": "function", "function": {"name": "..."}}`	Forces a specific named function.

tool_choice: none — know the history

Live testing on 2026-08-28 found tool_choice: "none" silently ignored on both local and staging — the model called the tool anyway, with no signal in the response that the constraint was dropped. That defect (and a related gap in named-function forcing on some model families) was root-caused and fixed (PR #1029). If you built a workaround around "none" not working, you can remove it.

Vision

Image content parts work. Send an image as a data URI or URL in a content array, the same shape as OpenAI's vision API:

r = client.chat.completions.create(
    model="auto",
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", "text": "What color is this?"},
            {"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}},
        ],
    }],
)

If you test this with a tiny image, read this first

A 1×1-pixel test image produced a wrong-looking answer during verification — that turned out to be a degenerate-input artifact (the image was too small for the vision encoder to resolve), not a dropped feature. A 64×64 solid-color control image answered correctly. Test with a real image before concluding vision is broken.

Anthropic Messages and OpenAI Responses shapes

POST /v1/messages (Anthropic's wire format) and POST /v1/responses (OpenAI's Responses wire format) are both supported, streaming and non-streaming, as of 2026-09-01.

Sending Anthropic's shape does not pin you to an Anthropic-served model: ModelBeat's own catalog-driven routing still decides which model serves the request, the same as it does for /v1/chat/completions.

import anthropic

client = anthropic.Anthropic(
    api_key="mb_live_...",
    base_url="https://api.modelbeat.ai/v1",
)

r = client.messages.create(
    model="auto",
    max_tokens=100,
    messages=[{"role": "user", "content": "Write a haiku about prepaid inference."}],
)

See Limits for the one billing caveat worth knowing before you rely on this: prompt-cache tokens are billed at the regular input rate, not Anthropic's cheaper cache-read rate.

SDKs

sdks/python and sdks/typescript are real, versioned clients — not internal tooling. Both are thin wrappers around the official OpenAI SDK: you get typed routing_info (tier + fallback flag; provider/model identity is None/undefined during the private beta — see Limits) and typed cost instead of untyped extra fields, plus a raw escape hatch to the underlying OpenAI client for anything the wrapper doesn't cover yet.

pip install modelbeat        # Python
npm install @modelbeat/sdk    # TypeScript

Both wrappers currently cover only chat.completions.create and completions.create. For /v1/embeddings or /v1/feedback, use the raw client. Full usage, error handling, and current limits are in each package's own README: sdks/python/README.md, sdks/typescript/README.md.

Not covered here

Structured outputs (response_format) — confirmed broken, not documented as working. See Limits.
Idempotency-Key — supported: send the same key on a retry and the request is billed once. Reusing a key for a different body, or one whose reservation already settled, is refused with 409. See the API reference.

Limits. What is supported, and what deliberately or currently is not.
API reference. Full schemas, with a playground.

Tool calling

Vision

Anthropic Messages and OpenAI Responses shapes

SDKs

Not covered here

Next

On this page