Anthropic API¶

Claude models are available two ways: through the OpenAI-compatible /v1/chat/completions endpoint like every other model, or through Anthropic's own Messages API at /v1/messages. The native endpoint accepts the same request and response shapes as api.anthropic.com, so an existing Anthropic client only needs its base URL changed.

from anthropic import Anthropic

client = Anthropic(
    base_url="https://llm.paydigital.shop",
    api_key="YOUR_KEY",
)

message = client.messages.create(
    model="claude-sonnet-4-5",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello"}],
)

Both x-api-key and Authorization: Bearer are accepted, so the official SDKs work unmodified.

Endpoints¶

Endpoint	Status
`POST /v1/messages`	Available
`POST /v1/messages/count_tokens`	Available
`GET /v1/models`	Available
`GET /v1/files`, `GET /v1/files/{id}`, `GET /v1/files/{id}/content`, `DELETE /v1/files/{id}`	Available
`POST /v1/files` (upload)	Not available yet — see below
`POST /v1/messages/batches` and the rest of the Batches API	Not available

What works on every Claude model¶

Verified against all twelve Claude models on this platform:

Multi-turn conversations, system prompts
Streaming (SSE) — content_block_delta through message_stop
Tool use, including tool_choice forcing a specific tool
Full tool round-trip — sending tool_result back and receiving the final answer
Vision — images as base64 in an image block
PDF documents — document blocks with a base64 payload
stop_sequences containing text
Token counting

What depends on the model¶

Three features are not uniform across models: temperature, extended thinking, and prompt caching. The per-model table is in Feature support by model — measured on this platform rather than copied from a vendor sheet.

The short version:

temperature is rejected by the newest models with temperature is deprecated for this model. Omit it there.
Extended thinking exists everywhere it is supported, but the newest models want {"type": "adaptive"} instead of {"type": "enabled"}.
Prompt caching works on some models, is intermittent on others, and never engages on the rest.

Not currently supported¶

Whitespace-only stop sequences. stop_sequences: ["\n\n"] and ["\n"] are rejected with HTTP 400 on every model:

stop_sequences: each stop sequence must contain non-whitespace

Sequences containing text — "STOP", "Observation:", "\nHuman:" — work normally. Agent frameworks that stop on a blank line need to use a marker with at least one non-whitespace character.

Message Batches. The endpoints do not exist. Requests to /v1/messages/batches return 404.

File uploads. Listing, metadata, download and delete behave as documented. Uploading a document for later reference by file_id is not available yet. Send documents inline instead — a base64 document block, as shown in Feature support by model — which works on every model today.

Prompt caching below the minimum. Anthropic does not cache a prefix shorter than 1024 tokens (2048 for Haiku). Below that, cache_control is accepted and silently ignored — the response is a normal 200 with cache_creation_input_tokens: 0. This is expected behaviour, not a failure.