Anthropic

VM-X uses the official @anthropic-ai/sdk to talk to api.anthropic.com. Native HTTP — no OpenAI-compat shim.

Connection config

Field	Required	Description
apiKey	yes	Anthropic API key. Create one at console.anthropic.com → API keys.

{
  "provider": "anthropic",
  "config": { "apiKey": "sk-ant-..." }
}

Endpoint passthrough

Anthropic is the native target for the /v1/messages endpoint — that cell is full input + output passthrough. The other two endpoints each take a single direct hop into Anthropic Messages (no internal ChatCompletion pivot):

Client request shape	What hits the wire
Anthropic Messages	Verbatim. cache_control, extended thinking, server tools, service_tier, top_k, refusal stop details all preserved.
Chat Completions	Direct Chat Completions ↔ Anthropic Messages converter. Cache markers / thinking / top_k / server tools survive via the __vmx_passthrough.anthropic envelope so they're re-attached on the wire body.
Responses	Direct Responses ↔ Anthropic converter (D5 — no ChatCompletion pivot). The Responses event vocabulary maps onto Anthropic SSE events one-to-one for streams.

The gateway's vmx envelope and __vmx_passthrough envelope are stripped before send so Anthropic's strict validator doesn't 400 with vmx: Extra inputs are not permitted.

Capabilities

Capability	Status
Streaming	✅
Function / tool calling	✅
tool_choice: 'none'	✅ — converted to Anthropic {type:'none'}; tools field is preserved
Vision (image content blocks)	✅
Prompt caching (cache_control)	✅ — billing model honours 1.25× / 2× cache-write multipliers (5m / 1h ephemeral)
Extended thinking	✅ — including adaptive thinking and display: 'summarized'
Reasoning tokens reported	✅ — surfaced as reasoning_tokens in usage on converted endpoints
Server tools	✅ — web_search_*, code_execution_*, bash_*, text_editor_*, computer_*
Service tiers	✅ — auto (default) / standard_only
Citations	✅ on the /anthropic/messages endpoint (response forwarded verbatim)
Beta opt-ins (anthropic-beta)	✅ — betas[] on the body lifts to the anthropic-beta HTTP header (T9)

providerArgs — common Anthropic-native fields

When using the Chat Completions / Responses endpoints, these are the escape hatches for Anthropic-native knobs that the OpenAI shape can't express:

{
  "vmx": {
    "providerArgs": {
      "top_k": 10,
      "thinking": { "type": "enabled", "budget_tokens": 5000 }
    }
  }
}

(When you're already on the /anthropic/messages endpoint, just put those fields at the top level of the request body — that's what Anthropic's SDK does natively.)

Audit row fields specific to Anthropic

Audit rows for Anthropic completions populate:

• cached_tokens — prompt cache reads
• cache_creation_input_tokens — total cache writes
• cache_creation_ephemeral5m_tokens / cache_creation_ephemeral1h_tokens — per-TTL breakdown (when reported on the response)
• server_tool_use_*_requests — counts per server tool invocation
• service_tier — the tier Anthropic actually billed at

The cost service applies the published cache-write multipliers (1.25× input rate for 5m ephemeral, 2× for 1h) and subtracts cache- creation tokens from the base prompt billing so you don't pay twice.

Models

Pass any Claude model id Anthropic exposes (claude-haiku-4-5, claude-sonnet-4-6, claude-opus-4-7, …). The default for new connections is claude-haiku-4-5.

Notes

• Beta opt-ins are accepted as a betas: ['…'] array on the body (the cross-format converters stow them via the __vmx_passthrough envelope so they survive Chat Completions / Responses input). Native Anthropic rejects betas as a body field, so the gateway lifts it off the body and emits anthropic-beta: <comma-separated> as an HTTP header at dispatch time. The Bedrock-Invoke variant of this fix uses anthropic_beta as a body field instead — see aws-bedrock-invoke.
• Caller-forwarded headers (T18): allow-listed headers on the client request (anthropic-beta, anthropic-version, …) are forwarded onto the Anthropic SDK call so opt-in headers always reach upstream regardless of which endpoint the body came in on.
• Upstream headers surface on non-streaming responses (T4): request-id and anthropic-ratelimit-* are mapped onto the audit row's x-request-id / x-ratelimit-* keys.
• tool_choice: {type:'none'} strips tools entirely on the cross-format paths — Anthropic's {type:'none'} semantics match OpenAI's "model must not call tools".
• Reasoning ordering (T16): when an OpenAI / Responses caller sends an assistant message that carries a reasoning extension, the converter re-emits the thinking (or redacted_thinking) blocks before the assistant text/tool blocks, matching Anthropic's signed-thinking continuity rules.
• Streaming Anthropic→Responses uses the direct converter; the Anthropic↔ChatCompletion stream converter is used by Gemini / Groq / Perplexity for their Anthropic-input cells (T17) — see the providers index matrix.