Anthropic (Claude) to OpenAI API Converter

The converter is a deterministic restructuring of a published schema — there is no model call. It reads your Anthropic Messages payload, applies a fixed set of mapping rules transcribed from the official Anthropic and OpenAI API references, and emits the equivalent OpenAI Chat Completions body. Every rule is a static table entry, so the same input always produces the same output, and the result is testable. Two directions are supported: a whole request body, and a whole response body.

Request: Anthropic → OpenAI

1. System prompt. Anthropic's top-level system (a string or an array of text blocks) becomes a leading {role:"system"} message, because OpenAI has no top-level system field.
2. Messages & content blocks. Roles pass through. A text block collapses to a string (or a text content part when mixed with images); an image block becomes an image_url part (base64 sources become data: URLs).
3. Tool calls. An assistant tool_use block becomes a tool_calls entry. Anthropic's input is a JSON object; OpenAI's arguments is a JSON string, so the converter runs JSON.stringify(input).
4. Tool results. A user tool_result block becomes its own {role:"tool", tool_call_id, content} message, placed after the assistant message that made the matching call so the turn order stays valid.
5. Tools & tool_choice. tools[].input_schema becomes function.parameters (copied verbatim); {type:auto}→"auto", {type:any}→"required", {type:tool,name}→{type:function,function:{name}}.
6. Scalars. max_tokens is carried over (rename to max_completion_tokens for o-series / GPT-5 targets via the toggle); stop_sequences → stop; temperature, top_p and stream pass through.

Response: Anthropic → OpenAI

1. The message is wrapped as {id, object:"chat.completion", model, choices:[{index:0, message, finish_reason}], usage}.
2. stop_reason → finish_reason: end_turn→stop, max_tokens→length, stop_sequence→stop, tool_use→tool_calls, refusal→content_filter, and pause_turn→stop (flagged approximate).
3. usage: input_tokens→prompt_tokens, output_tokens→completion_tokens, and total_tokens = input_tokens + output_tokens (OpenAI reports the total; Anthropic does not).

As a credibility check, two worked conversions ship as fixtures and are reconciled against the references on every load, and the response direction carries an independent arithmetic invariant — total_tokens must equal prompt_tokens + completion_tokens, recomputed on a separate code path. Anthropic-only features that have no faithful OpenAI target (thinking, cache_control, service_tier) are listed in the notes rather than silently dropped.

• Anthropic API Reference — Messages (request & response fields)
• Anthropic — Tool use (input_schema, tool_use, tool_result, tool_choice)
• OpenAI API Reference — Chat Completions (messages, tools, finish_reason, usage)

The mapping rules were last cross-checked against the official references on 2026-07-11. This page is reviewed whenever either provider ships a breaking request- or response-schema change.