Skip to main content

​
Format Conversion

Openference supports three wire formats:
  • OpenAI (chat/completions, responses, embeddings)
  • Anthropic (messages)
  • Gemini (via OpenAI compatibility layer on many routes)

​
Passthrough vs conversion

When the model route is configured with the exact format you send, we pass the body through unchanged (best performance and fidelity). When formats differ, we convert:
  • OpenAI ↔ Anthropic (full messages, tool calls, system prompts, streaming events)
  • Gemini is usually reached via its OpenAI-compatible surface

​
What is preserved

  • Roles and content
  • Tool / function calling roundtrips (definitions, calls, and results — in both streaming and non-streaming responses)
  • System instructions (all of them, joined, mapped to system for Claude)
  • Sampling parameters
  • Streaming event shapes (converted to the client-expected format, with a stable response id, correct finish_reason, and a guaranteed terminal event even if the upstream stream ends abruptly)

​
OpenAI → Anthropic conversion matrix

When you send OpenAI chat-completions requests to a route served by an Anthropic Messages upstream: Dropped (no Anthropic equivalent): frequency_penalty, presence_penalty, seed, logprobs, n, parallel_tool_calls, and inbound assistant reasoning_content (thinking blocks must carry a provider signature to be replayed, which the OpenAI shape does not preserve).

​
Stop reason mapping

​
Usage semantics

Anthropic’s input_tokens excludes cached tokens; OpenAI’s prompt_tokens includes them. Converted responses report prompt_tokens = input_tokens + cache_read_input_tokens + cache_creation_input_tokens, with prompt_tokens_details.cached_tokens carrying the cache-read subset — the same arithmetic OpenAI clients expect.

​
Anthropic → OpenAI conversion

The reverse direction (Anthropic clients on OpenAI-served routes) preserves the same fields, plus thinking blocks round-trip as reasoning_content. Mid-stream Anthropic errors are surfaced as sanitized terminal error frames so SDKs fail cleanly instead of hanging.

​
What to expect

You can use your preferred client SDK and model ID without caring which upstream serves it. If a capability (e.g. vision, specific tool schema) is supported by the upstream model, it will work. See the source of truth for any given model on the Models page.