Formats

The same models answer in three request formats. Use whichever your tool already speaks; the key and the billing are the same.

Format	Path	Who sends it
OpenAI chat completions	`POST /v1/chat/completions`	opencode, Hermes Agent, SillyTavern, the OpenAI SDKs
Anthropic messages	`POST /v1/messages`	Claude Code, the Anthropic SDKs
OpenAI responses	`POST /v1/responses`	Codex and tools built on the Responses API

Chat completions

Streaming is server-sent events with "stream": true, ending in [DONE]. Usage arrives on streams without you asking for it.
The output limit can be sent as max_tokens, max_completion_tokens or max_output_tokens.
Images go in as image_url parts, as a URL or a data URI, on models that list supports_vision.
reasoning_effort is moved to the nearest value the model supports. Reasoning text comes back on the message as reasoning or reasoning_content.
A developer message is treated as system.
These fields are accepted and ignored: n, best_of, service_tier, provider, route, transforms, plugins, extra_body.

Mapped: system, max_tokens, temperature, top_p, stop_sequences, your own tools, tool_choice, image blocks and document blocks.
Dropped: thinking, top_k, metadata and Anthropic's server tools.
cache_control is kept. A system prompt of 4,000 characters or more with no breakpoint is marked for caching for you.
Usage reports input_tokens, output_tokens and cache_read_input_tokens.
POST /v1/messages/count_tokens returns an estimate and spends nothing.

Mapped: input, instructions, max_output_tokens, temperature, top_p, stream, function tools, tool_choice, text.format and image input.
previous_response_id is refused with a 400: send the conversation each time. Hosted tools are dropped.