LLM parameter mapping

What it does

One table: each request setting in each of the four formats — OpenAI Chat Completions, OpenAI Responses, Anthropic Messages and Gemini generateContent. Search it by the setting or by a field name you already have, such as max_tokens or stopSequences.

Try it: the output token limit, tool choice, JSON output.

Where the table comes from

It is not written by hand. For each setting, the converter's own writer writes a request once without the setting and once with it, and the cell is the fields that appear. The table therefore says exactly what the LLM request converter does, and it changes when the converter does. A setting a format cannot carry says "Not carried", with the writer's reason, as the converter reports it.

Reading it

• The output limit has four names: max_completion_tokens in Chat Completions, max_output_tokens in Responses, max_tokens in Messages, where it is required, and generationConfig.maxOutputTokens in Gemini.
• Stop sequences are stop in Chat Completions, stop_sequences in Messages and generationConfig.stopSequences in Gemini. Responses has none.
• Tool choice "must call a tool" is required in both OpenAI formats, {"type": "any"} in Messages and ANY in Gemini.
• Parallel tool calls off is parallel_tool_calls: false for OpenAI and disable_parallel_tool_use inside tool_choice for Messages. Gemini has no setting for it.

What it does not know

• A setting's range can differ between providers. For example, Anthropic's Messages API takes a temperature from 0 to 1, and Salyro accepts one from 0 to 2. The table names fields, not their limits: the LLM request linter checks those for a given model.
• Settings that only one provider has, such as a provider's own safety settings, are not in the table. The converter reports them as it meets them.