Parameter mapping

How every OpenAI chat-completions parameter behaves on each provider lane, and how to control dropping with drop_tuning_params.

The gateway speaks the OpenAI chat-completions wire format to every provider. OpenAI-compatible targets (OpenAI, Azure OpenAI, self-hosted vLLM and other custom OpenAI-compatible endpoints) bypass the policy on this page entirely: the request body is forwarded byte-for-byte apart from the gateway-only drop_tuning_params field, which is consumed before forwarding, so whatever the provider accepts, you get. For the translated lanes (Anthropic, Bedrock, Gemini, Vertex) the gateway builds the provider-native request itself, and every parameter follows an explicit policy:

• mapped: translated faithfully to the provider’s native field.
• dropped: a tuning parameter with no equivalent on the lane. By default it is removed, the request proceeds, and the drop is signaled (see below). It is never silently ignored.
• refused: the request depends on the parameter functionally; honoring the request without it would change what you observably get. These always fail with a clear 400 unsupported_parameter error naming the parameter, the lane, and why.

drop_tuning_params

Send "drop_tuning_params": false in the request body to opt into strict mode: any parameter the lane cannot map then refuses instead of dropping. The default is true. The field itself is consumed by the gateway and never reaches a provider.

{
  "model": "anthropic/claude-haiku-4-5",
  "messages": [{"role": "user", "content": "hi"}],
  "seed": 42,
  "drop_tuning_params": false
}

{
  "error": {
    "type": "unsupported_parameter",
    "code": "unsupported_parameter",
    "message": "refusing to drop 'seed' for anthropic/claude-haiku-4-5: drop_tuning_params is false and Anthropic has no sampling seed. Remove it, set drop_tuning_params to true, or use a model that supports it"
  }
}

Drop signals

When a parameter is dropped, the drop is visible in three places:

• Response body: extra_fields.params_dropped lists the dropped parameter names (on streams, on the final usage-bearing chunk).
• Response header: X-LangWatch-Params-Dropped carries the same list, comma separated.
• Trace: the gateway span records langwatch.gateway.params_dropped.

The table

Parameter	anthropic	bedrock	gemini / vertex	Notes
audio	dropped	dropped	dropped
frequency_penalty	dropped	dropped	mapped
function_call	refused	refused	refused
functions	refused	refused	refused
logit_bias	dropped	dropped	dropped
logprobs	refused	refused	mapped
metadata	dropped	dropped	dropped
modalities	dropped	dropped	dropped
n	dropped	dropped	dropped	n: 1 passes; n > 1 is reduced to one completion and signaled
parallel_tool_calls	dropped	dropped	dropped
prediction	dropped	dropped	dropped
presence_penalty	dropped	dropped	mapped
prompt_cache_key	dropped	dropped	dropped
prompt_cache_retention	dropped	dropped	dropped
reasoning_effort	mapped	mapped	mapped	bedrock: mapped for Anthropic (thinking) and Nova; refused for other families; a customer VPC endpoint maps Anthropic only
response_format	mapped	mapped	mapped	anthropic: json_schema enforced; json_object refused (no parseable-JSON guarantee); bedrock: json_schema enforced; json_object refused (no parseable-JSON guarantee); a customer VPC endpoint enforces json_schema for Anthropic models only; gemini: json_object and json_schema both enforced
safety_identifier	dropped	dropped	dropped
seed	dropped	dropped	dropped
service_tier	dropped	dropped	dropped	anthropic: “auto” passes natively; other values dropped; bedrock: dropped (AWS validates a different enum); gemini: dropped
stop	mapped	mapped	mapped
store	dropped	dropped	dropped
stream_options	mapped	mapped	mapped
temperature	mapped	mapped	mapped
tool_choice	mapped	mapped	mapped	anthropic: none/auto/required/named mapped; named must exist in tools; allowed_tools refused; bedrock: none maps by removing tools (Converse has no none mode); named must exist in tools; allowed_tools refused; gemini: none/auto/required/named mapped; named must exist in tools; allowed_tools refused
tools	mapped	mapped	mapped
top_logprobs	refused	refused	mapped
top_p	mapped	mapped	mapped	anthropic: dropped with a signal when temperature is also set (Anthropic models reject the pair); bedrock: dropped with a signal when temperature is also set on Anthropic models (they reject the pair); other families take both
user	dropped	dropped	dropped
verbosity	dropped	dropped	dropped
web_search_options	dropped	dropped	dropped
any other key	dropped	dropped	dropped	vendor params with no mapping on translated lanes

The codex lane

Codex models (your own ChatGPT subscription, model prefix openai_codex/) speak the Responses API. The gateway builds the outgoing body itself, so every provider-bound field you send is mapped, dropped or refused by the table below, exactly as above. The codex backend accepts only a fixed set of fields and answers 400 Unsupported parameter to anything else, which is why every key the table does not list is dropped rather than forwarded. The gateway also pins three invariants the backend requires: the bare model name, stream on, and store off. Dropped fields are signaled on the X-LangWatch-Params-Dropped header and the gateway span; the response body is the backend’s own SSE stream, so extra_fields.params_dropped is not injected into it. drop_tuning_params works on this lane too.

Parameter	codex	Notes
background	refused	the background response id the caller expects to poll would be silently absent
include	mapped
input	mapped
instructions	mapped
max_output_tokens	dropped	the codex backend refuses an output cap
max_tool_calls	refused	the tool-call cap would be silently unenforced
metadata	dropped	the codex backend refuses request metadata
model	mapped	pinned to the bare model name
parallel_tool_calls	mapped
previous_response_id	refused	the codex lane pins store: false, so the response chain the caller asked to continue would silently not continue
prompt_cache_key	mapped
prompt_cache_options	dropped	the codex backend manages its own prompt cache
prompt_cache_retention	dropped	the codex backend manages its own prompt cache
reasoning	mapped
safety_identifier	dropped	the codex backend refuses caller identifiers
service_tier	dropped	the codex backend has no service tiers
store	mapped	pinned off (the backend is stateless)
stream	mapped	pinned on (the backend is SSE-only)
stream_options	mapped
temperature	dropped	the codex backend refuses sampling options
text	mapped
tool_choice	mapped
tools	mapped
top_logprobs	refused	the requested log probabilities would be silently absent from the response
top_p	dropped	the codex backend refuses sampling options
truncation	dropped	the codex backend manages its own context truncation
user	dropped	the codex backend refuses caller identifiers
any other key	dropped	the codex backend accepts a fixed set of fields

reasoning_effort budgets

On lanes that map reasoning_effort to a thinking budget (Anthropic models without adaptive thinking, on both the anthropic and bedrock lanes), the budget is derived from your output cap: budget = 1024 + ratio x (max_tokens - 1024) with ratios minimal 0.025, low 0.15, medium 0.425, high 0.80, xhigh 0.92, max 1.0. Claude Opus and Sonnet 4.6+ use the provider’s adaptive thinking with your effort passed through (minimal maps to low). Gemini thinking models receive the effort natively. Thinking tokens count inside usage.completion_tokens.