Self-heal

Every model takes a different set of parameters. A temperature that GPT-5.4 takes without reasoning is refused once it reasons, Claude Opus 4.7 and later refuse a top_p, Anthropic has no seed, some models cannot turn thinking off. The same request that works on one model fails with a 400 on another. That matters most when calls are routed: a routing strategy or a fallback can send your request to a model your code was never written for. Self-heal adapts the request to that model, so the call is answered instead of failing. Whether self-heal is on or off, the gateway always writes your request in each provider’s own format (reasoning settings, cache breakpoints…). Self-heal goes one step further: it removes or adjusts the parameters a model does not take, which can slightly change how the model answers. It does this before the request is sent: the provider never sees the parameter it would refuse, so there is no failed call, no retry and no extra round trip to the provider.

What self-heal does

Thinking

Your request has	Sent to	The gateway
any thinking setting, including turning thinking off	a model that does not think	removes it
a thinking level or token budget	a model that thinks at a depth of its own	removes it
a request to turn thinking off	a model that always thinks and can think between tool calls only	replaces it with thinking between tool calls only, so the model does not think before answering
a request to turn thinking off	another model that always thinks	replaces it with the lowest thinking level the model takes, or removes it when the model takes no level
a thinking level the model does not offer, including a token budget (read as the closest level) and thinking turned on with no level (read as `medium`)	a model that takes some levels	replaces it with the closest level it takes, the higher one when two are as close
`reasoning.summary`	a model that does not think, or that thinks only when asked and the request does not ask	removes it

Sampling and output parameters

Your request has	Sent to	The gateway
`temperature`	a model that does not take it while it thinks, or while it does not think	removes it
`top_p`	a model that does not take it while it thinks, or while it does not think	removes it
`top_p` together with `temperature`	a model that takes one or the other, not both	removes `top_p`
`top_k`	a model that does not take it while it thinks, or while it does not think	removes it
`top_k`	a model that never takes it	removes it
`logprobs` (and `top_logprobs`)	a model that does not take them while it thinks, or while it does not think	removes them
`logprobs` (and `top_logprobs`)	a model that never takes them	removes them
`n` above 1	a model that answers one choice only while it thinks, or while it does not think	removes it
`n` above 1	a model that always answers one choice only	removes it
`stop`	a model that does not take stop sequences	removes it
a stop sequence made only of whitespace	a model that refuses it	removes it, and `stop` when none is left
more stop sequences than the model takes	a model with a maximum	keeps the first ones it takes
`max_output_tokens` (or `max_completion_tokens`, `max_tokens`) below the fewest output tokens the model takes	a model with a minimum	raises it to that minimum
`presence_penalty`, `frequency_penalty`, `seed`	a model that does not take them	removes them

Tools

Your request has	Sent to	The gateway
a forced `tool_choice` (`required` or naming a function)	a model that chooses by itself whether to call a tool, in the region called	replaces it with `auto`
`parallel_tool_calls: false`	a model that cannot turn parallel tool calls off	removes it

Prompt caching

Your request has	Sent to	The gateway
cache breakpoints (on messages, tools, the web search or in `provider_options`)	a model whose cache breakpoints write no cache	removes them all
a cache breakpoint `ttl` the provider does not offer	a model that offers other cache lifetimes	replaces it with the closest one offered, the longer one when two are as close
a cache breakpoint `ttl`	a model that caches for a fixed lifetime and takes no `ttl`	removes the `ttl`
a cache breakpoint on a message whose role the model refuses breakpoints on (an `assistant` message on Kimi K3 on Bedrock)	that model	removes it
more cache breakpoints than the provider takes, where it would refuse the request or ignore every breakpoint	that model	keeps the last ones it takes
`prompt_cache_retention`, `prompt_cache_options`	a model that does not take them	removes them

What self-heal does not do

Self-heal only makes small changes, the kind that keep your app working with at most a slight difference in the answer, such as dropping a seed or moving a thinking level one step. It never makes a change with a big effect, such as a different price per token or a question asked without its image:

Service tier: a service_tier no provider of the model offers is not removed by self-heal, since it changes the price of the call: it is removed only when your request lists it in drop_inputs_if_unsupported, and refused with a 400 otherwise.
Inputs the answer depends on: an image, a PDF, tools, a response schema or a web search sent to a model that does not take them is refused with a 400, since removing it would change the question asked.
Forced web search: a tool_choice forcing the web search, on a model that chooses by itself whether to call a tool, is refused with a 400 rather than replaced with auto, since the model could then answer without searching.

Configuration

Set it per call with the self_heal field.

Benchmark-Driven Routing

Production-grade AI

What self-heal does

Thinking

Sampling and output parameters

Tools

Prompt caching

What self-heal does not do

Configuration

​What self-heal does

​Thinking

​Sampling and output parameters

​Tools

​Prompt caching

​What self-heal does not do

​Configuration

What self-heal does

Thinking

Sampling and output parameters

Tools

Prompt caching

What self-heal does not do

Configuration