Reasoning effort

Effort is a dialect-neutral hint for how much reasoning a model should spend. The value is carried by Sampling and translated by each codec.

API surface

type Effort string

const (
	EffortNone    Effort = ""
	EffortMinimal Effort = "minimal"
	EffortLow     Effort = "low"
	EffortMedium  Effort = "medium"
	EffortHigh    Effort = "high"
	EffortXHigh   Effort = "xhigh"
	EffortMax     Effort = "max"
)

func (e Effort) Valid() bool

EffortNone is a valid unset value. An arbitrary string is not valid according to Valid and should be rejected by a caller before it reaches a codec.

sampling := model.Sampling{Effort: model.EffortHigh}
if !sampling.Effort.Valid() {
	panic("unknown effort")
}

Not every provider and model supports every effort level. The OpenAI Chat Completions codec maps effort to reasoning_effort and the Responses codec to reasoning.effort, and both preserve each level exactly, though Chat Completions sends the field whenever an effort is set while Responses emits it only for a model whose Capabilities.Thinking is set. The Anthropic codec picks between two request shapes from the model’s declared Capabilities.ThinkingDialect, so effort alone is not enough there. Gemini rejects xhigh and max with UnsupportedEffortError, and Bedrock Converse rejects every non-empty level with the same error. A codec that gates on Capabilities.Thinking omits the reasoning request entirely for a model that does not advertise it, rather than reporting an error.

Effort alone is therefore not enough for a provider with more than one reasoning request shape. An Anthropic-family model must also declare its dialect with model.WithThinkingDialect(...), which sets Thinking for you; model.WithThinking() leaves the dialect undeclared, and the Anthropic codec then fails closed rather than guessing.

Proof

• Source: inference/model/effort.go, inference/model/sampling.go
• Tests: inference/model/effort_test.go

Related: Sampling, Capabilities.