Effort levels

Reasoning models expose a "how hard should I think" dial, and every API spells it differently. OpenGateway normalizes a single internal enum and translates it to whatever the upstream model expects — regardless of which inbound API you used.

The normalized enum

text

off · minimal · low · medium · high · max

The default is roughly medium. If the upstream model does not support an effort knob, the gateway drops it silently — it never hard-fails a coding client over an unsupported parameter.

How each API signals effort

Tier	OpenAI Chat	OpenAI Responses	Anthropic (4.6+)
Off	`reasoning_effort: "none"`	`reasoning.effort: "none"`	omit `thinking`
Minimal	`"minimal"`	`"minimal"`	`output_config.effort: "low"`
Low	`"low"`	`"low"`	`"low"`
Medium	`"medium"`	`"medium"`	`"medium"`
High	`"high"`	`"high"`	`"high"`
Max	`"xhigh"`	`"xhigh"`	`"xhigh"` / `"max"`

Legacy thinking is deprecated

Anthropic's legacy thinking: {type:"enabled", budget_tokens} is deprecated. Modern clients send thinking: {type:"adaptive"} with output_config.effort. OpenGateway accepts both inbound and buckets the legacy form into the normalized enum.

Setting effort

typescript · openai chat

await client.chat.completions.create({
  model: "claude-sonnet-4-7",
  messages: [{ role: "user", content: "Find the bug." }],
  reasoning_effort: "high",
});

typescript · openai responses

await client.responses.create({
  model: "claude-haiku-4-8",
  input: "Diagnose this stack trace.",
  reasoning: { effort: "high" },
});

typescript · anthropic messages

await client.messages.create({
  model: "claude-sonnet-4-7",
  max_tokens: 1024,
  thinking: { type: "adaptive" },
  output_config: { effort: "high" },
  messages: [{ role: "user", content: "Review this diff." }],
});

Cross-API translation

The point of the normalized enum is that effort survives a protocol mismatch between the inbound API and the upstream model:

Claude Code (Anthropic thinking) → routed to an OpenAI-wire model ⇒ emitted as reasoning_effort.
Codex (OpenAI Responses reasoning.effort) → routed to an Anthropic-wire model ⇒ emitted as thinking / output_config.effort.

The branded SDK exposes a single effort option that maps to whichever wire your call uses.