[Feature]: Support Anthropic adaptive thinking and effort parameter for Opus 4.6

[shaharsha]

Summary

Anthropic's Claude Opus 4.6 introduces adaptive thinking (thinking.type: "adaptive") as the recommended way to use extended thinking. The previous thinking.type: "enabled" + budget_tokens approach is deprecated on Opus 4.6 and will be removed in a future model release.

OpenClaw currently only supports the deprecated format. There is no way to use adaptive thinking or the new effort parameter (low | medium | high | max).

Why it matters

• The deprecated thinking API will eventually break on newer Anthropic models.
• Adaptive thinking lets the model decide when and how deeply to reason, which improves cost efficiency and enables automatic interleaved thinking between tool calls -- especially useful for agentic workflows.
• The effort parameter gives users a cleaner control surface than fixed token budgets.

References

• Adaptive thinking docs
• Effort parameter docs
• Related: [Feature]: Support for Anthropic Opus 4.6 #9811 (add Opus 4.6 to model registry)

[tyler6204]

FYI: Claude Opus 4.6 was added to the built-in model catalog in #9853 (merged). Leaving this issue open since it’s specifically about supporting adaptive thinking + the parameter for Anthropic/Opus 4.6.

[tyler6204]

Clarification: the missing word above was the Anthropic "effort" parameter (low|medium|high|max).

[GodsBoy]

Status Update

I've created PR #10217 that adds documentation and tests for adaptive thinking support.

Current Implementation Status

Good news: Adaptive thinking support is already implemented in @mariozechner/pi-ai 0.52.6:

• For Opus 4.6+: Uses thinking: { type: "adaptive" } with the effort parameter
• For older models: Falls back to budget-based thinking (thinking: { type: "enabled", budget_tokens: N })

OpenClaw's HEAD already has pi-ai 0.52.6 as a dependency, so the feature is available in the repo. However, the current npm release (2026.2.3-1) is on pi-ai 0.51.3 which doesn't have this feature.

What PR #10217 Adds

1. Documentation: Explains adaptive thinking in docs/tools/thinking.md
2. Tests: Verifies the thinking level to effort mapping works correctly

Thinking Level Mapping

OpenClaw Level	Anthropic Effort
minimal	low
low	low
medium	medium
high	high
xhigh	max

Remaining Work

Once PR #10217 is merged, this issue can be closed. Users will get adaptive thinking support automatically when:

1. A new npm release is cut (includes pi-ai 0.52.6+)
2. They use Claude Opus 4.6 (anthropic/claude-opus-4-6)
3. They set any thinking level (/think:high, etc.)

[1kuna]

bump

[Chloecvvvvvv]

Great to see PR #10217 moving forward. The thinking level to effort mapping looks solid.

One thing worth noting for production use: adaptive thinking changes token consumption significantly. With fixed budget_tokens you could predict costs, but with adaptive mode the model decides how much reasoning to do — costs can vary a lot.

Practical tip — combine effort with a max_tokens cap for cost control:

{
  "thinking": { "type": "adaptive" },
  "effort": "high",
  "max_tokens": 16384
}

Also for anyone running Opus 4.6 through OpenRouter or other proxies — make sure your proxy passes through the thinking and effort parameters. Some proxies strip unknown params which silently disables adaptive thinking.

On a related note, AllyHub already handles the adaptive thinking mapping automatically based on task complexity. Still in closed alpha, feel free to DM me if anyone wants an invite code to try it out.

[rdslw]

Sonnet 4.6 also supports adaptive thinking, as described here: https://www.anthropic.com/news/claude-sonnet-4-6