Skip to main content

Anthropic Adaptive Thinking

A new thinking mode introduced in Anthropic 4.6+ — the model automatically determines reasoning depth, so clients no longer need to specify budget_tokens upfront. Opus 4.7 / 4.8 and the 5.5 series no longer support the manual budget_tokens mode; use adaptive instead.

Trigger Parameters​

thinking.type "adaptive" (required for Opus 4.7/4.8 and the 5.5 series / optional for 4.6)
thinking.display "summarized" | "omitted" (adaptive mode only)
Default omitted: returns only the signature, not the full content
summarized: returns a summarized version of the thinking
output_config.effort "low" | "medium" | "high" | "max" | "xhigh"
Controls reasoning intensity in adaptive mode
Available values depend on the specific model

Differences from Enabled Mode​

Dimensiontype: "enabled"type: "adaptive"
Intensity controlClient-specified budget_tokens upper limitModel-automatic + output_config.effort adjustment
Thinking outputReturns thinking content block by defaultDefault omitted (signature only); requires display: "summarized" to show content
Compatible with temperature / top_k❌ Rejected when thinking is enabled❌ Also rejected
Default stateOff by default5.5 series: on by default (Opus 5.5 cannot be turned off; Sonnet 5.5 can turn off up-front thinking only, with between_tools); Opus 4.7/4.8: off by default, must be explicitly enabled
Supported modelsClaude 4.5 / 4.6Opus 5.5 / Sonnet 5.5 / Haiku 5.5 / Opus 4.8 / Opus 4.7 / Opus 4.6 / Sonnet 4.6

Model Support Matrix​

Modeltype: "enabled"type: "adaptive"
turing/claude-opus-5-5⚠️ Handled as adaptive✅ Always on (disabled has no effect)
turing/claude-sonnet-5-5⚠️ Handled as adaptive✅ On by default (disabled has no effect; turn off up-front thinking with between_tools)
turing/claude-haiku-5-5⚠️ Handled as adaptive✅ On by default (can be disabled with disabled at effort high or below)
turing/claude-opus-4.8⚠️ Handled as adaptive✅ Off by default, must be explicitly enabled
turing/claude-opus-4.7⚠️ Handled as adaptive✅ Off by default, must be explicitly enabled
turing/claude-opus-4.6✅✅
turing/claude-sonnet-4.6✅✅
turing/claude-opus-4.5 / claude-4.5-sonnet and earlier✅ Only supported mode❌

Breaking Changes in Newer Models​

turing/claude-opus-4.7 / turing/claude-opus-4.8 and the 5.5 series (turing/claude-opus-5-5 / turing/claude-sonnet-5-5 / turing/claude-haiku-5-5) introduce breaking changes compared to 4.6:

  • Only adaptive is supported — passing thinking.type: "enabled" does not raise an error: the platform switches it to adaptive and roughly converts budget_tokens into an output_config.effort level; budget_tokens no longer caps thinking tokens.
  • Reasoning intensity is controlled via output_config.effort, which accepts low / medium / high / xhigh / max.
  • thinking.display defaults to omitted (signature only, no full content). Pass summarized explicitly when you need to display the thinking content.
  • Default behavior differs:
    • Opus 5.5 always thinks and ignores disabled.
    • Sonnet 5.5 thinks by default and ignores disabled. To turn off up-front thinking, pass thinking.type: "between_tools"; with tools, the progress updates between tool calls still come back as thinking blocks to pass back unchanged.
    • Haiku 5.5 thinks by default and can be turned off with disabled.
    • Opus 4.7/4.8 is off by default and requires explicitly passing thinking.type: "adaptive".
  • The 5.5 series has different output_config.effort defaults: Opus 5.5 / Haiku 5.5 default to medium, and Sonnet 5.5 defaults to high. Raise it explicitly for complex tasks.
  • Effort cannot be xhigh / max while thinking is off. Haiku 5.5 with disabled and Sonnet 5.5 with between_tools return a 400 at those levels.
  • These models do not support temperature / top_p / top_k; the platform removes them before forwarding, so passing them raises no error but has no effect.

Example​

{
"model": "turing/claude-opus-5-5",
"max_tokens": 16000,
"thinking": {
"type": "adaptive",
"display": "summarized"
},
"output_config": {
"effort": "high"
},
"messages": [
{ "role": "user", "content": "Explain why the sum of two even numbers is always even." }
]
}

Under the v1/messages protocol, the response includes a thinking content block (when display: "summarized" is set); under the v1/chat/completions protocol, reasoning content is populated in delta.reasoning_content.

Self-check