Claude Adaptive Thinking & Effort Control¶

★★★★★ Intermediate

Adaptive thinking lets Claude decide per-turn how much reasoning to use. Shipped February 2026 with Opus 4.6. The March 2026 default effort downgrade caused a measurable regression documented across thousands of real sessions.

How Adaptive Thinking Works¶

thinking: {type: "adaptive"}  →  model decides reasoning depth per request
thinking: {type: "enabled", budget_tokens: N}  →  fixed budget (deprecated on 4.6)

Three layers of control:

Layer	What it controls
`thinking.type`	`adaptive` or `enabled` (fixed, deprecated on 4.6)
`output_config.effort`	Guidance level: `max`, `high`, `medium`, `low`
`budget_tokens`	Exact token cap (legacy, deprecated on 4.6 models)

Effort behavior:

Level	Behavior
`max`	Always thinks, no constraints. Opus 4.6 / Sonnet 4.6 only
`high`	Always thinks, deep reasoning (default pre-March 3, 2026)
`medium`	Moderate thinking, may skip for "simple" queries (default post-March 3)
`low`	Minimal thinking

The March 3, 2026 change: default silently lowered from high to medium (effort level 85). No changelog entry. Confirmed post-hoc by Boris Cherny (Anthropic).

Quantified Regression (March 2026)¶

From 6,852 Claude Code session logs, 234,760 tool calls:

Metric	Before (Jan 30 - Feb 12)	After (Mar 8+)	Delta
Median thinking depth	~2,200 chars	~600 chars	-73%
Reads per edit	6.6	2.0	-70%
Edits without prior Read	6.2%	33.7%	+5.4x
Stop-hook violations	0 total	173	-
User interrupts per 1K tool calls	0.9	11.4	+12.7x
Reasoning loops per 1K tool calls	8.2	21.0	+2.6x
Estimated monthly API cost	$345	$42,121	+122x

Cache invalidation bugs (v2.1.100): Two separate bugs also inflated token costs: 1. Billing-sentinel string-replacement breaks cache prefix → full context re-billed 2. Resume/continue tool-injection at wrong position → cache invalidated on every resume

Typical cost inflation from cache bugs: 10-20x per session with resume.

Affected Models¶

Model	Adaptive thinking	`enabled` + `budget_tokens`
Claude Opus 4.6	Default mode	Deprecated (still works)
Claude Sonnet 4.6	Supported	Deprecated
Claude Mythos Preview	Only mode - cannot disable	Not supported
Claude Opus 4.5 / Sonnet 4.5	Not supported	Still works

Restoring Reasoning Depth¶

Claude Code (interactive)¶

/effort max         # session-only
/effort             # check current level

Environment variables (permanent)¶

# Linux/macOS
export CLAUDE_CODE_EFFORT_LEVEL=max
export CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1

# Windows PowerShell
$env:CLAUDE_CODE_EFFORT_LEVEL="max"
$env:CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING="1"

settings.json¶

{
  "env": {
    "CLAUDE_CODE_EFFORT_LEVEL": "max",
    "CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING": "1"
  }
}

API (Opus 4.6 / Sonnet 4.6)¶

{
  "model": "claude-opus-4-6",
  "thinking": {"type": "adaptive"},
  "output_config": {"effort": "max"},
  "max_tokens": 32000
}

Legacy path (deprecated, still functional):

{
  "thinking": {"type": "enabled", "budget_tokens": 31999}
}

Hybrid Model Stack (Cost Mitigation)¶

Task	Tool	Rationale
Architecture, security review, complex debugging	Claude Opus 4.6 + effort=max	Quality gap still significant
Bulk generation, routine refactoring	DeepSeek V3.2	~20x cheaper, ~90% quality
Long context, terminal tasks	Qwen 3.6-Plus	1M context, strong on terminal benchmarks
Batch dataset processing	DeepSeek V3.2	Cost-effective at scale

Gotchas¶

MAX_THINKING_TOKENS makes every request a thinking request (issue #5257). Even "what time is it" gets a thinking block. Check for this env var before troubleshooting unexplained costs.
ultrathink keyword conflicts with MAX_THINKING_TOKENS (issue #18072): the keyword is ignored when the env var is set.
Raising max_tokens alone does nothing. Adaptive mode won't use the extra space - the bottleneck is the effort level, not the output ceiling.
Downgrading Claude Code to v2.1.87 avoids the v2.1.100 cache invalidation bugs, but also loses features. Evaluate the trade-off.
budget_tokens is deprecated on 4.6 models and will be removed in a future release. Migrate to output_config: {effort: ...} now.
CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1 increases cost 3-5x vs medium default. Budget accordingly.