Anthropic Drops Extended Thinking Toggle for Opus and Sonnet 4.6
Key insights
- Anthropic is removing the type: 'enabled' extended thinking parameter for Claude Opus 4.6 and Sonnet 4.6, enforcing adaptive thinking as the only mode.
- Claude Opus 4.7 already returns a 400 error on the deprecated format, confirming there is no backward-compatible fallback path.
- Community data shows Opus 4.6 thinking 67% less since Claude Code defaulted to effort-medium in March 2026.
Why this matters
Developers who built reproducible agentic pipelines around fixed budget_tokens allocations now face a forced migration with no opt-out, breaking any workflow that assumed stable reasoning depth as an invariant. The 67% reduction in observed thinking time is a concrete performance regression for tasks where deep reasoning was load-bearing, not optional, and no official mitigation has been published. More broadly, Anthropic centralizing control over reasoning depth signals that cost efficiency and infrastructure predictability will increasingly override developer configurability as the model generation advances.
Summary
Anthropic is retiring the extended thinking type: 'enabled' API parameter for Claude Opus 4.6 and Sonnet 4.6, making adaptive thinking the mandatory default across both models. The shift is already enforced on Claude Opus 4.7 and later, which return a 400 error on the deprecated format, and the legacy parameter is now being removed from the 4.6 generation as well.
The practical impact lands hardest on Claude Code users. Since Claude Code moved to effort-medium as the default reasoning setting in March 2026, community benchmarking shows Opus 4.6 thinking 67% less than it did under the previous configuration. Developers who set fixed budget_tokens values to get consistent, reproducible reasoning depth in agentic pipelines now face a migration to the adaptive thinking API before the cutoff.
Essentially: (Anthropic, Claude Code users) are on opposite sides of a tradeoff between cost efficiency and reasoning predictability.
- The deprecated type: 'enabled' parameter is already a hard error on Opus 4.7+, signaling no rollback path.
- Fixed budget_tokens allocations used for reproducible agentic workflows are the primary migration target.
- The 67% reduction in observed thinking time reflects both the API change and Claude Code's March 2026 default shift to effort-medium.
The deprecation formalizes Anthropic's preference for dynamic resource allocation over developer-controlled reasoning depth, a tradeoff that will define how agentic workflows are designed going forward.
Potential risks and opportunities
Risks
- Agentic pipeline operators using Opus 4.6 with fixed budget_tokens for legal, financial, or scientific reasoning tasks could see silent quality regressions before they detect the behavioral shift.
- Claude Code extension developers who shipped effort-override configurations to enterprise customers face breaking changes with no vendor-provided migration tooling, creating support liability in the next 30-60 days.
- Teams that benchmarked model selection decisions on Opus 4.6 reasoning depth may now be running a materially weaker configuration than their evaluations assumed, invalidating prior accuracy or reliability claims.
Opportunities
- Observability vendors targeting LLM pipelines (Langfuse, Helicone, Braintrust) can position adaptive-thinking monitoring as a required migration tool for teams auditing reasoning depth regressions.
- Consulting firms and AI engineering shops with Anthropic API expertise can capture near-term demand from enterprises needing to audit and migrate fixed budget_tokens workflows before the deprecation cutoff.
- Competing model providers (OpenAI with o-series, Google with Gemini thinking modes) that offer stable, developer-controlled reasoning budgets gain a concrete differentiation argument for teams prioritizing reproducibility over cost optimization.
What we don't know yet
- No official deprecation timeline or hard cutoff date has been published for the 4.6 parameter removal, leaving migration urgency unclear for production teams.
- Whether the adaptive thinking API can be constrained to match the minimum reasoning depth previously achievable with fixed budget_tokens is not addressed in available documentation.
- Community-measured 67% thinking reduction covers Claude Code workflows specifically; impact on direct API users with custom effort settings has not been independently quantified.
Originally reported by reddit.com
Read the original article →Original headline: Anthropic Deprecating Extended Thinking Toggle for Claude Opus 4.6 and Sonnet 4.6 — Adaptive Thinking Enforced by Default Going Forward