Anthropic deprecated manual extended thinking budgets for claude-opus-4-7" class="entity-chip">Claude Opus 4.7, returning a 400 error for fixed budget requests. Opus 4.6 and Sonnet 4.6 still support manual configuration but it is deprecated and will be removed in a future release.
Key facts
- Opus 4.7 returns 400 error for fixed thinking budget requests.
- Opus 4.6 and Sonnet 4.6 manual thinking deprecated, removal planned.
- Users report quality drop after adaptive thinking became default.
- Anthropic has not published benchmarks comparing adaptive vs fixed thinking.
- Fourth API/subscription change in six weeks from Anthropic.
Anthropic has deprecated manual extended thinking budgets for its latest Claude models, forcing adaptive thinking as the default. The change, documented on the Anthropic platform page, means users can no longer specify a fixed token budget for reasoning via thinking: {type: "enabled", budget_tokens: N}. For Claude Opus 4.7, such requests now return a 400 error. Opus 4.6 and Sonnet 4.6 still accept the parameter but the company warns it is deprecated and will be removed in a future release. [According to Anthropic's documentation] The recommended approach is adaptive thinking with the effort parameter: thinking: {type: "adaptive"}.
The User Backlash
Users on Reddit have reacted negatively, particularly those using Claude Code for complex agentic workflows. One user, CaffeineBrogrammer, reported that disabling adaptive thinking in Claude Code restored quality levels that had dropped after the feature was introduced. They posed three pointed questions: if adaptive thinking improves performance as Anthropic claims, why did quality drop when it was enabled? Why have no benchmarks shown better results with adaptive thinking? And why are users who pay API rates — where Anthropic profits from higher token usage — being prevented from using a fixed, higher budget? [According to Reddit user CaffeineBrogrammer]
The Structural Tension
This move creates a contradiction for Anthropic's API business. API users pay per token; a fixed extended thinking budget means more tokens consumed and more revenue for Anthropic. By removing that option, the company is effectively turning down paying customers who want to spend more for higher quality. The only plausible explanations are: (1) adaptive thinking actually produces better results on average, making fixed budgets unnecessary for most use cases; (2) compute constraints force Anthropic to optimize inference efficiency at the expense of user choice; or (3) the company is prioritizing subscription margins over API revenue. The Reddit thread leans toward explanation (2), citing the broader pattern of Anthropic tightening subscription usage with the Agent SDK credit split announced May 13. [According to the Reddit thread and devtoolpicks.com analysis]
What the Benchmarks Don't Show
Anthropic has not published comparative benchmarks showing adaptive thinking outperforming fixed budgets. The company's documentation simply states adaptive thinking is "recommended" without citing metrics. Users report that for tasks requiring deep reasoning — like complex code generation or multi-step agent workflows — fixed budgets often produced more reliable outputs. The lack of transparent evaluation erodes trust, especially given the one-way door: users cannot opt back into fixed budgets even if they see quality degradation. [Per user reports on Reddit]
The Broader Pattern
This is the fourth subscription or API change from Anthropic in six weeks: the third-party agent ban on April 4, the brief Claude Code removal from Pro on April 21, the Agent SDK credit split on May 13, and now the forced adaptive thinking. Together, they signal a company tightening control over how its models are used, likely driven by compute cost pressures. The agentic workflows that made Claude Code popular are also the most expensive to serve, and Anthropic is systematically closing the arbitrage between subscription pricing and actual compute cost. [According to devtoolpicks.com analysis]
What to watch
Watch for Anthropic to publish benchmark results comparing adaptive thinking against fixed budgets, or for a backlash-driven reversal on Opus 4.6/Sonnet 4.6 deprecation timeline. Also watch for OpenAI to capitalize by offering fixed reasoning budgets as a differentiator.
