Claude Opus 4.7 sparks power-user backlash

The core problem here is trust, not raw capability. If a model keeps visibly revising itself mid-answer, the benchmark story stops mattering to people paying for predictable output.

• –Anthropic’s release pitch emphasizes stronger coding, better vision, and more consistent agentic workflows, but this user is judging the model on proof quality and response stability.
• –Adaptive thinking can be useful when it allocates compute well; in practice, it can look like indecision, looping, or overthinking if the defaults are off for a given task.
• –Usage limits matter here: every extra correction pass burns tokens, so “better reasoning” can feel like a hidden cost increase on a $20 plan.
• –For research-heavy users, especially in math and physics, coherence and follow-through often matter more than headline benchmark gains.
• –If Anthropic wants this release to land with power users, it needs clearer effort controls and more predictable behavior on complex proofs and multi-step reasoning.