Detailed Analysis
A vocal segment of Claude power users is pushing back against Anthropic's rollout of "adaptive thinking" — an automatic, model-controlled reasoning mode — arguing it strips them of granular control over when extended Chain-of-Thought (CoT) reasoning is engaged. The post, representative of broader user frustration surfacing on forums and community threads, centers on a reported degradation in output quality when Claude's thinking is left to the model's own discretion rather than explicitly triggered by the user. The author alleges that adaptive thinking was quietly tested via "Opus 4.6," a designation consistent with iterative internal releases, and that this experiment produced widespread missing thinking blocks and shorter, lower-quality responses — a tradeoff that is unacceptable for professional, multi-step workflows involving studying, planning, creative ideation, and agentic auditing tasks.
The friction here is structural. Anthropic's extended thinking architecture, introduced at scale with Claude 3.7 Sonnet and expanded through Claude 4-series models including Sonnet 4, Sonnet 4.5, and Opus 4, was originally designed with explicit user and developer control in mind. The API allows fine-grained configuration via parameters such as `"thinking": {"type": "enabled", "budget_tokens": 10000}`, with budgets ranging from a minimum of 1,024 tokens up to 128,000 tokens for the most capable models. This design philosophy — giving operators and users the ability to dial reasoning effort precisely — is what made extended thinking a reliable professional tool. The user's complaint that reverting to always-on adaptive inference is "literally 4-5 lines of code" to fix underscores the perceived gap between what the infrastructure already supports and what the consumer-facing claude.ai interface exposes. The author explicitly distinguishes between Claude Code (which retains configurable effort levels) and the claude.ai web app, noting that the latter is preferred for non-coding intellectual work precisely because it avoids the token overhead of agentic scaffolding.
The underlying concern reflects a broader tension in AI product design between model autonomy and user agency. Adaptive thinking, by design, lets the model self-determine when deeper reasoning is warranted — a feature that reduces unnecessary latency and token cost for simple queries. However, for users whose workflows are built on the consistent availability of extended reasoning, model-discretionary activation introduces unpredictability that undermines trust. Benchmarks affirm the performance stakes: extended thinking mode achieves 96.2% on MATH 500 problems and 93.2% on IFEval versus 90.8% in standard mode. When reasoning is suppressed — even adaptively — in contexts where a user knows it is needed, the accuracy gap is not merely theoretical. The author's mention of agentic projects that "need thinking to function properly" points to a class of user whose entire workflow architecture is predicated on reasoning being reliably accessible, not probabilistically available.
This dispute situates Anthropic within a wider industry dynamic that emerged following OpenAI's o1 release in late 2024, which normalized the idea of reasoning models as a distinct product tier. Anthropic's response — integrating extended thinking directly into its flagship models rather than bifurcating into separate reasoning and non-reasoning products — was architecturally elegant but created an interface design challenge: how to surface control of that reasoning to end users in a way that satisfies both casual users (who benefit from adaptive efficiency) and power users (who require deterministic activation). The author's ultimatum framing — treating the loss of toggleable thinking as an "unsubscribe moment" — signals that for a meaningful segment of paying subscribers, the value proposition of Claude is inseparable from its reasoning transparency and user-controllability. Whether Anthropic treats this as a niche complaint or a signal about the professional user segment's needs will likely shape how the claude.ai interface evolves over the remainder of 2026.
Read original article →