Detailed Analysis
Anthropic appears to have quietly rolled out a behavioral change to Claude in which the model engages in some degree of internal reasoning or "thinking" even when the dedicated adaptive thinking toggle is disabled. The observation, surfaced by a user on what appears to be a social platform, is accompanied by a screenshot suggesting that Claude's thinking process is now visibly occurring or consuming tokens outside of the explicitly activated extended reasoning mode. The precise scope of the change — whether it applies across all Claude tiers, specific API configurations, or only certain interfaces — is not detailed in the post.
The significance of this development, if confirmed, lies in how it alters the token economy for Claude users and developers. Extended thinking in large language models is computationally expensive, as the model generates intermediate reasoning steps that consume input/output tokens before producing a final response. If Claude is now performing reasoning steps even in standard mode, users on consumption-based API pricing could see increased costs without having deliberately opted into the feature. The author's sardonic reference to "selling your soul" for subscriptions reflects a real concern among developers and power users about token bloat and the opacity of billing implications when model behavior shifts without explicit user action.
This development fits within a broader trend of AI providers deepening the default capabilities of their models in ways that blur the line between standard and premium features. Anthropic has been expanding Claude's reasoning capabilities significantly, with extended thinking having been introduced as an opt-in feature for tasks requiring deeper analysis. Gradually normalizing reasoning behavior — even at a reduced level — in standard interactions mirrors strategies seen across the industry, where differentiation between model tiers becomes less about discrete feature gates and more about degree and depth of capability. The move also reflects competitive pressure from models like OpenAI's o-series, which made chain-of-thought reasoning a central product identity.
The casual and humorous framing of the post, combined with the absence of an official Anthropic announcement, suggests this may be an undocumented or incremental change rather than a formal product release. Anthropic has historically made iterative updates to Claude's behavior without detailed changelogs, which can create friction for developers who rely on predictable model behavior and cost structures. As reasoning-capable models become the norm rather than the exception, transparency around when and how thinking processes are invoked will likely become an increasingly important expectation from enterprise and developer communities.
Read original article →