Detailed Analysis
A Reddit post captures a user's frustration after discovering that Claude Opus's extended thinking output had seemingly changed overnight from detailed, verbose reasoning traces to condensed, summarized thinking blocks—a style the poster derisively compares to OpenAI's approach to displaying model reasoning. The complaint centers on a perceived sudden shift in how much of the model's internal chain-of-thought is exposed to the user, with the poster noting that the previous format allowed them to catch inconsistencies, spot instances of the model being "lazy," and simply enjoy reading the more expansive reasoning process. The new format, by contrast, is described as a "dry ass summary line" that strips away the texture and utility of the original thinking blocks.
This complaint sits at the intersection of a broader tension in how AI labs handle the visibility of model reasoning. Anthropic's Claude models, particularly in their "extended thinking" or reasoning modes, have historically exposed relatively raw chain-of-thought text to users, which power users have come to rely on as a diagnostic tool—not just for transparency, but for actively debugging model behavior, catching hallucinations mid-stream, and understanding why a response took a particular shape. OpenAI, by contrast, has generally opted to summarize or obscure raw reasoning traces in products like o1 and o3, partly citing safety and competitive concerns (preventing distillation of reasoning patterns by competitors) and partly to present a cleaner, more polished user experience. When users perceive Claude moving toward the OpenAI model of summarized reasoning, it reads as a regression in transparency rather than a product improvement, especially for a user base that specifically gravitated toward Claude for its more legible internal reasoning.
The reaction also reflects how sensitive power users are to any change in model behavior, even absent official communication from Anthropic. Sudden shifts in output style—whether due to A/B testing, backend prompt changes, quiet model updates, or infrastructure adjustments—are common in production LLM systems, but they land differently with users who have built workflows around specific behavioral quirks of a model version. The absence of changelog transparency or an official acknowardgment (evident in the "research context" turning up nothing) compounds the frustration, since users are left to speculate whether this is an intentional product decision, a cost-saving measure (shorter thinking traces reduce token usage and inference cost), a safety-motivated filtering of reasoning content, or simply a bug.
More broadly, this incident is illustrative of a recurring friction point in the AI industry: the tension between exposing raw model cognition (which builds trust and enables debugging but risks revealing unsafe, embarrassing, or exploitable reasoning) versus presenting sanitized, user-friendly summaries (which protect the company and simplify the interface but sacrifice depth and trust for technical users). As reasoning models become more central to coding, research, and agentic workflows, the format and fidelity of exposed thinking will likely remain a contested design decision, with labs like Anthropic needing to balance the demands of its most technical, high-engagement users against broader considerations of cost, safety, and IP protection around chain-of-thought reasoning.
Read original article →