Claude's reasoning scaled back
An AMD AI director posted that Anthropic reduced Claude’s effective reasoning depth — median 'thinking' tokens fell from about 2,200 to 600 and pre‑edit reads dropped from ~6.6× to ~2× (x.com). The post framed those numeric drops as a capacity tuning that made the model 'dumber' in practice, raising questions about cloud model capacity choices versus local hardware tradeoffs (x.com).
Anthropic’s Claude appears to be spending less time “thinking” on some coding tasks, according to measurements posted by an Advanced Micro Devices artificial intelligence director. (x.com) The post said Claude’s median “thinking” tokens fell from about 2,200 to about 600, while “pre-edit reads” dropped from roughly 6.6 times to about 2 times. A second post said those shifts looked like a capacity tuning change rather than a new model release. (x.com 1) (x.com 2) In large language models, “thinking” tokens are the extra internal steps a model uses before answering, like scratch paper before writing the final line. Anthropic introduced that feature as “extended thinking” in February 2025 and said developers could set a “thinking budget” to control how long Claude spends on a problem. (anthropic.com) Anthropic has kept pushing Claude toward longer, more autonomous coding sessions. Its February 18, 2026 research said the longest Claude Code sessions had nearly doubled in three months, from under 25 minutes to more than 45 minutes, and that software engineering made up nearly 50% of agentic activity on its public application programming interface. (anthropic.com) Claude Code is built to read a codebase, edit files, run tests, and ship changes, so a drop in how much context it rereads can change how carefully it plans multi-file work. Anthropic says the product “reads your codebase, makes changes across files, runs tests, and delivers committed code.” (anthropic.com) Anthropic’s own product line also frames reasoning depth as a selling point. In November 2025, the company said Claude Opus 4.5 handled “complex, multi-system” bugs and “long-horizon, autonomous tasks,” and in February 2026 it said Sonnet 4.6 improved long-context reasoning and agent planning while becoming the default model for Free and Pro users in Claude.ai. (anthropic.com 1) (anthropic.com 2) The tradeoff is cost and capacity. Anthropic’s current model pages still market fine-grained controls for thinking effort, 1 million-token context on newer models, and prompt caching discounts of up to 90%, all features aimed at balancing performance, latency, and price on shared cloud infrastructure. (anthropic.com 1) (anthropic.com 2) Anthropic did not appear to publish a public announcement tied to the specific token and reread changes cited in the posts. Without a company statement, the public evidence so far is limited to user-side measurements and commentary about how Claude behaved in practice. (x.com) (x.com) That leaves a narrower question than “is Claude worse”: how much reasoning Anthropic lets users buy, and how often that budget shifts underneath a cloud product that many developers now use as a daily coding tool. (anthropic.com) (anthropic.com)
