Anthropic Explains Claude’s “Dumb Phase”
Anthropic has confirmed that recent complaints about Claude’s degraded reasoning were valid, but the cause wasn’t model “nerfing” — it was three product‑layer changes that unintentionally crippled performance. A lowered default reasoning effort, a caching bug that wiped Claude’s short‑term memory every turn, and a restrictive verbosity instruction collectively made the model seem dumber, more repetitive, and less capable. All issues have now been fixed in version v2.1.116, and Anthropic is rolling out new safeguards, internal testing requirements, and expanded evaluation suites to prevent future regressions. The company also reset usage limits to compensate affected subscribers.
→ Read the full story on VentureBeat →
Editorial Note: The recent case of Anthropic and the sudden “degradation” of Claude is a stark reminder that in the world of Large Language Models (LLMs), there is no such thing as a free lunch. While the official explanation points to “technical errors” and “caching bugs”, a harsher economic reality lies between the lines.
As OpenAI pursues an aggressive cash‑burn strategy for market dominance, Anthropic is clearly struggling to balance quality with surging operational costs. Reducing reasoning effort and imposing strict verbosity limits are not just UI optimizations - they are direct levers used to cut expensive GPU compute cycles. Anthropic may have genuinely underestimated the compounded impact of these changes or executed them poorly, but the underlying motivation remains a clear drive to reduce costs.
When user demand reaches a critical peak, even the most promising AI firms face a choice: dilute the product’s intelligence to remain financially viable or risk infrastructure collapse. This incident proves that the professional community is highly sensitive to stealth optimization. In an era where we rely on AI for mission‑critical code and engineering, transparency about the “effort” behind a response is no longer a luxury - it is a prerequisite for trust.