Post by Mohammad Dayem A.

Stella Laurenzo didn't tweet "Claude feels dumber." She ran 6,852 sessions and proved it. 234,760 tool calls. 17,871 thinking blocks. File reads before each edit dropped from 6.6 to 2. Reasoning effort down 67%. Stop-hook violations: zero to roughly ten per day. Her conclusion: "Claude has regressed to the point it cannot be trusted to perform complex engineering." When Opus 4.7 shipped, three things changed without announcement. The budget_tokens parameter now returns a 400 error, replaced by a vague "effort" scale. The tokenizer grew 35%, which means the same prompts now cost 35% more on identical pricing. Extended thinking became hidden by default. Boris Cherny, the Claude Code lead at Anthropic, confirmed it: defaults changed in February and March. If you're running AI in production and the model quietly reduces how much thinking it does, your benchmarks drift silently. Error rates change. Cost models stop matching actuals. You find out when something fails, not when something changed. The fix is one flag: --effort max in the command line, or /effort high in terminal. Thirty seconds. But the flag only helps if you noticed. Most operators won't notice until a client does. What Stella built is what every operator should have before touching a model update: a structured baseline. Not a sense that it used to work better. A number that tells you it did. Every model update should have a baseline before and after. Like a deployment. Like a budget review. If you're running Claude in production today, you need to be wary of these changes by Anthropic.