## Ask HN: Agents get dumber before release of new model version? Yes, this is a real phenomenon that has been discussed extensively on Hacker News and confirmed by Anthropic's engineering team. Here's what I found: ### **The Main Hacker News Thread** **Title:** "I've always seen people complaining about model getting dumber just before the new one drops" **Link:** https://news.ycombinator.com/item?id=47797110 **Key observations from the thread:** - **User gck1** reported that Claude Opus 4.6 was "acting like it was Sonnet 2" just hours before the 4.7 release - very verbose, extremely fast, and "dumb" - **Multiple theories** were proposed for why this happens: 1. **Compute reallocation**: VRAM being freed up for new model staging causes aggressive quantization of current models 2. **System prompt bloat**: Accumulation of safety/behavioral instructions that create "cognitive burden" 3. **Infrastructure changes**: Taking machines off production stack for staging new models ### **Anthropic's Official Response** **Engineering postmortem:** https://www.anthropic.com/engineering/april-23-postmortem Anthropic confirmed **three specific issues** that caused perceived degradation before Opus 4.7 release: 1. **March 4**: Changed default reasoning effort from "high" to "medium" to reduce latency (reverted April 7) 2. **March 26**: Bug in caching optimization that cleared thinking history every turn instead of once (fixed April 10) 3. **April 16**: System prompt instruction to reduce verbosity that hurt coding quality (reverted April 20) **Key quote:** "We take reports about degradation very seriously. We never intentionally degrade our models... All three issues have now been resolved as of April 20." ### **Legal/Consumer Protection Discussion** Users debated whether this practice is legal. Some argued that: - Model names are "marketing product names" not guarantees of specific weights - Terms of service don't promise capability or performance - Consumer protection laws may still apply to "material information" about products ### **Bottom Line** The phenomenon is real and documented, but Anthropic attributes it to **unintentional engineering changes** rather than deliberate degradation. The timing coincidence with new releases appears to be due to infrastructure reallocation and internal testing preparations.