The Manufactured-Problem Cycle in Frontier Model Releases

The Claude Opus 4.7 release on April 16, 2026 shipped with a striking property: nearly every documented complaint about Claude Opus 4.6 had a directly matching fix in 4.7. X-high effort answered the silently-lowered effort default. The /ultrareview command answered complaints about shallow review passes. Long-context retrieval was 'completely overhauled' just after users complained about retrieval regressions. The one-to-one correspondence raises a question about cause and effect. One reading is the charitable one: real customer feedback drove a real prioritization cycle, and Anthropic executed well. Another reading, articulated by Nate Herk: 'every complaint has a directly correlating matching fix... it's almost like they're creating holes just so they can fill them and look like the hero.' Independent of intent, the commercial dynamic is concerning. If a vendor benefits from product regressions (because the recovery release becomes a marketing event), the incentive structure rewards manufactured problems. This is structurally similar to planned obsolescence in consumer hardware, except the regression is invisible — there is no physical wear, just a configuration change pushed silently to production. The contrarian counterweight comes from Nick Saraev: 'AI doesn't make things possible anymore, just slightly more profitable.' His argument is that GPT-3 in 2020 was the last true zero-to-one moment in LLMs, and that 3-4% benchmark jumps between releases — whether genuine improvement or staged recovery — do not justify flipping infrastructure each cycle. Personality and behavior differences between models matter more than headline benchmarks for most production work. See Anthropic vs OpenClaw: The Third-Party Harness Lockout of April 2026 for a related Anthropic commercial-strategy episode.