One user questions benchmarks as a measure of peak AI capability, arguing that AGENTS.md instructions can have enormous effects on less polished models. They say V4.1 does “well” with minimal setups and prompts, but scenario-specific AGENTS.md optimization can produce “massive gains.”