Prompt nerfing and parameter lockdown: why AI users feel models got worse
Prompt nerfing is the user-facing feeling that a model or product became less responsive, less direct, or less capable after a change. Sometimes it is real. Sometimes it is an interaction between defaults, parameters, safety behavior, tooling, and expectations.
What changed in the conversation
Community posts around Opus 4.7 include claims that non-default temperature, top_p, or top_k values are being rejected. Those claims need official confirmation, but the user concern is real: when defaults become stricter, expert users can feel like the model has been nerfed.
Prompt nerfing is not one thing
Perceived degradation can come from model routing, safety tuning, system-prompt changes, hidden context, rate-limit pressure, tool failures, or parameter restrictions. A serious team should not rely on vibes. It should rerun representative tasks, compare artifacts, and measure retries, latency, edits, and final quality.
How to test it
- Keep a stable benchmark prompt set.
- Record model, tool, and parameter settings.
- Compare final artifacts, not just subjective feel.
- Separate model quality from agent-harness behavior.
- Track token usage per accepted artifact.
TRH angle
If users feel a model got worse, they often compensate by prompting more, retrying more, and adding more context. That can increase token waste even when the actual root cause is unclear. Token recovery helps turn the complaint into measurable workflow evidence.