Your model didn't get worse — the wrapper around it did (and you can control that)

Short summary

When an LLM seems to regress after an update, the cause is usually not the model weights but the silent wrapper layer — the router, system prompt, reasoning effort, and verbosity caps that change per product without announcement. The author recommends pinning the wrapper by forcing model routes instead of leaving Auto selection, citing OpenAI's GPT-5 router and Anthropic's April 2023 post-mortem as evidence. Real weight changes come with new model names; silent behavior shifts are wrapper-level and controllable.

• •LLM behavior regressions are usually wrapper changes, not weight changes
• •Wrappers (routers, system prompts, reasoning caps) change silently per product
• •Force explicit model routes instead of Auto to prevent quiet model demotion

Generated with AI, which can make mistakes.