For me, the “watering down” began with Sonnet 4 and GPT-4o. I think we were at peak capability when we had:
- Sonnet 3.7 (with thinking) – best all-purpose model for code and reasoning
- Sonnet 3.5 – unmatched at pattern matching
- GPT-4 – most versatile overall
- GPT-4.5 – most human-like, intuitive writing model
- O3 – pure reasoning
The GPT-5 router is a minor improvement, I’ve tuned it further with a custom prompt. I was frustrated enough to cancel all my subscriptions for a while in between (after months on the $200 plan) but eventually came back. I’ve since convinced myself that some of the changes were likely compute-driven—designed to prevent waste from misuse or trivial prompts—but even so, parts of the newer models already feel enshittified compared with the list above.
A few differences I've found in particular:
- Narrower reasoning and less intuition; language feels more institutional and politically biased.
- Weaker grasp of non-idiomatic English.
- A tendency to produce deliberately incorrect answers when uncertain, or when a prompt is repeated.
- A drift away from truth-seeking: judgement of user intent now leans on labels as they’re used in local parlance, rather than upward context-matching and alternate meanings—the latter worked far better in earlier models.
- A new fondness for flowery adjectives. Sonnet 3.7 never told me my code was “production-ready” or “beautiful.” Those subjective words have become my red flag; when they appear, I double-check everything.
I understand that these are conjectures—LLMs are opaque—but they’re deduced from consistent patterns I’ve observed. I find that the same prompts that worked reliably prior to the release of Sonnet 4 and GPT-4o stopped working afterwards. Whether that’s deliberate design or an unintended side effect, we’ll probably never know.