Same experience here. I even built a Gem with am elaborate prompt instructing it how to be concise, but it still gives annoying long-winded responses and frequently expands the scope of its answer far beyond the prompt.
The same happened with GPT-3.5. It was so good early on and got worse as OpenAI began to cut costs. I feel like when GPT-4.1 was cloaked as Optimus on Openrouter, it was really good, but once it launched, it also got worse.
LLMs, on the other hand, operate under different incentives. It’s in a company’s best interest to initially release the strongest model, top the benchmarks, and then quietly degrade performance over time. Unlike traditional software, LLMs have low switching costs, users can easily jump to a better alternative. That makes it more tempting for companies to conceal model downgrades to prevent user churn.
Counterexample: 99% of average Joes have no idea how incredibly enshittified Google Maps has become, to just name one app. These companies intentionally boil the frog very slowly, and most people are incredibly bad at noticing gradual changes (see global warming).
Sure, they could know by comparing, but you could also know whether models are changing behind the scenes by having sets of evals.
We can tell it’s getting worse because of UI changes, slower load times, and more ads. The signs are visible.
With LLMs, it’s different. There are no clear cues when quality drops. If responses seem off, users often blame their own prompts. That makes it easier for companies to quietly lower performance.
That said, many of us on HN use LLMs mainly for coding, so we can tell when things get worse.
Both cases involve the “boiling frog” effect, but with LLMs, users can easily jump to another pot. With traditional software, switching is much harder.