I’m not saying this is confirmed, but it would explain a lot of what people are noticing with Codex and ChatGPT lately. A lot of degradation benchmarks seem to use API access, not the subscription product. So when people say “the model hasn’t degraded,” they may only be proving that the API version still performs well. From a cost perspective, it would also make sense. Serving millions of subscription users at a flat monthly price is very different from metered API usage. If the ChatGPT/Code...