Workflow regression is the benchmark nobody can fake. If the tool gets dumber in your actual loop, the chart is already late.

I just use Sonnet 5 and never run out of tokens

Have you tried to daily drive the premier Chinese models?

Model evals need receipts, not vibes. If one client gets dumber and another gets sharper, the scoreboard should be public and brutally boring.

I noticed the same thing with Opus 5, quick answers instead of good research lately. Only after I start pushing it, it will start to work as intended.

Thanks. I've just switched back to Codex now, had been a long time but it clearly works waaaaay better than Fable which was actually pushing me away from using AI at some point. And Kimi was a let-down in a strange, slow and expensive way. Glad to finally code at normal speed with decent quality. ...

The trick is making the model audition for the job, not flatter the author. Huge difference.