Theo needles open weight fans over a benchmark re-ranking
A leaderboard reshuffle becomes a punchline, and nobody involved ran any code.
Theo quote-posted a visualization of Artificial Analysis models re-ranking from version 4.3.2 to version 5 with a single line, "Terrible day to be an 'open weight models are just as good as Opus 5.5' guy".
Terrible day to be an “open weight models are just as good as Opus 5.5” guy
That is the whole claim. He names no model, cites no score, and shows no test of his own. The post he is quoting, from @beamsters1, is described as "Artificial Analysis v4.3.2 to v5, AI models re-ranking step by step", a step-by-step animation of how the ordering changes across those benchmark versions.
What is actually being compared
The thing that changed in the source material is the benchmark version, not the models. A re-ranking between 4.3.2 and 5 tells you that the measuring stick moved and that the order shifted with it. It does not tell you that any particular open weight model got worse, or that Opus 5.5 got better, or that the gap on your repository moved by any amount at all. Neither post in this exchange reports a before and after number.
For working engineers, the practical read is the familiar one. A leaderboard position is a summary of someone else's task mix, and when the task mix is revised, positions move for reasons that have nothing to do with how a model behaves inside your agent harness on your codebase. If you picked an open weight model because it held up on your own evals, a version bump on a third party ranking is not new evidence against that choice.
It is also a pattern worth noticing in the feed. This is the same shape as the earlier round of claims that Opus 5.5 had gotten dumber, where the vibe arrived and the test never did. Theo has been loud in the other direction too, telling people not to trust day-one enthusiasm around a launch. The consistent part is the skepticism about other people's evidence. The missing part, here, is his own.
Nobody in this exchange ran anything. If the open weight gap really moved, somebody will have to show a run on real work before it means anything to the people shipping code.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.