Theo ends the week calling Opus 5.5 the greatest model ever
The nerf argument closes out with one side on feel and the other on a benchmark Theo himself called the worst of all time.
Theo ended the week on Opus 5.5 with a verdict that is hard to misread. "Opus 5.5 is the greatest model of all time and I'm still struggling to find the ceiling," he posted. That lands days after Thariq asked, in a reply to Lydia Hallie, whether it was just him or Opus 5.5 is dumber, a question we covered earlier when nobody attached a test to it.
Opus 5.5 is the greatest model of all time and I'm still struggling to find the ceiling.
The only number anyone brought is NerfBench, which tracks whether a model degrades after release. Theo had already cited it against the dumber claim, noting it showed Opus 5.5 scoring 5.8% smarter since launch, and then immediately turned on it.
The benchmark nobody defends
"This might legitimately be the worst benchmark of all time," Theo wrote in the same post. So the evidence on the pro-model side is a measurement its own loudest citer calls garbage, and the evidence on the degradation side is a feeling that several people share and none of them have reproduced in public.
This might legitimately be the worst benchmark of all time
That is the whole shape of the argument. One engineer with strong opinions about model quality says the ceiling is still out of reach after sustained use. Another engineer says it feels worse. Neither posted a run you can check.
For working engineers
The nerf question comes back every few weeks and it has never once been settled by anything other than vibes, because the thing people are actually measuring is their own prompting drift, their context window, and what they happened to ask that day. A degradation claim is only worth acting on if someone reruns the same task against the same harness and shows the diff. Nobody in this thread did, and the one benchmark that exists to answer the question is being treated as a punchline by the person quoting it.
If you felt Opus 5.5 got worse this week, you are in company, and that is all you are in. There is no test behind the feeling yet, and until there is, treat both the complaint and the greatest-of-all-time call as what they are, which is two experienced users reporting different impressions of the same model.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.