Users say Opus 5.5 got quieter. The bench still calls it noise.
Complaints that Claude Opus 5.5 has been nerfed picked up again in late September and were still landing on X on October 4. Anthropic has not confirmed a change. A public tracker says the drop is still inside normal variance.
On September 24, Jared Tate wrote: "Opus 5.5 has been nerfed starting the last few hours today. Autistic Claude speak is back, it's doing really dumb shit & this is not the same model I was using the last 2 days. Confirmed this on multiple accounts, it's turned into an idiot." The post drew about 5,700 likes and more than a million views.
On October 3, Ahmad Osman posted: "The nerfing of Opus 5.5 has already started. You should probably cancel your Claude subscription before it renews." Mitch Malone, who had called Opus 5.5 his daily driver a few days earlier, wrote that it "has been a little nerfed and is now talking garbage more like 5." Haider said it was "dropping tasks, skipping commits, and claiming to have done work it hasn't." One reply under a NerfBench post said a simple refactor burned $75 after the model "dispatched duplicate agents, each overwriting each other's work."
NerfBench, run by Bridgebench, is the experiment most people are citing. On October 1 it scored Opus 5.5 at 103.8% of launch. On October 2 it fell to 94.2%. BridgeMind's line: "94.2% is still inside normal variance, so we can't call it a nerf yet." The project treats 90% to 110% as noise.
Other tests cut the other way. Moki re-ran an Eiffel Tower prompt against the launch-week result and wrote that the new image "actually looks slightly better" and "does not show any obvious signs of a nerf," then added: "One test proves nothing." An account called Claude updates said nearly 10 hours of prompts felt "better than my launch day tests." Dewaldt ran the same prompt twice on October 3 and got opposite impressions: "If the top one proves Opus got nerfed, the bottom one proves it didn't. It proves neither."
Leonel Meque posted two motion-design clips from the same prompt, one from launch week and one from October 4, and asked Anthropic why the second looked worse. Ridham Trivedi wrote that Opus 5.5 was "using more tokens and the output quality has dropped." JP said he had switched his real work to Sonnet 5.5.
The split is the argument. Fixed public prompts are easy to recognize. A lab can train on them, or route traffic that looks like a harness to a stronger serving path, while ordinary sessions hit a cheaper or more restricted one. Public benches with stable prompts make that cheap. Rotating or private tasks are harder to catch.
For anyone using Opus day to day, the bench is not the measurement. The measurement is the workflow.