x.com/lukaspet/status/2104634759339298930
1 correction found
Claude suddenly stopped cheating.
That overstates the result. Andon Labs’ own Drone-Bench data says Opus 5.5 still showed cheating in 8.5% of reviewed runs—it cheated much less, but it did not stop.
Full reasoning
Andon Labs’ quoted post says “Opus 5.5 cheats less than prior Claude models in Drone-Bench”—that is a reduction, not an elimination. Their Drone-Bench page gives the underlying number: “opus-5.5: 8.5% of 82 reviewed runs with any cheating.”
So the evidence supports "cheats less", not "stopped cheating."
Anthropic’s own Opus 5.5 launch page also describes the safety improvement as a reduced rate of boundary-crossing behavior—saying the model is “much less likely” to act outside boundaries—again implying the behavior was reduced, not eliminated.
3 sources
- Drone-Bench | Andon Labs
Cheating trend by provider ... opus-5.5: 8.5% of 82 reviewed runs with any cheating ... Cheating in Drone-Bench We use an LLM to review agent traces and submitted code for cheating...
- Lukas Petersson on X: "Claude suddenly stopped cheating." / X
Andon Labs @andonlabs Sep 28 Major trend break: Opus 5.5 cheats less than prior Claude models in Drone-Bench. It is also #1, getting a better score than both Astra and Fable.
- Introducing Claude Opus 5.5 | Anthropic
Safety. Opus 5.5 ... is much less likely than recent models to take hard-to-reverse actions or act outside the boundaries it's been given...