GPT-6 Sol dropped an hour after Opus 5.5, so it got the same 9 real Windows tasks.
Same code, same machine, judged on the files on disk: Opus 5: 9/9 · 84 steps Opus 5.5: 7/9 · 109 steps GPT-6 Sol: 7/9 · 150 steps One run each. Clip: Opus 5.5 vs Sol, same task, real time.
I installed opus 5.5 into a fly.
It turned 10 sol into 31.6 sol. Astra 6 in the fly next to it went to 2.9. same body. same feed. same 10 SOL. only the brain was different. every fly got two buttons: → RED: ape the ping right now → GREEN: wait for the rule ASTRA went first.
One early tester finished a 680,000-line code migration with Claude Opus 5.5 in less than a day
Work Anthropic says would have taken an engineering team weeks. And on typical workloads it costs 40% less to run than Opus 5.
Claude has released Opus 5.5.
Claudeai I swear your old models get dumber once you launch a new model Sonnet 5 used to be on fire helping me code with little to no mistakes Now give a very small task and he would f things up unless you use a "better" model
You have reached the end of the archive
All of Claude Opus 5