A Claude Opus 5.5 case · #9
Boris Cherny · Anthropic · X ·
The proof process found concurrency bugs that are hard for people to spot, and a few short requests produced 16 code fixes.
#9 · Hours-long jobs finished alone
Claude Agent SDK, a development tool for building agents, had its code formally verified in Lean. This method checks mathematically whether code behaves as designed.
The proof process found concurrency bugs that are hard for people to spot, and a few short requests produced 16 code fixes.
#10 · Hours-long jobs finished alone
Fifty Opus 5.5 agents were launched at once and put to work for 2 hours 27 minutes. The weekly usage limit went down by only about 18%.
Work that used up the limit with just 5 agents on Fable 5.1 ran at ten times the scale.
#11 · Hours-long jobs finished alone
Given a single prompt and a city-building game to make, it worked without a break for about 1 hour 30 minutes.
With no human stepping in, it planned, built and fixed things on its own for over an hour.
#12 · Hours-long jobs finished alone
The same Excel model for analyzing a corporate merger was given to both Opus 5 and Opus 5.5. Opus 5.5 took 63 minutes. Opus 5 took 93 minutes.
It produces real-world finance deliverables more thoroughly, faster and at half the cost.