Theo says Opus 5.5 threw out Astra's Rust port, not fixed it
The 10 hour result stands, but the mechanism behind it was a from-scratch rewrite in a new crate.
Theo has corrected the story behind his TypeScript compiler port to Rust. He earlier said Opus 5.5 got the port working in 10 hours after months of stalled attempts, and read that as the model picking up where GPT-6 Astra left off and unblocking whatever Astra was stuck on. That is not what happened.
In an update, Theo says Opus decided all of Astra's code was slop and made an entirely new crate with a from-scratch rewrite. By his count it made more progress in 10 hours than Astra did in two weeks.
It actually decided all of Astra's code was slop, and made an entirely new crate with a from-scratch rewrite.
What the original claim was
Theo had been working on the port on and off for about four months. He reached roughly 35% of tests passing with GPT-5.6 Sol and roughly 85% with GPT-6 Astra, and said both models stalled hard at those numbers and ran in loops with no meaningful progress for days at a time. He then gave Opus 5.5 a single instruction, "/goal finish the port and make it faster", something he said he had never been able to try before because he had never had enough Anthropic tokens for a job that size.
The outcome he reported at the time was the port working in 10 hours followed by 24 hours of grinding on performance. That part of the result is unchanged. What changed is the explanation. The 85% test pass rate Astra reached was not a foundation Opus built on, it was a codebase Opus discarded.
Why the distinction matters
One model resuming another model's half-finished work and pushing it over the line is a claim about handoff between agents, and it would suggest you can bank partial progress from a cheaper or earlier model and pay a stronger one only for the last mile. A model deciding the existing code is not worth keeping and starting again is a different and narrower result. It says the strong model can do the whole job in one pass when you can afford the tokens, and it says nothing encouraging about the value of the two weeks that came before.
It is also a reminder of how easy it is to misread what an agent actually did from the outside. Theo watched the result, not the process, and drew the natural conclusion. The correction only surfaced on inspection.
This is one engineer, one codebase, one port, and Theo is reporting it himself rather than measuring it against anything. Treat the 10 hours as an anecdote with a receipt attached, not a benchmark.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.