Opus 5.5 finished Theo's four month Rust port in 10 hours
Two other models stalled for days at 35% and 85% of tests passing, and one weekly limit covered the run.
Theo says Opus 5.5 got his ts-rust port working in 10 hours, then spent the following 24 hours grinding on performance. He had been at the port on and off for about four months.
I can not believe how quickly Opus unblocked the work Astra was stuck on.
The numbers are the story. He got to roughly 35% of tests passing with GPT-5.6 Sol and roughly 85% with GPT-6 Astra. Both models, he says, stalled hard once they hit those figures and ran in loops with no meaningful progress for days at a time. He handed Opus 5.5 a single instruction, "/goal finish the port and make it faster", and it cleared the wall.
The unsticking claim, not the quality claim
What Theo is describing is not that one model writes prettier Rust than another. It is that two models plateaued in the same way, burning days on a task where the remaining work was the hard part, and a third kept making progress. That is a different axis from benchmark scores, and it is the one that decides whether a long running port is a project or a hobby. In his words the port "may have made this port an actually viable project".
Treat this as one engineer's result on one codebase, because that is what it is. A TypeScript compiler rewritten in Rust is a large, mechanically brutal, very well specified job, which is close to the ideal shape for an agent left running on a goal. Nobody has shown the same pattern holds on vaguer work.
The billing detail
Theo notes he had never had "enough Anthropic tokens" to point a Claude model at a port like this, and that Opus 5.5 now feels practically unlimited to him. The whole run fit inside roughly one weekly limit, with his other accounts still full.
That did not last. A day later he reported killing a Claude account, mostly on Opus, while also doing other work, and updated that to roughly 2.5 accounts down. So the ceiling exists, it just sits somewhere north of rewriting the TypeScript compiler in Rust.
For working engineers the practical read is about the pricing shape rather than the model. A subscription with a weekly cap changes what you are willing to let an agent chew on, because a 34 hour grind that might fail costs nothing extra beyond the cap. Per token billing makes the same experiment a decision you have to justify before you run it, which is why several long shot ports never get attempted at all.