Why Was Opus 4.5 a Dividing Line for Coding Agents?

David Heinemeier Hansson calls November 24, 2025 the dividing line, because Opus 4.5 returned work uncannily close to code he would have written himself. He did not try it on the 24th. He tried it on the 26th, gave it a couple of tasks, and leaned back. The summer before had been an autocomplete mess, and he is not sure Opus 4.5 was much smarter than Opus 4. What changed was the harness: it could instrument a computer, use tools, and check its own work.

People at Shopify and at other large companies hit the same shock over the Christmas break, once they had a couple of weeks to try it. By December he was already asking whether the next task would work, and it did. He remembers thinking that if Opus 4.5 were the last model, he could live with it for 20 years. The path was recognizable, the code was something he wanted to merge, and Ruby came out looking beautiful. Rust was a different question.

Early spring brought sub-agents and harnesses that split a task. Work that used to take Opus a long time dropped to a fifth of the time, or a tenth, because eight sub-agents could run at once. Even then he still had to drive. He named the destination, steered when it wandered, and audited what came back.