Overnight laptop runs took over most of one Rust developer's Opus coding work

"Flash Next does not decode faster, it just doesn't overthink as much as 27b," the author wrote ^[11]. Decode rate stayed roughly the same between the two models and prefill stayed slow; total time fell because the model emitted fewer thinking tokens ^[12].

At 10 to 15 tokens per second ^[5], an eight-hour overnight run tops out at roughly 288,000 to 432,000 output tokens ^[16]. That is the budget for the whole feature, and any reasoning tokens the model spends re-deriving the plan come out of the same pool as the diff.

The parity claim rests on one number from someone else's harness: 34 for the 27b at the xhigh setting against 32 for Opus 4.6 on the Artificial Analysis index ^[7]. That is a two-point gap, about 6 percent relative ^[15]. For it to transfer to your repository, the index's task mix would have to resemble agentic edits across a large Rust codebase, and the post does not break the score down by language or task type. The cited figure is also for xhigh, which by the author's own account is the configuration where the 27b burns the most time thinking ^[7][11].

"I thought finally the time has come where I no longer need a cloud AI subscription," he wrote ^[10]. He still runs a frontier model over the local model's output: Opus 5 reviews what the 27b writes ^[14]. On the first serious task the local model produced a plan and one-shotted the implementation, the review comments were mostly nits, and the 27b fixed them in the next commit ^[9].

Running locally also means owning the failure modes. The sessions looped a few times, crashed mid-session a few times, and needed hand-holding to resume, and he attributes the crashes probably to his power supply ^[8].

The material has no arithmetic on either side of the swap. There are no dollar figures for the subscription he dropped or the hardware he runs, and no count of tasks completed or abandoned ^[17]. The environmental claim is asserted, not measured: he writes that running on the laptop with no cloud dependency "is way greener and wastes less energy" ^[13]. He dates his Opus 4.6 conversion to around February or March 2026 and says about six months have passed ^[2], which places the writing in roughly August or September 2026 ^[19].

What makes the tradeoff work for him is the kind of work he does. As he put it, "Most of the coding I do is open source, so the model taking more time than a frontier cloud model wasn't an issue for me" ^[20].