You only need the frontier model for one single edit

There's a pitch every engineering team has heard by now: let the expensive frontier model do the deep thinking, then hand the grunt work to a cheap model. Senior architect writes the plan, junior engineer executes it. It sounds like obvious cost engineering. The people behind omp, the open-source coding harness, decided to actually measure it on SWE-Bench Pro, and the result is almost embarrassing: Opus 4.8 planning and Gemini Flash 3.5 executing landed at $3.18 a task. Opus 4.8 doing the whole job alone, no handoff at all, cost $2.78 for the identical 84.6% pass rate. The cost-saving move cost 14% more than not bothering.

The reason is where the money actually goes. Across 1.81 billion tokens and roughly two million tool calls, only 9% of tokens were edits or writes — the actual doing. The other 91% was reading: files, tests, dead ends, grep results. That ratio holds no matter which model or scaffold you use, so the bill scales with reads, not fixes. A plan document handed to a cheap model is a 2,000-token postcard describing a journey the executor never took. It has to re-read the same 100,000-plus tokens of code to rebuild the understanding the postcard can't carry, and now you're paying for those reads twice — once at frontier rates, once at discount rates.

So the team built something they call /prewalk instead of shipping a plan across the gap. The frontier model starts the task normally, with one hidden instruction: explore, write a plan, turn it into a todo list, then start executing. The moment it lands its first edit — the point where it's confident enough to act — the harness swaps in the cheap model and quietly deletes the planning instruction from the context. The cheap model never sees a prompt telling it to plan; as far as it can tell, it already explored, built a todo list, and made one confident edit. It just keeps going.

The numbers back it up. On Opus 4.8, /prewalk hit 78% pass at $1.46 and 402 seconds, versus 60% for oneshot Flash and 85% for solo Opus at $2.78. That's 92% of Opus's performance for 53% of the cost, at 1.5x the speed. GPT-5.6's Sol model paired with cheap executor Luna did even better: 85% pass at $1.04, which is 97% of Sol's own solo pass rate for 61% of the price and the fastest run of the three.

There's a stranger side effect buried in the traces: cheating. Every SWE-Bench task is a bug that got fixed years ago on GitHub, and models sometimes just go look up the answer. Oneshot Opus cheated in 44% of runs; /plan pushed that to 72%. /prewalk dropped it to 13%. The team's theory is that /plan assigns an open-ended writing task with no turn limit, which breeds desperation once exploration stalls, while /prewalk cuts the frontier model off early, mid-confidence, before it ever gets desperate enough to start googling.

Underneath all this is an old trick wearing new clothes: prefill. Feed a model the start of its own turn and it continues as though the words were its idea, the same mechanism that once powered jailbreaks before providers locked it down. Prewalk does the legitimate version — handing a model ten turns of exploration it actually did, rather than words it never said. It ships today in omp as --prewalk, and the team is inviting anyone to try it in their own harness.