If providers charge much less for text an AI has already seen, could that flip which coding-assistant setup is actually the cheapest?

This explores whether cheap cached input tokens could change which agent harness looks cheapest. That matters because the corpus suggests harnesses differ more in cost than in accuracy. Across three production coding harnesses and five models on SWE-bench Verified, swapping harnesses changed pass rates by no more than ordinary rerun noise but changed token cost by up to threefold Does switching between coding harnesses change model accuracy?. The gap had a specific source: fixed overhead from system prompts and tool definitions, re-sent on every step and multiplied by the number of steps. The corpus doesn't test cached pricing directly, so what follows is an inference. That repeated, identical opening chunk of each request is exactly what prompt caching discounts. So the main thing that makes a heavy harness expensive is also the part caching makes cheap, and the threefold gap measured at list prices would probably shrink, perhaps a lot.

A long-running case study shows how big this effect can be. Over 115 days of a persistent agent setup, 82.9% of all tokens were cache reads Do persistent agents really cost less per token?. When most of what you send is reused context, counting raw tokens describes your bill badly. The authors argue the useful measure becomes cost per finished piece of work, not cost per token. Applied to harnesses, a comparison that counts tokens without separating cached from fresh ones may rank them in the wrong order.

Caching doesn't cancel out every difference, though, because harnesses save money in different ways. One large automated search found four mechanisms that cut token traffic by 44.7–49.0% without losing performance: smarter action execution, compressing the context, trimming tool outputs, and handing reading off to helpers Can agent harnesses be automatically optimized across many environments?. Some of these remove fresh tokens, which caching can't discount, so they keep their value. Compressing the context is different. It rewrites earlier context, and caching only pays off when the start of each request stays exactly the same. This point is also an inference rather than something the corpus measures. A harness that compresses aggressively can therefore save on paper while losing its cache discount. A controlled study found that context management pays off most when the context window is tight, and that planning shifts from helping weaker models to cutting costs for stronger ones Which coding harness components matter most in different conditions?. Which harness is cheapest was already conditional before pricing entered the picture.

There's also a measurement trap. Rerunning the same harness on SWE-bench Verified flips individual task results about as often as switching harnesses does. On typical task-set sizes, even a 13-point accuracy gap is detected only about half the time Do harness comparisons on SWE-bench Verified measure real differences?. If cost is where the real differences show up, cost comparisons need the same care: repeated runs, matched budgets How should we measure gains from automatic harness evolution?, and separate counts for cached and fresh tokens.

The short answer is probably yes, but the corpus contains no direct experiment that reprices harnesses under cache discounts. The evidence that does exist suggests the cost gap comes mostly from repeated overhead, which is the part caching discounts. That means the cheapest harness at list prices may not stay cheapest once caching is in play.