Most coding agents have an effort setting, usually running from low to max, that decides how long the model thinks before it starts editing files. It's tempting to leave it on max and forget about it. The catch is that thinking is billed as output tokens, so every step up costs money.
Cognition's FrontierCode 2.0, released on 27 October, puts numbers on that trade. Open-source maintainers wrote its 132 tasks, and each pull request is graded on whether the maintainer would merge it. For every effort level a model offers, the benchmark reports the score and the average API cost of a run.
The last step up buys tokens, not many points
Sonnet 5.5 is the clearest example. At xhigh it used 80k tokens per task; at max it used 515k. The cost went from $3.31 to $19.94, and the score went up 6.21 points.
Most models aren't that extreme, but the pattern holds. The top level costs between 1.1x and 6x what the one below it costs, and for 11 of the 17 models it adds two points or less. Opus 5.5 is typical: max costs 2.5x as much as xhigh and scores 2.07 points higher. FrontierCode doesn't publish error bars, so it's hard to say whether a gap that size is real.
On three models, more effort actually scored worse: Fable 5.1 did worse at max than at xhigh, GPT-6.1 Sol did worse at xhigh than at high and Kimi K3 did worse at max than at high.
So the top setting isn't automatically the best one. Whether it's worth it depends mostly on what the model costs to begin with.
On cheap models, max is worth it
Haiku 5.5 is the counterexample. Going from xhigh to max adds 5.44 points and 14 cents per task, for a total of $0.37. GPT-6.1 Sol gains 6.62 points at max and still comes in at $1.03.
In relative terms those steps are much like Opus's. What they're applied to is very different. Moving Opus from xhigh to max adds $7.64 to a $5.18 task, while on a model that costs under a dollar a task the same move costs cents. The points are worth the same either way; the bill isn't.
Our recommendations follow from that. For each model we pick the cheapest level that scores within 90% of its best. Eight models end up at their top level, and five of them cost under $1.10 a task: Haiku 5.5, GPT-6.1 Sol, GPT-6 Luna, GLM-5.3 Flash and DeepSeek V4.1 Flash. The other seven do better further down, among them Opus 5.5 at xhigh, GPT-6 Astra at xhigh and Grok 4.7 at medium.
Sometimes the better setting is a different model
We also skip any level that a cheaper model from the same vendor matches within a point. Same vendor means the same agent and the same subscription, so swapping models is just a settings change.
Sonnet 5.5 fails that check at every level. Haiku 5.5 matches its low, medium and high settings for less. Opus 5.5 at high scores 34.53 for $2.42, within a point of Sonnet at xhigh, and Opus at xhigh beats Sonnet at max outright. Opus's per-token price is twice Sonnet's, but it needs far fewer tokens: 88k at xhigh, against 515k for Sonnet at max.
Fable 5.1 goes the same way. Its best result, 32.44 at xhigh for $8.68, is beaten by Opus 5.5 at high for 28% of the cost. OpenAI's lineup does it too: GPT-6.1 Sol matches GPT-6 Astra at low, medium and high, so Astra is only worth running at xhigh or max.
When a vendor sells several models, choosing an effort level is really choosing a model.
Switching vendors can save more still, though it's a bigger change. Grok 4.7 at medium scores 25.94 for $8.03; GPT-6.1 Sol at low scores 25.6 for 34 cents. But Grok was tested in Grok Build and Sol in Codex, each vendor's own agent, so moving from one to the other means changing your tooling and your subscription as well. That's why we list those swaps separately.
What could change this
The biggest unknown is noise. Without error bars, some of the one-point calls above could turn out to be ties. FrontierCode is also just one benchmark, 132 tasks judged on whether the PR would be merged, and another suite, or Artificial Analysis's broader index, could rank the levels differently.
Flat-rate plans are the other caveat, since there you spend usage limits rather than dollars. Usage generally scales with tokens, so the effort advice should mostly carry over: Sonnet at max uses 6x the tokens it uses at xhigh. The model swaps are harder to call, because vendors don't publish how a plan counts Opus tokens against Sonnet tokens.
If you pay API prices, here's what we'd do today. Run Opus 5.5 at xhigh rather than max. Pick Opus over Sonnet or Fable. Leave Haiku, Luna and GPT-6.1 Sol on max, where the top setting costs cents.