Cost per successful task is the right unit, but it assumes "successful" means the same thing across models, and none of the benchmarks you're citing really guarantee that. AutomationBench, FrontierCode and DeepSWE are all graded against some verifier, and two models can land the same aggregate score while failing on completely different slices of the task distribution — one nails the common case and blows the tail, the other does the opposite. That matters for the cost math because in production you don't just eat a failure, you escalate it to something more expensive, and a model that fails on a different slice than the one it replaced can trigger that escalation path just as often even with a better average score. So cost per successful task for a single model isn't really the number that matters, it's cost per successful task blended across whatever tier a failure falls back to, and that's a harder thing to measure than anything in either announcement.