Asana cuts AI cost 76× in OpenAI test case

The Asana test is presented as an official OpenAI case study, not an independent benchmark, and the numbers are explicitly described as test values with confidence rated medium. That framing matters because the mechanism behind the saving is procedural: the agent had been resending the record of pages it had already read at full price on every call, and the fix was to change how that record is kept and to batch-and-discard screenshots. Model choice was revisited as well.

The companion data point in the same roundup is LegalOn, which reported cutting Codex costs by 65% by routing different tasks to different models. Together the two examples point at a pattern where cost control comes from matching model and context handling to the task, rather than from a single stronger model.

A separate figure in the roundup gives a sense of scale on the research side: work that would take one to two months by hand was completed in about a week by letting the AI run experiments, across 144 tests with 4 models. The cost angle connects to the broader theme of the day — that how AI is used and governed, more than the AI itself, is where results diverge.