Asana cuts model costs 76x in browser tests with GPT-6.1 Sol

Asana has cut the cost of running its browser agent by a factor of 76 and made it five times faster in tests, according to an OpenAI News item. The company used a model identified as GPT-6 Astra inside Codex to achieve those results, which the report frames as a way to offer customers more capable models.

The item repeats the same two performance figures without publishing benchmark methodology, the size or shape of the test suite, token prices, latency baselines, or the date the work was done. Because of that, the numbers should be read as test results reported by the publisher rather than a guarantee of production performance.

For context, browser agents typically operate by looping: they read a page, decide on an action, and execute it through a browser automation layer, which historically means many model calls per task and therefore high inference cost. An improvement of this magnitude would matter most for high-volume, multi-step automation, where per-task savings compound quickly. It is also worth noting that cost and speed gains can come from several places at once, including a cheaper model, fewer reasoning steps, caching, or a change in how the agent is orchestrated, and the report does not separate those factors.

Why it matters: teams building browser automation should expect continued pressure on per-task inference cost, and may want to re-test their own agent loops against newer models or harnesses rather than assuming current price-performance holds. That is inference from the reported figures, not a stated claim about any specific product or customer outcome.