According to OpenAI, Asana reported significant improvements to its browser agent during testing. By using GPT-6 Astra in Codex, Asana cut model costs by 76x and achieved a 5x speed increase in tests. The stated aim was to provide customers with more capable models.
Why it matters
Large reductions in cost and gains in speed for an agent workflow can affect how feasible it is to deploy such agents at scale, though the figures described here come from tests.
Who should care
Enterprises building browser-based or agentic workflows, and teams evaluating model and tooling choices for cost and performance, may find the reported results relevant.