Asana cut browser agent costs by 76x and improved speed 5x by switching to GPT-4.1 Sol inside OpenAI Codex. The gains came from targeted model selection, not architectural overhaul.
The cost reduction matters because browser agents are notoriously expensive to run at scale. Every page interaction, every DOM parse, every decision loop burns tokens. A 76x cost drop changes the unit economics of deploying agents to real enterprise customers, not just demos.
Read the full piece for the specific benchmarks, the task types where Sol outperformed predecessor models, and how Asana structured its Codex integration. The methodology is the story, not just the headline number.
[READ ORIGINAL →]