What changed
OpenAI published an Asana browser-agent case study on October 9. It reports a 144-run study in which a workflow using GPT-6.1 Sol averaged $0.47 in estimated model costs, 76 times below the original setup on another model. GPT-6 Astra in Codex helped investigate and optimize the agent. Those are different roles: the model improving the software and the model operating the resulting workflow.
Source: Asana cuts model costs 76x in browser tests with GPT-6.1 Sol · 2026-10-09 · Primary source context · attributed account
Why it matters
Silver Owl’s reading is that the useful story sits in the machinery around the model. A browser agent must carry enough of its earlier work forward to avoid becoming an expensive forgetful assistant. Changing how it remembers pages can change the economics of the whole task. For a small team, that makes the purchasing question more demanding than choosing the lowest token price. An agent that repeatedly loses its place may consume both money and the operator’s afternoon. The case offers a reason to inspect the workflow before buying a bigger allowance; it does not establish what another company will save.
Source: Asana cuts model costs 76x in browser tests with GPT-6.1 Sol · 2026-10-09 · Primary source context · attributed account
What to check
The study used a public demo catalog, not an open-ended sample of customer work. Before treating its headline as a forecast, compare the same task, success criteria, history settings and billing assumptions across candidate setups. Count failed attempts and human repair time as well as successful runs. Ask which changes survive a different website or a longer session. We have checked OpenAI’s page and its date, but have not rerun the experiment or verified Asana’s production savings. The exact study and deployment dates remain unstated.
Source: Asana cuts model costs 76x in browser tests with GPT-6.1 Sol · 2026-10-09 · Primary source context · attributed account