What just changed
On October 9, 2026, OpenAI published the Asana case. The title cites a 76x cut in model costs in browser tests with GPT-6.1 Sol. The summary says that with GPT-6 Astra in Codex, Asana's browser agent became 76x cheaper and 5x faster, again "in tests".
Two numbers, one caveat. And a nudge to revisit how you pick models.
Where the gain is clearest
A browser agent loops: read a page, decide, click, repeat. Every step spends model calls, so cost compounds. Per OpenAI, Asana cut that cost 76x and time 5x in its tests. A ratio like that changes what an agent can afford to do on every request.
Another detail from the summary: the goal is to offer customers more capable models. The savings fund capability rather than just trimming a bill.
What the source does not say
The relayed text publishes no success rate, no test protocol, no production figure and no breakdown of which model does which work. The title names Sol while the summary names Astra in Codex, so the extract cannot attribute each gain to one model. No quality comparison between them is provided.
Cautious reading: 76x and 5x describe tests, not a real invoice.
Pricing and operational implications
- Measure cost per successful task, not price per million tokens.
- Time runs end to end: a 5x faster agent changes the user experience.
- Keep a reproducible test set, because that is the exact frame in which Asana reports its numbers.
For a multi-model architecture
Here, the possible lesson is that one agent can be redesigned around what actually costs money, the repeated action loop. With no published detail, treat it as something to test, not a recipe.
Three levers to activate this week
- Record 20 real browser flows and replay each with two model configurations.
- Log cost, duration and success rate per flow, not just the average.
- Set a quality threshold before looking at the bill, so you do not optimize blindly.
Is your agent still too expensive on every click?
If you follow the latest AI technology, I publish a deep dive every day on frontier models, hardware, robotics, automations and AI-generated music. 👉 Get the next one straight in your inbox — sign-up takes ten seconds.