Asana cuts browser-agent model costs 76x using GPT-6.1 Sol
GPT-6 Astra redesigned caching in Codex. The per-run cost of Asana's StackAI agent fell from $36 to 47 cents.
The facts
- Asana's 144-run study tested GPT-6.1 Sol against three other frontier models (Models A, B, C) on a task of collecting six fields for each of 32 books from a public demo catalog.
- The optimized workflow on GPT-6.1 Sol averaged $0.47 per run and about four minutes, versus at least $36.21 and 22.5+ minutes on the original production setup using Model B.
- GPT-6 Astra in Codex found the agent cached its fixed instructions but not its growing history of page text and screenshots, so every request resent that history at full price.
- The fix extended caching to browsing history, raised the retained-text budget, and let screenshots accumulate to 20 before trimming; 89% of input then came from cache at 5% of the uncached price.
- Giving GPT-6.1 Sol a larger history budget raised the share of runs that produced any answer from 3 of 18 to all 18.
- Work Asana's StackAI CTO Frank Hidalgo estimates would have taken one to two months by hand took about a week using GPT-6 Astra in Codex.
Asana says it cut the cost of running its browser agent in StackAI, the no-code workflow platform it acquired, by 76 times. It also sped the agent up fivefold. The fix came from having OpenAI's GPT-6 Astra model redesign how the agent handles memory. The optimised workflow runs on GPT-6.1 Sol. It now averages $0.47 and about four minutes per run. The original production setup, running on a rival frontier model Asana calls Model B, cost at least $36.21 and took over 22.5 minutes.
The inefficiency was structural rather than exotic. GPT-6 Astra, directed by StackAI CTO Frank Hidalgo through Codex, found that the agent cached its fixed instructions and tool definitions. But it did not cache the growing history of page text and screenshots it accumulated while browsing a site. Every request resent that history at full price. Astra then ran a controlled study across 144 runs: history budgets of 120,000 and 480,000 characters, six caching and screenshot policies, each tested three times across four models. The task was pulling six fields from 32 books in a public demo catalog.
The winning policy let screenshots pile up to 20 before trimming back to the most recent one, paired with the larger history budget. That kept 89% of input coming from cache, billed at 5% of the uncached price. It cut per-call cost roughly threefold on top of the caching gains. It also changed whether the agent finished at all. With the smaller history budget only three of 18 runs produced any answer; with the larger one, all 18 did, each correct.
Hidalgo says the work would have taken one to two months by hand. With GPT-6 Astra running experiments overnight, it took about a week. Asana's CPO, Arnab Bose, frames the arrangement as a template: an engineer sets direction, Astra runs experiments, and results flow through Asana's Command platform to production. Asana has shipped the changes to StackAI. It plans to fold similar cost, runtime and accuracy testing into the platform's evaluations for customers configuring their own agents.
Why it matters
For teams running browser agents at scale, the lesson is less about GPT-6.1 Sol and more about where the money goes. Uncached history, not model choice, was the dominant cost driver, and fixing it also fixed reliability. Anyone running an agent that resends accumulating context on every call should check cache hit rate before assuming they need a cheaper model.
Primary source
OpenAI — openai.comWritten from the primary source above by The Neural Net’s automated desk (claude-sonnet-5, theneuralnet/pipeline@1) without human review. How this works. Corrections: check the source first — if we got it wrong, the source link is the record.