Asana Slashes Browser Agent Costs 76x with GPT-6.1 Sol

Browser agent costs are the hidden tax on AI automation, and Asana just showed how steep that tax can be — and how quickly smart engineering can erase it. According to a case study published by OpenAI on October 11, 2026, Asana cut the model costs of its StackAI browser agent 76-fold while making it five times faster, dropping the estimated per-run cost from at least $36.21 to about $0.47. The twist: most of the savings came from fixing how the agent managed its own memory, not from switching to a newer model.

StackAI is the automation platform Asana acquired, and it lets customers build workflows that navigate websites, fill out forms, and gather information without writing code. At Asana's scale, small inefficiencies in those workflows compound fast. StackAI CTO Frank Hidalgo directed GPT-6 Astra, running inside OpenAI's Codex coding tool, to investigate the browser agent, test improvements, and compare results across 144 runs. Work that would have taken an engineer one to two months by hand took about a week, according to OpenAI's account of the study.

The study's job for the agent was deliberately concrete: collect six fields for each of 32 books in a public demo catalog, then deliver a table, a price total, and the cheapest and most expensive books. Asana's technical report, summarized by coverage on CryptoCompass, ran the same task across four models and a dozen configurations per model, with the main experiments taking place in September 2026.

The caching bug that was burning millions of tokens

GPT-6 Astra's investigation surfaced the central inefficiency quickly. The agent cached its fixed instructions and tool definitions, but it did not cache the growing history of page text and screenshots it gathered during a run. Every request therefore resent that entire history at full price. Making matters worse, the agent dropped older screenshots and trimmed text at nearly every step. Each edit altered the history, so simply caching the history would not have helped, and losing those facts risked forcing the agent to revisit pages it had already read.

The fixes Astra tested were architectural, not cosmetic. The team extended caching to the browsing history, raised the history budget to 480,000 characters, and changed the screenshot policy so images accumulated to 20 before being trimmed back to one, rather than being discarded at nearly every step. On GPT-6.1 Sol alone, that revised policy cut costs fourfold, from $1.97 to $0.47 per run, because roughly 89 percent of the input tokens were then served from cache at about five percent of the uncached price, reported by Zetik.

The larger history budget also made the agent dramatically more reliable. With the restricted history budget, GPT-6.1 Sol produced a correct answer in only three of 18 runs. With the expanded budget, all 18 runs returned the correct answer. Memory was not just a cost lever; it was a quality lever.

Most of the 76x came from engineering, not the model

Breaking down the headline number is important, because the 76-fold comparison blends two different effects. As reported by RohitAI's analysis of Asana's technical report, the original production setup on the anonymized rival Model B cost at least $36.21 per run, including spending on unfinished runs stopped at a step limit. Workflow optimization alone, keeping the same model, brought that down to $1.24 — a 29-fold reduction. Switching the optimized workflow to GPT-6.1 Sol then added another 2.6-fold reduction to reach $0.47. In other words, most of the gain was engineering discipline, with the model switch as a meaningful but secondary contributor.

That nuance matters for anyone budgeting AI agents. The figures are company-reported means of three runs per condition, and the cheapest settings were selected after testing twelve configurations per model, according to RohitAI. With three repetitions per main condition, the study is evidence about one workflow, not a general ranking of models. Nobody should read it as a claim that GPT-6.1 Sol is 76 times cheaper than every alternative in every setting.

Still, the direction of travel is clear. Asana has already shipped the browser navigation updates in StackAI, according to coverage by Brocker.org, and the company is building tools to make such experiments easier to repeat. The team plans to integrate this kind of testing into the platform's evaluation system, so both customers and internal teams can assess cost, runtime, and answer quality when configuring agents. The company also now deploys GPT-6 Astra within Codex for pre-release testing of product features, where the agent explores the platform, experiments with inputs, and flags bugs for human QA reviewers to examine.

Why this matters for the agent economy

Asana chief product officer Arnab Bose framed the experiment as a preview of human-and-agent teams working together, with an engineer setting the direction, GPT-6 Astra running the experiments, and the results flowing through the company's Command system into production. That pattern — agents doing the repetitive optimization loop under human direction — is exactly the workflow the study itself validated.

For the broader ecosystem, the lesson is that context architecture may matter more than raw model choice. One of the open questions raised by analysts covering the study is whether developers are overspending on inference when a redesign of memory management could unlock multiples of efficiency. Similar cost-per-action work is appearing across the industry, from Microsoft's Decision-1 model for scoring AI decisions cheaply to OpenAI's Decisions API, which trades verbose prose for fast typed verdicts. The pattern is consistent: as agents move from demos to production workloads, every token gets counted, and the winners will be the systems that manage browser agent costs most efficiently.

The full case study is published at OpenAI's site, with detailed breakdowns from CryptoCompass and Brocker.org.