GPT-6.1 Sol Browser Agent: How Asana Cut Costs 76x

A GPT-6.1 Sol browser agent at Asana now completes a research task for a small fraction of what it used to cost, and the reason has little to do with switching models. In a customer story published by OpenAI on October 9, 2026, and a detailed engineering write-up from Asana, StackAI's Frank Hidalgo explains how the team made the agent 76x cheaper and 5x faster. Most of the gain came from one idea: stop breaking the prompt cache.

• Result: 76x lower cost and 5x faster on GPT-6.1 Sol, with 89% of input tokens read from cache, according to Asana.
• Experiment: 144 runs across six caching and history policies, plus a 12-run follow-up.
• Fix one: cache the growing conversation history, not only the system prompt and tools.
• Fix two: prune old screenshots in large batches so the cached history stays valid.

Why Was the Browser Agent So Expensive?

A browser agent works in a loop. It takes a screenshot, decides on an action, performs it and repeats, and every step sends the full history back to the model. As the history grows, so does the bill.

Prompt caching is meant to help. On the models Asana tested, tokens read from cache cost between 0.05x and 0.1x the standard input price. But a cache only matches when the start of the prompt is identical to the previous call. The original agent trimmed old screenshots from its history on every step. Each trim changed the earlier part of the prompt, so the cache missed and the whole history was billed at full price again.

"Caching alone is not enough," Hidalgo writes. How the history is managed decides whether the cache is ever used.

What Were the Two Fixes?

The first fix was to place a cache marker on the latest tool result, so that the entire history is cached and not just the fixed instructions at the top. In Asana's words: "Cache the growing history, not just the system prompt."

The second was to stop editing history on every call. The agent now keeps up to 20 screenshots and then prunes back to 1 in a single step. That means about 19 consecutive calls reuse the cached history before one call pays to rebuild it. The team also raised the history budget from 120,000 to 480,000 characters.

Hidalgo sums up the rule this way: "Keep the history append-only; when pruning is necessary, prune in large batches."

How Much Did It Save?

The test task asked the agent to collect six fields for each of 32 books from a public demo catalog, with answers scored against a reference set of 192 facts.

On the model the agent already used in production, the fixes alone made runs 29x cheaper and 4x faster. Moving the optimized agent to GPT-6.1 Sol brought the total to 76x cheaper and 5x faster than the original setup. An analysis by OrcaRouter of the published figures puts that at about $0.47 and four minutes per run, against a baseline of at least $36.21 and 22.5 minutes.

Two points of context are worth keeping. These are the companies' own measurements on one task, and the larger share of the improvement came from the workflow change, with the newer model adding the rest. I find that the most useful part of the result, because it means the technique is not tied to one vendor.

What Should Agent Builders Take From This?

The lesson generalizes to any long-running agent. Treat the prompt as an append-only log, put the cache boundary at the end of that log, and batch any cleanup. It is a design habit that costs nothing to adopt.

For readers choosing a model for this kind of work, our comparison of GPT-6.1 Sol and GPT-6 Astra covers pricing, and our guide to what GPT-6 Astra can automate covers computer use in more depth. More in our AI coverage.

Sources: OpenAI — Asana browser agent customer story — October 9, 2026; Asana — How we cut a browser agent's cost 76x — October 8, 2026; OrcaRouter — analysis of the figures — October 9, 2026.