Seventy-six times cheaper. Five times faster. Every run correct. That is the headline number from an OpenAI case study published on October 9, and the obvious reading is that a newer model did the work.
The obvious reading is wrong. And the gap between the headline and the detail is the most useful lesson about AI costs I have seen this year.
What Asana actually changed
The case study on OpenAI's site describes a browser agent inside StackAI, the no-code platform Asana acquired. Customers use it to visit websites, fill in forms and collect information. At Asana's scale, small inefficiencies in those runs multiply into real money.
StackAI's CTO, Frank Hidalgo, pointed GPT-6 Astra in Codex at the agent's code and asked it to find out why runs were expensive. The answer was not exotic. The agent cached its fixed instructions and tool definitions, but not the growing history of page text and screenshots it collected. Every new request resent that entire history at full price.
Worse, the agent trimmed old screenshots and text at almost every step. Each trim changed the history, so even switching caching on would not have helped. And when useful facts were trimmed away, the agent sometimes had to revisit pages it had already read.
Where the 76x really came from
Asana ran a 144-run study across four frontier models, two memory budgets and six caching and screenshot policies. The winning policy let screenshots pile up to 20 before cutting back to the most recent one, which kept the history stable for longer stretches. Combined with a larger memory budget, that became the optimized workflow.
Now the important part. On Model B, the competitor model Asana was already using in production, the fixes alone cut cost from at least $36.21 per run to $1.24. That is a 29x reduction on the same model, at the same price per token. Switching to GPT-6.1 Sol took it from $1.24 to $0.47, a further 2.6x.
Multiply those and you get roughly 76x. But the split is clear: the architecture fix delivered the order of magnitude, and the model switch delivered the last step.
On GPT-6.1 Sol, 89 percent of input came from cache at 5 percent of the uncached price. That is not intelligence. That is plumbing.
Why this changes how you read every AI cost report
When a team tells me its AI costs are too high, the conversation almost always turns to vendors. Which model is cheaper, which provider will negotiate, whether open-weight models would save money. I covered how AI costs are getting metered a few weeks ago, and the reflex response to a bigger bill is always to shop around.
The Asana numbers say look inward first. Most agent workflows were built quickly, by people proving a concept, not optimizing a production system.
They resend context they do not need to resend. They throw away memory they will need again. They retry steps because they forgot what they saw.
Each of those habits is invisible in a demo and very visible on an invoice.
There is a quality angle too. With the smaller memory budget, GPT-6.1 Sol produced an answer in only 3 of 18 runs. With the larger budget, it answered all 18, correctly.
The cheapest configuration was also the most reliable one, because it stopped forgetting. Cost and quality were not a trade-off here. They had the same root cause.
The quieter lesson: who ran the experiment
Hidalgo estimates the investigation would have taken him one to two months by hand. With GPT-6 Astra running experiments, it took about a week. His description of the method is worth repeating: he set a goal before going to bed and reviewed the results in the morning.
The agent mapped the code, proposed fixes, refactored the system so many configurations could run in parallel, and ran the full study. A human chose which fixes to test and approved what went to production. Every request and result was logged in Command, Asana's delivery platform, so the whole study could be reviewed afterward.
That is the shape of useful AI work in most companies right now. Not autonomous agents replacing a department, but one expert directing a tireless experimenter, with a complete paper trail. Hidalgo put it well: shipping speed is no longer the bottleneck, human attention is.
What I would do with this on Monday
Pull the logs for your three most expensive AI workflows. For each one, find out how much of every request is repeated context, and whether that context is being cached. If nobody on your team can answer that within a day, that is your first finding.
Then set a baseline before you change anything: cost per completed task, time per task, and the share of runs that actually finish correctly. Most teams only track the monthly bill, which tells you nothing about where the waste sits.
Only after that, compare models. A cheaper model on a wasteful architecture is still a wasteful architecture. If you want help auditing and rebuilding an agent workflow like this end-to-end, difrnt.ai (difrnt.ai) builds and optimizes custom AI systems for exactly this kind of problem.
The model got the headline. The cache did the work.