The gain was not simply a model switch
Asana reported a study of a browser agent that carries out tasks through a web interface. Its team investigated traces, removed redundant context and improved reuse within a controlled evaluation.
For the optimized GPT-6.1 Sol workflow, average estimated model usage cost fell to $0.47 per run, a 76-fold reduction compared with the initial setup. Runs were also about five times faster. These are internal experimental findings, not industry-wide expectations.
Why browser agent histories get expensive
A browser agent observes pages, processes interface state, chooses actions and repeats until a task finishes. Sending a growing conversation history on every model call can inflate token use.
Stable instruction blocks can sometimes be cached, while duplicated or unnecessary messages can be omitted. The cumulative benefit can be substantial when similar requests recur many times during the same task.
Prompt engineering becomes systems engineering
The team paid attention to actual prompt content, ordering and cache behavior. Small structural changes can have large consequences when executed repeatedly within agent loops.
This is a useful reminder that price per million tokens is not the whole story. Cost per successfully completed task, combined with accuracy, retries and elapsed time, is a better comparison metric.
The limits of the result
Caching depends on provider rules, prompt structure and expiration windows. Removing history can also harm performance if a model loses information needed to make the next decision.
Teams should replicate tests on representative tasks and measure success rates before claiming savings. The Asana case shows a promising optimization strategy, not a guaranteed 76x discount for every browser agent.
Official sources and visuals
Original research images and charts are at the links above. The AUXER cover is independent editorial artwork rather than an official promotional image.