Future AI Pulse
Quick answer: A browser agent’s efficiency depends on how it retains page evidence, not only which model it uses. Asana’s case study describes changes to caching, text history and screenshot removal. Its headline comparison combines workflow optimization with a different model, so it should not be read as a universal model-price ratio.
What the study compared
OpenAI’s October 9, 2026 case study describes a 144-run study collecting six fields for each of 32 books from a public demo catalog. It reports an optimized GPT-6.1 Sol workflow averaging $0.47 in estimated model cost and about four minutes per run, compared with the original production setup on another model.
The account attributes the changes to larger retained text history, caching browsing history and removing screenshots in batches. On Sol alone, it reports a fourfold estimated cost reduction from $1.97 to $0.47. These are vendor-published study results, not measurements reproduced by Future AI Pulse.
Retained evidence can prevent repeated work
If an agent drops the page details it needs, it may have to revisit the page. If it retains everything without a plan, requests can become unnecessarily large. The practical question is which evidence supports the remaining task and which content is merely repeated interface noise.
For a catalog extraction, retain the fields already collected and the source location for each entry. Keep enough page context to resolve ambiguity. A list of values with no connection to the page is harder to verify than a smaller evidence record with clear references.
Change history policies in a controlled comparison
Use a public test site or an authorized staging environment. Fix the task and expected output before changing how the agent handles history. Run the baseline and candidate policies on the same inputs, and include repeated runs so one lucky result does not decide the outcome.
Track whether the agent finishes, whether all fields are correct, how often it revisits pages and how much time the task takes. Record model usage separately from other service costs. Include unfinished runs rather than comparing only successful examples.
Task: collect the specified fields from this authorized test catalog. Keep a source reference for every collected item. Do not infer missing values from similar products. Record completion, incorrect or missing fields and repeated page visits. Compare the baseline with one history-policy change at a time. Use actual usage records for cost estimates. Do not describe a failed or capped run as a successful extraction.
Cache reuse needs stable inputs
A cache can reuse eligible prior computation only under the service’s actual rules. Changing a request’s history may affect what can be reused. Inspect the real request and usage records rather than assuming that repeated-looking prompts receive the same treatment.
Keep a versioned record of the policy being tested: retained text, screenshot handling and when evidence is removed. When a page changes, ensure the agent can distinguish current observations from older material. Reusing stale information is not an efficiency improvement.
Correctness is the gate
For structured output, compare every expected item and field with a reference. Inspect the mistakes, not just an aggregate completion count. An agent that quickly fills every row with guesses has not completed an extraction task correctly.
Protect sensitive browsing data according to the environment’s retention requirements. A performance experiment should not expand access or retain private screenshots simply to make its benchmark look better. The workflow above is an editorial suggestion, not an implementation of Asana’s system.
Related: our cost-per-deliverable guide covers rework, and the job-state article examines repeated polling.
Illustrative workspace photograph: Yen Vu / Unsplash, under the Unsplash License. Not Asana product media.
Sources
Source checked October 14, 2026. Written with AI assistance. No hands-on benchmark or product test is claimed.