A browser agent can make each request smaller and still make the finished job more expensive. Removing old observations may force it to revisit pages, lose facts or break reuse of earlier input. Asana's browser-agent study offers a useful way to test that tradeoff: measure what a complete, correct task costs, then inspect the request history that produced it.
The reported 76-fold model-cost reduction and roughly fivefold runtime improvement compare an optimized GPT-6.1 Sol workflow with the original Model B production setup. Both architecture and model changed. GPT-6 Astra in Codex helped investigate and run the engineering study; it was not the runtime model behind the cheapest browser configuration.
Define completion before calculating savings
The main study included 144 runs across four models, six history/screenshot policies, two history allowances and three repetitions per condition. Every run collected six fields for 32 books, or 192 facts, from one public demonstration catalog. A reference answer made correctness inspectable. Three competing models remained anonymous, which limits provider-level comparisons.
Record unanswered runs as failures, not cheap successes. With a 120,000-character history allowance, GPT-6.1 Sol answered in only three of 18 runs; at 480,000 characters it answered correctly in all 18, according to Asana. That does not establish reliability on authenticated applications, changing layouts or ambiguous business tasks. It establishes a large task-specific sensitivity to retained evidence.
Preserve useful history and the stable prefix
Prompt caching reuses eligible computation for an identical request prefix; it does not replay a finished answer. Asana's original agent cached fixed instructions but repeatedly removed old screenshots and trimmed text. Those edits changed earlier history, undermining reuse of the growing record. A smaller prompt could therefore create more repeated processing.
The selected batch-pruning policy kept up to 20 screenshots before reducing them to the newest one. Combined with the larger text allowance, it produced longer append-only stretches and an 89 percent cached-input share for GPT-6.1 Sol. More memory and less frequent deletion worked together. Caching without batch pruning was more expensive than no caching on three of four models at the larger allowance.
Compare matched configurations, not just the headline
Within GPT-6.1 Sol, the new policy reduced cost fourfold; that is a different denominator from the 76-fold cross-model, old-versus-new-system claim. Model B also improved under the revised architecture. Some original runs reached step limits, so their measured durations were lower bounds for completing the job. Neither result is a neutral leaderboard.
Reproduce the comparison with fixed tasks, reference outputs and provider usage counters. Vary model, screenshot retention and text allowance separately, then repeat promising settings. The published three repetitions support broad patterns, not precise small differences. Keep request traces, configuration versions and failed outputs outside the active prompt so reviewers can audit evidence after pruning.
Set boundaries before shipping the optimization
Keeping every screenshot sometimes lowered per-call cost in follow-up tests, but long workflows can exceed context limits, and cache economics vary by provider. Compaction can bound history while losing details and changing the prefix. Set step, token and total-spend limits, and include retries, human repair and cache expiration in the cost of a finished task.
Asana says the browser changes reached StackAI production, but broad customer outcomes and long-term error rates were not published. Pilot realistic workflows, measure completion, correction time, duration and cache share, and protect screenshots, credentials and customer records. Caching changes computation and billing, not data-governance obligations. Expand only when the complete workflow becomes cheaper without sacrificing correctness or control.