Asana Slashes StackAI Browser Agent Costs by 76x Using GPT-6.1 Sol and Prompt Caching
Asana has reduced the operational costs of its StackAI browser agent by 76 times and accelerated execution speed by fivefold through architectural refactoring and the deployment of OpenAI’s GPT-6.1 Sol. The optimization, completed in a one-week internal study, brings the estimated model cost down to $0.47 per run, providing a practical framework for enterprises managing the high token expenditures of autonomous AI agents.
How StackAI Reduced Browser Agent Costs to $0.47 per Run
To address the compounding costs of its browser automation workflows, Frank Hidalgo, StackAI CTO, directed GPT-6 Astra in Codex to investigate the agent’s architecture and test improvements. The core inefficiency lay in how the agent handled context. While the system cached its fixed instructions and tool definitions, it failed to cache the growing history of page text and screenshots it gathered during execution. Consequently, every request resent that history at full price.
Furthermore, the agent dropped older screenshots and trimmed text at nearly every step. Because prompt caching only reuses the longest unchanged prefix of a request, each edit altered an earlier part of the prompt, breaking cache reuse on almost every call. Losing those facts also forced the agent to revisit pages it had already read, compounding the token waste.
Hidalgo reviewed the proposed fixes and implemented three critical changes: extending caching to the agent’s browsing history, increasing the amount of text it could retain, and removing screenshots in batches rather than at every step. The optimized policy allowed screenshots to accumulate to 20 before cutting back to the most recent one, keeping earlier history unchanged for longer stretches between removals.
The team validated these changes through a 144-run study that tested history budgets of 120,000 and 480,000 characters across six caching and screenshot policies. "This would have taken me one to two months by hand. With GPT-6 Astra in Codex, it took about a week: I’d set a /goal before going to bed and review the results in the morning," Hidalgo noted in the StackAI engineering blog.
Why GPT-6.1 Sol Replaced the Original Model for Asana’s Production Workflows
The 144-run study tested GPT-6.1 Sol against three other frontier models, including the original production model (Model B). For Model B, the optimization reduced the estimated model cost from at least $36.21 per run to $1.24 per run, a 29x reduction. However, migrating the optimized workflow to GPT-6.1 Sol yielded an additional 2.6x cost reduction, bringing the final price to $0.47 per run. Every run in the optimized GPT-6.1 Sol workflow completed the task and returned the correct answer in about four minutes.
The financial advantage of GPT-6.1 Sol is driven heavily by its aggressive caching discounts, which are critical for agentic workloads where system prompts and context are reread constantly.
| Model | Input per 1M Tokens | Cached Input per 1M Tokens | Output per 1M Tokens |
|---|---|---|---|
| GPT-6.1 Sol | $2.00 | $0.10 | $10.00 |
| GPT-6 Astra | $10.00 | $1.00 | $50.00 |
GPT-6.1 Sol’s cached input is priced at a 95% discount off the standard input rate, a slight improvement over the 90% discount offered by its predecessor, GPT-6 Sol, according to Artificial Analysis. Despite the lower price, GPT-6.1 Sol achieves near-Astra intelligence, gaining 4 points in the Intelligence Index over GPT-6 Sol and making significant gains in agentic knowledge work. At max effort, it costs less than a quarter of GPT-6 Astra per Intelligence Index task.
"This is what teams of humans and agents look like in practice. An engineer set the direction, GPT-6 Astra ran the experiments, and the results went through Command to production," said Arnab Bose, Chief Product Officer at Asana.
Why Asana is Prioritizing Agent Economics Following the $75M StackAI Acquisition
The urgency behind this optimization is directly tied to Asana’s strategic pivot following its acquisition of StackAI for $75 million in May 2026. StackAI is a no-code platform that enables companies to design and deploy custom AI agents that navigate websites, fill out forms, and pull data across enterprise systems like ERP and CRM platforms.
Asana’s stated goal is to transform its platform into an "operating system for human-agent teams." However, at Asana’s enterprise scale, the compounding costs of running autonomous agents on expensive frontier models without strict context management are financially unsustainable. The transition from simple, deterministic task automation to complex, multi-step agentic workflows requires rigorous cost controls. Asana’s internal study reflects a broader industry realization that the viability of AI agents depends just as much on context-window management and caching strategies as it does on raw model intelligence.
What the Asana Optimization Means for Enterprise AI Developers
For IT professionals and developers building autonomous agents, the Asana case study highlights a necessary shift from pure prompt engineering to aggressive context-window management. The 20:1 screenshot pruning ratio is a specific, actionable technique for computer-use agents that preserves cache integrity while preventing context overflow.
Furthermore, the results validate a tiered model routing strategy. Developers should deploy a highly capable, cost-efficient model like GPT-6.1 Sol for the bulk of agentic work, reserving flagship models like GPT-6 Astra only for the hardest multi-step technical tasks where the cheaper model measurably fails, as recommended by Exodata. Because agent workloads run mostly on cached tokens, leveraging models with deep cache discounts collapses real-world costs far below the headline per-token rate.
The optimized workflow on GPT-6.1 Sol completed every run and returned the correct answer, proving that aggressive context pruning and strategic model selection can make complex browser automation economically viable for enterprise deployment at under 50 cents per task.
