Asana Cuts Browser Agent Model Cost 76× With GPT-6.1 Sol

Asana says it made a StackAI browser agent 76× cheaper and 5× faster in a controlled internal study. The headline compares its original production workflow on an anonymized Model B with a revised workflow running GPT-6.1 Sol. The lower bill came from preserving prompt-cache prefixes, retaining more browsing history, pruning screenshots in batches, and changing models. OpenAI did not announce a new API price.

What changed in the 144-run test

Asana tested six cache/history policies at two history budgets on four models, with three runs per condition, then ran a 12-run follow-up. Every configuration gathered six fields for each of 32 books in a public demo catalog. GPT-6 Astra in Codex helped investigate, instrument, execute, and analyze the experiments; humans reviewed the findings before changes shipped to StackAI. The Asana study and OpenAI account describe the work.

Browser-agent setup	Model	Estimated model cost per run	Runtime	What the comparison isolates
Original production setup	Anonymous Model B	At least $36.21	At least 22.5 minutes	Baseline; some runs hit the step cap unfinished
Optimized workflow	Same Model B	$1.24	About 4× faster than baseline	Workflow changes alone: about 29× lower cost
Optimized workflow	GPT-6.1 Sol	$0.47	About four minutes	Workflow plus model switch: reported 76× lower cost, 5× faster

These are Asana’s estimated per-run model costs from provider token counters, not published token list prices or our own benchmark. Dollar figures are rounded means of three runs for the compared configurations. The rounded $36.21/$0.47 endpoints divide to about 77; Asana reports 76× using its underlying run data. The capped baseline makes its cost and time lower bounds; it did not always complete. Model B is unnamed, so identifying it as any specific competitor would be speculation.

Why the cache mattered more than a rate-card change

The agent had cached fixed instructions and tool definitions but repeatedly rewrote earlier browsing history: it dropped the previous screenshot and trimmed text almost every step. That broke the unchanged prompt prefix a cache needs. Asana cached the growing history, raised the history budget from 120,000 to 480,000 characters, and removed screenshots in batches only after accumulating 20. On Sol, its optimized setup read 89% of input from cache. With the larger budget, changing the cache and pruning policy reduced Sol’s estimated run cost from $1.97 to $0.47, about 4×.

Asana also reports that with the smaller history budget, only 3 of 18 Sol runs produced an answer; with the larger budget, 18 of 18 did, each correct against the prepared reference. The study’s narrow task, small per-condition samples, varying call counts, and internal execution limit generalization. The follow-up found that no pruning could be cheaper per call on this short task; long-running jobs still need context and spend limits.

Pricing impact: no API cut, different billing roles

The live pricing table above pulls current OpenAI API rates from our maintained dataset. GPT-6.1 Sol ran the production-task comparison; GPT-6 Astra in Codex helped design and run the experiment. Do not add Astra’s token rate to each Sol browser run or infer the experiment’s total cost from the $0.47 figure. The study does not disclose its complete Astra/Codex spend, tool fees, engineering review, infrastructure, or final customer invoice. More than 272,000 input tokens in an OpenAI request also trigger a higher long-context rate for the full request; service tiers and regional processing can change the bill. See OpenAI API pricing for those rules.

Verified GPT-6.1 Sol rate card

The official model documentation and OpenAI price table list Standard short-context prices per 1 million tokens at $2 fresh input, $0.10 cached input, $2.50 cache writes, and $10 output. Input above 272,000 tokens reprices the whole request to $4/$0.20/$5/$15. Batch and Flex halve Standard; Fast is 2×. OpenAI now lists Ultrafast at 6× Standard: $12/$0.60/$15/$60 for short context and $24/$1.20/$30/$90 for long context. Eligible regional processing adds 10%. The 76× Asana result changed neither the Standard gpt-6.1-sol row in our canonical pricing feed nor the API feed.

For a cross-provider control, compare a model you can actually run on the same browser task using the Anthropic pricing page and our token calculator. Asana’s anonymous Model B must not be treated as an Anthropic model. Our earlier AI agent economics guide explains why token spend, retries, tools, and human review belong in a complete cost-per-task calculation.

Labs boundary: model route included, browser claim unverified here

Our Labs includes the exact openai/gpt-6.1-sol route on a narrow 49-task deterministic text suite, but it does not reproduce Asana’s browser-agent workflow. The coverage note records the missing pinned browser build, anonymous Model B identity, prompts, screenshots and cache markers, raw runs, acceptance rubric, and complete token, tool, and infrastructure ledger. GPT-6 Astra was the experiment-design agent in Codex, not a scored browser-run model.

What this means for browser-agent teams

Teams with long, repetitive histories and expensive image/tool transcripts may benefit most: cache hits can reduce input spend without weakening the model. Teams with short tasks or unstable prompts may see smaller gains; careless pruning can cause re-navigation or wrong answers. Neither buyer should assume a 76× saving simply by switching its model ID.

1. Replay a fixed set of representative browser tasks with pinned model, tools, step cap, and acceptance criteria.
2. Log input, cached-input reads, cache writes, output, screenshots, retries, tools, elapsed time, and whether each run actually finished correctly.
3. Keep history append-only between deliberate batch-pruning points; test a budget large enough to retain task-relevant facts.
4. Compare total cost per accepted completed run, including failed runs and review time, before routing production traffic.

If you want an independent open-model cost control—not a substitute for Sol or a claim about Asana’s unnamed models—you can check Novita’s current hosted-model options. Verify browser-tool support, context, cache billing, and quality on your own task.

Affiliate disclosure: AI Pricing Guru may earn a commission from the sponsored link at no extra cost to you. It does not affect this analysis.

Bottom line: Asana’s result is strong evidence that cache-friendly agent design can change the economics of a browser workflow. It is not a 76× API price cut, a model-only win, or a transferable cost guarantee. Sources: Asana’s full methods, OpenAI’s customer story, GPT-6.1 Sol model documentation, and API pricing, checked October 9, 2026.