← Back to feed

Asana ran a 144-trial controlled study across four frontier models to find the optimal configuration for its browser agent workload. GPT-6.1 Sol with prompt cache tuning came out cheapest by a wide margin, at $0.47 per run — 76 times cheaper than the original production setup.

OpenAI published the findings on October 9, 2026.

The Numbers

The original production configuration ran on a model Asana calls “Model B,” priced equivalently to GPT-6.1 Sol. Before optimization, some runs hit the step limit before completing. After optimization on Model B, cost dropped to $1.24 per run. The optimized workflow on GPT-6.1 Sol reduced that further by 2.6x, reaching $0.47.

On latency: the original setup on Model B took at least 22.5 minutes per run. The optimized workflow on GPT-6.1 Sol averaged four minutes — a 5x improvement.

Every run in the optimized workflow completed the task and returned the correct answer.

What Changed

The key optimization was cache utilization: 89% of GPT-6.1 Sol’s input tokens were served from cache at 5% of the uncached price. Across the run, that single change reduced per-call cost 3x. Asana also introduced a larger history budget and an updated screenshot policy, together contributing a further 4x reduction from $1.97 to $0.47 on GPT-6.1 Sol alone.

The process used GPT-6 Astra running inside OpenAI Codex to design and run the optimization experiments. Astra was the research engine; GPT-6.1 Sol became the production target.

The Four-Model Field

Asana’s 144-run study compared:

• GPT-6.1 Sol — final optimized cost: $0.47/run (~4 minutes)
• Model A — smaller model from a competing lab, half the price of GPT-6.1 Sol
• Model B — original production model, same price as GPT-6.1 Sol; optimized to $1.24/run
• Model C — updated version of Model B, same price

GPT-6.1 Sol outperformed the cheaper Model A on quality at the tested workload, while significantly undercutting Model B and C on cost after the cache optimization was applied.

Context: GPT-6.1 Sol Pricing

GPT-6.1 Sol launched September 29, 2026 at $2 per million input tokens and $10 per million output tokens — one-fifth the price of GPT-6 Astra, and matching GPT-6 Sol’s published pricing. The cache discount at 5% of input price makes inference-heavy agentic loops substantially cheaper than headline rates suggest.

What It Means

The Asana study is a template for the cost engineering pattern now standard in production AI: run the frontier model through a structured optimizer (Astra, in this case), find the cache-maximizing prompt structure, then deploy on the cheaper sibling. The result is not quality degradation — every run completed correctly. The result is a 76x reduction in inference spend.