Asana cut browser-agent model costs 76x using GPT-6.1 Sol in Codex

Asana cut model costs 76-fold in browser tests by running its browser agent on GPT-6.1 Sol through Codex, the work-management company disclosed. The same tests showed the agent completed tasks 5x faster than before.

The two numbers — 76x cheaper, 5x faster — come from Asana's own internal testing, not an independent benchmark. They describe what happened when Asana swapped its browser agent onto GPT-6 Astra running in Codex.

What did Asana actually change?

Asana moved its browser agent — software that operates a web browser on a user's behalf — onto GPT-6 Astra in Codex. The stated purpose was twofold:

• Reduce the per-task cost of running the agent
• Increase the speed at which the agent completes tasks

The tests measured both dimensions. Cost per run dropped by a factor of 76. Task completion speed improved by a factor of 5.

Asana framed the cost and speed gains as a way to offer customers more capable models. If inference is 76x cheaper, Asana can either pass savings to customers or spend the same budget on more demanding agent work.

Why 76x matters for browser agents

Browser agents are among the most token-intensive AI deployments. Each task can involve dozens of page reads, element selections and action confirmations, all routed through a frontier model. Cost per task, not raw model quality, has been the binding constraint on how capable an agent a vendor can afford to ship.

A 76x cost reduction changes that arithmetic. At that scale, workloads that were commercially unviable — long-horizon browser automation, bulk agent runs across many users — move into the range a SaaS vendor can price into a subscription. The 5x speed gain compounds the effect, because faster task completion means fewer tokens burned per minute of wall-clock waiting.

The gains also explain why Asana tied the tests to offering customers "more capable models": cheaper inference lets a vendor upgrade the underlying model without raising prices.

What the numbers do and don't show

The disclosure carries clear limits. The figures are Asana's own test results, and the source does not specify the baseline model the 76x and 5x comparisons were measured against, the test suite used, or how many tasks the evaluation covered.

What the numbers do establish is direction and magnitude. A 76x cost gap between two model configurations is large enough that baseline details, while relevant for exact replication, are unlikely to reverse the conclusion that the new setup is dramatically cheaper for Asana's browser-agent workload.

What comes next

Asana says the improvement is headed toward its customers. The company's stated goal in running the tests was "to offer customers more capable models," which positions GPT-6.1 Sol in Codex as the production path for its browser agent rather than a lab experiment. If the 76x cost figure holds outside tests, Asana has headroom to ship more capable agents at existing price points — and rivals building browser agents on more expensive inference will face pressure to match the economics.