Asana cuts browser agent costs 76x with GPT-6.1 Sol

Background and Context

Asana, the project management software company, has rebuilt its browser agent with GPT-6.1 Sol inside Codex. The details come from a customer case that OpenAI published on 9 October 2026. In Asana's tests, the agent ran 76 times cheaper and 5 times faster. Asana says the gain gives it room to offer customers more capable models without paying a high price on every call.

The public material needs careful reading. The figures come from a test setting. The summary does not describe the task set, the baseline model, or the way costs were counted. The two multiples are therefore a signal of scale. They are not a promise that other teams can copy on their own workloads.

A browser agent is expensive by design. It reads a page, either as structure or as a screenshot, and decides what to click or type. It then reads the new page state and repeats the loop until the task ends. A single task can take a dozen steps or several dozen. Each step sends the page content and the history back into the model, so input tokens pile up and delay builds step by step.

Deep Analysis

Cost and speed are two views of one problem. More steps and longer context raise the bill, and they also lengthen the wait for the user. A 76x drop in cost, which is nearly two orders of magnitude, rarely comes from swapping in a cheaper model alone. It more likely comes from several effects added together. On the model side, a stronger model may finish the same task in fewer steps. It may read page information more accurately and retry less often after an error. On the engineering side, the surrounding system can be rewritten. Teams can trim the page content that goes to the model. They can reuse cached context. They can move steps that plain deterministic code handles out of the model calls altogether.

The case also shows a loop that is easy to miss. Asana did this work in Codex, a coding agent. An AI tool was used to improve the running efficiency of an AI product. That suggests the coding agent took part in the refactoring, not only the model that sits inside the final product. The public information does not say how much of the 76x came from the model and how much came from the rewrite. This split matters. If most of the gain is engineering, other companies can chase it with any capable model. If most of it is the model, the gain depends on a specific supplier. Until more data appears, caution is the right stance.

Industry Impact

The case touches the sharpest tension in agent products. Enterprise software vendors want to ship agents as features. Per-call model costs can turn an attractive feature into a loss at scale. When the cost of one task falls to a small fraction of the old level, several things change at once. First, a feature can open to more users by default instead of sitting behind an expensive plan. Second, the vendor can spend the saved budget on a stronger model for key tasks and buy a higher success rate. This is what Asana means by offering customers more capable models. Third, a 5x gain in speed changes how the feature feels. An action that finishes in seconds can sit inside a daily workflow. An action that takes minutes can only run as a background job. For a work management platform, the value of a browser agent is to repeat chores across systems for the user. Examples include tidying tasks, syncing status, and moving information between web tools that have no open interface. These cases are sensitive to both delay and per-task cost. The effect on the wider market runs on three levels. For model suppliers, customer stories now show real spend and time per task, not only benchmark scores. Competition is shifting from the strongest model to the best usable model for the price. For peer software firms such as Notion, Atlassian and Salesforce, who are also pushing agent features, a proven cost drop makes high inference cost a weaker excuse for pricing. Agents will become standard sooner, and the gap will move to data, permissions and workflow depth.

For independent browser agent start-ups, the pressure is higher. A large platform can now run agents cheaply and already holds the user's business data. General browser automation tools must then show clearer strength in reliability, coverage or a vertical niche. The role of Codex in the story also matters. It suggests that model suppliers are bundling developer tools with the model, so one product set covers the path from building to tuning.

Outlook

Several signals deserve watching. The first is detail. If Asana or OpenAI shares task success rates, sample size, the comparison conditions against the old system, and the split between model and engineering gains, the case becomes far easier to judge and copy. The second is pricing. If the saving reaches customer-facing plans and the agent opens to more users, the saving has become product strategy. If it only lifts internal margin, the market effect is smaller. The third is reliability. Cheaper runs do not remove the old risks. Sites change their layout, login checks block flows, permission limits apply, and wrong actions can cause harm. Enterprise buyers will judge these features on stability and auditability.

The fourth is the response from other model makers. If rivals answer with their own cost and speed data, the market may move fast into a contest built around cost per task. The lasting value of the case is not the two multiples. It is the direction they point. An agent becomes a daily base feature, not a demo, only when its running cost falls low enough. Asana's test is one early data point on that path, and it still needs wider evidence.