GPT-6.1 Sol makes computer use agents affordable

Computer use agents, AI models that click, type, and read a screen the way a person does, just dropped in price by about seven times. OpenAI's GPT-6.1 Sol, released at the company's DevDay event on September 29, scores within 2 points of the flagship Astra model on the OSWorld 2.0 benchmark, and does it for roughly $1.30 per completed task instead of $9.30.

That price drop matters more than the model name does. A computer use agent does not need an API. It opens a browser, finds the login box, clicks the right buttons, and reads what comes back, the same way an employee would. For a small business running on a mix of a CRM, a supplier portal, a licensing site, and a scheduling tool with no integration between any of them, that has always been the appeal. Cost has been the catch.

Why computer use agents got cheap enough to try

OSWorld 2.0 is a standard test for AI that operates real software through a screen rather than through code. GPT-6.1 Sol scores 71.4 percent at its highest reasoning setting. Astra, OpenAI's top model, scores 73.5 percent. The quality gap is small. The cost gap is not: Sol finishes a typical task for about $1.30 against Astra's $9.30, by OpenAI's own published numbers. On a separate automation test, AutomationBench, Sol also edged past Anthropic's Opus 5.5 while costing roughly a third as much. The model is live now in ChatGPT Work, in Codex, and through the API as gpt-6.1-sol.

Where this actually helps a small business

• Supplier or distributor portals that only offer a web login, no API, for checking stock or placing reorders
• State or local licensing and permit portals that require manual form entry
• Directory and review listings that need the same update pushed to each site by hand
• Legacy scheduling or practice management software that predates any integration option

How to pilot one without the risk

1. 1

   Pick one task, not five

   Choose a single repeatable task inside one portal, ideally something low stakes like checking order status, not something that moves money or edits a customer record.

2. 2

   Write down the exact steps

   List every click and field a person currently follows. If you cannot write the steps in order, the task is not ready for an agent yet.

3. 3

   Run it with a human checking every output

   For the first couple of weeks, have someone review what the agent did before it counts as finished. Treat it like onboarding a new hire, not flipping a switch.

4. 4

   Track the real cost per task

   Compare the token cost against the minutes a person currently spends on the same task. At roughly a dollar a task, the math only works if it happens often enough to matter.

OpenAI says a faster version of Sol is coming to Codex within weeks, and pricing on these models tends to keep falling, so get the task working without over-building around today's exact setup. If part of your business still runs through a browser tab because nobody ever built the integration, this might be the year the math finally works in your favor. Happy to take a look at where that applies in your own workflow.