Ari Weinstein

James Sun says the moat startups still have against OpenAI is speed and pain. He and Ari Weinstein were startup founders before they joined OpenAI, and he is still bullish on startups.

Two openings remain. There is still a huge amount to build. OpenAI has thousands of people and still feels stretched thin, because the teams are not very big for all the problems. A founder who deeply understands a problem, knows the industry, understands how regulation works, knows the go-to-market motion, and knows how to service customers is still looking at a green field. Tools through the APIs, including computer use, let that founder accelerate.

Many vertical industries are also adopting AI much more s

Ari Weinstein says companies should automate tedious workflows first, and the boring work is the opportunity. ERPs, CRMs, and every other system of record are the place to start.

Procurement is a clear example. A company orders lab equipment each year, and teams fill a giant spreadsheet with a thousand rows. An operations team moves the spreadsheet into internal procurement software, the order lines get finalized, and a procurement team then reaches out to vendors. Computers can automate the whole sequence.

Finance is the next category. A company keeps many ledgers, including ledgers for expenses, and the end of the month brings a lot of reconciliation. A lot of the financial software has

Ari Weinstein says Astra moved computer use forward at the edges even though computer-use models were already strong before Astra. It is a little faster on average because it uses fewer thinking tokens to do the same task. It is better at creative work, and it is more reliable around the edges.

In the launch video, someone talks to the model while it draws in Microsoft Paint, and Astra is way better at that sort of work. Weinstein thinks the improvement probably transfers to other creative tasks. He trusts it a little more. A small difference that is hard to perceive still matters between something being 99 percent reliable and 97 percent reliable, and Astra sits a little higher in a way th

Ari Weinstein says computer use is not yet as fast as an expert user. Over the past couple of months it has become faster than the average person at finishing tasks, in most cases. The next frontier he describes is literally superhuman speed: as fast as expert computer users, or faster. He expects products to feel more real-time once that happens, and he expects people to hand over chores they still do by hand because the effort to start will be lower.

What still slows the agents down, in his account, are small cuts on the model, on inference, on the harness, and on how the screen is represented.

Shawn Wang says a leading AI podcast, friends of the show, is famous for claiming computer use has not advanced in the last two years. Ari Weinstein says the claim is a few months old, and he hopes the view has changed, because computer use is 180 degrees different than it was.

The change he describes is the last year of models. At Sky, after Shortcuts at Apple, the models could start a task and then get stuck. They now debug, try again, and check what is working. In Codex, computer use often writes JavaScript and runs many actions at once instead of one action at a time. The model may use a screenshot, accessibility, or Playwright, depending on the task.

Ari Weinstein ordered from a meal-prep service he had joined to eat healthy. It lets him specify grams of chicken and grams of rice, and he can customize each order to a high degree of granularity. The form was so complicated that one order took him two hours. Computer use finished that order in 15 minutes, eight times faster than he could, and saved him the two hours, on GPT-6.1 Sol. He calls that kind of task the one that feels really powerful.

Shawn Wang puts flight booking, shopping, and even a game in the same group of chores. His own first use is YouTube. A/B tests and community posts are not available through an API, so the agent has to run them in a virtual machine. He says none of

Ari Weinstein says GPT-6.1 Sol is a fifth of the cost of Astra, and a seventh of the cost when the comparison is computer use. He ties that price, and the model's speed, to why the new model is particularly good for computer use. The same keynote recap includes Dots, a personal assistant with computer-use features, and an Agents API so developers can build on the computer use already in Codex and ChatGPT.

The Decisions API is a different, smaller model. It runs inference in parallel and it does not reason, which makes it fast and also less good at long-horizon tasks. Bringing that speed together with the computer-use models is still an open research question.