The Frontier Model Everyone Is Discussing Can Use a Mouse. That Is the Part That Matters to You
OpenAI released GPT-6 Astra on 3 September. It is available in ChatGPT on paid plans and through the API as gpt-6-astra, on Microsoft Azure and AWS Bedrock. Enterprise administrators have to switch it on; access is off by default at launch, and the cybersecurity capabilities sit behind a separate vetted-access programme.
Most of the coverage focused on intelligence claims. The specification that should interest a manufacturer is duller and more consequential: OpenAI positions the model as state of the art on computer use and browsing, and reports leading results on benchmarks built around operating software the way a person does – navigating interfaces, clicking through screens, completing multi-step workflows across professional tools.
If that holds up outside the benchmark, it addresses the single most persistent obstacle to automation in a mid-size plant.
The API wall
Every automation project in manufacturing eventually meets the same wall. The data you need lives in a system that will not talk to anything.
The MES installed in 2011 whose vendor no longer supports integration. The customer portal where orders arrive, which has no API and never will. The customs or logistics system your forwarder uses. The quality system that exports to PDF and nothing else. The supplier who sends order confirmations as email attachments in a format that changes without notice.
The usual answers are all bad. Pay the vendor a large sum for a connector. Have someone retype the data. Or leave the process manual and accept that the automation stops at that boundary. An agent that can operate a graphical interface directly sidesteps the integration question entirely, because it uses the software the way your staff already does.
That is a genuinely different proposition from the last decade of integration work, and it is worth taking seriously.
Why the scepticism is also earned
Robotic process automation promised exactly this in 2018, and the promise broke on contact with reality. Scripted GUI automation is brittle: a vendor moves a button, a dialogue box appears, a page loads slowly, and the script fails silently or does something wrong. Maintaining those scripts consumed the savings they generated.
A model that interprets the screen rather than following recorded coordinates should be more robust to that kind of change. Should be. Benchmark scores are run in controlled harnesses, and OpenAI notes itself that results depend on the harness and system configuration used – which is a fair disclosure and also a warning about how far those numbers transfer to your supplier portal at four in the afternoon.
Then there is cost. Reported API pricing sits at roughly $10 per million input tokens and $50 per million output, with reductions for cached input and batch processing. An agent that reads screens continuously generates a lot of tokens. A process someone does twice a day is probably worth it; one running constantly against a slow interface may not be, and that arithmetic is worth doing before the pilot rather than after the first invoice.
How to test it without betting anything
Pick the single most annoying manual data transfer in your operation. The one where someone reads from one screen and types into another, every day, and everyone agrees it is ridiculous. Time it. Then have the agent do it, supervised, for two weeks, with a human checking every result.
You will learn three things quickly: whether it handles your actual interfaces or only clean ones, what it costs per run at your volume, and what it does when something unexpected appears on screen – which is the behaviour that decides whether this is usable in production or a demo that works until it does not.
Two deployment details from the launch are worth noting for anyone in a regulated or IP-sensitive environment. Zero Data Retention is available for eligible API customers, which matters if the screens your agent reads contain customer drawings or pricing. And the fact that access is off by default for enterprise workspaces is a small gift: it means the decision to enable it is yours to make deliberately, rather than discovering later that half your engineering team has been using it on production data.
“Content generated using AI”