GPT-6 Astra: New model with screen control and hardware design

What's new

OpenAI launched GPT-6 Astra on September 3, 2026. Enterprise customers in the Daybreak program got access first, followed by paid ChatGPT subscriptions and the API, which developers use to build the model into their own solutions.

Astra operates programs through the screen the way a human would, instead of requiring a special technical connection. According to DataNorth, OpenAI reports 72.6 percent on OSWorld 2.0 (a computer-use test), 57.9 percent on Terminal-Bench 4.0 and 74.1 percent on DeepSWE v1.1.

What's smart

The most striking story is hardware. OpenAI's hardware team says it used Astra to design a device that lets you talk to the Codex coding agent. It was done as a side project in under a week. According to OpenAI, Astra handled the enclosure, the circuit board (choosing components and drawing the traces), the firmware (the software inside the device) and the user interface.

A contract manufacturer made a small batch of prototypes in under a month. It took several rounds of revisions, because Astra did not always account for manufacturing requirements, tolerances and assembly details. The sources do not say how much humans did themselves, or whether the prototypes work fully. The figures are OpenAI's own.

OpenAI's hardware teams also use Codex and ChatGPT for CAD models (3D drawings), production drawings and simulations. According to OpenAI, token usage, meaning chunks of text, has grown 23-fold since February.

Cheaper or better?

In the API, Astra costs $10 per million input tokens and $50 per million output tokens in standard mode. In fast mode, the prices are $20 and $100. The monthly price for regular ChatGPT subscribers is not stated in the sources, and the same goes for Danish pricing and availability in Denmark and the EU.

According to DataNorth, Astra matches its competitor Claude Fable 5.1 on general reasoning and coding, and is stronger in computer automation and cybersecurity. According to OpenAI's figures, Astra spends about 47 percent less time on tasks than its predecessor GPT-5.6 Sol: roughly 40 versus roughly 75 minutes per task on OSWorld 2.0.

In CAD, Astra at its highest reasoning level scores 0.594 on the BenchCAD 2.0 test, which measures whether a model can recreate 3D parts and identify connections on circuit boards. A newer internal OpenAI model scores 0.875, but it has not been named, and no details have been shared.

One caveat: according to DataNorth, OpenAI acknowledges that Astra's written reasoning is harder to monitor than that of GPT-5.6 Sol, and calls it a serious concern for control and oversight.

What it's good at

Astra is aimed at work where the AI has to click around in programs on its own and solve tasks from start to finish, for example routine tasks on a computer or coding. Cybersecurity is also mentioned as a strong area.

For companies that develop physical products, the hardware example suggests that AI can speed up the early design phases. But the prototypes still required several rounds, so humans and manufacturers are far from redundant.