Updates·September 4, 2026, 09:32
GPT-6 Astra: OpenAI model that operates software on its own
AI-generated from the sources listed, with automated source checks.
OpenAI has launched GPT-6 Astra, a model that, according to the company's own figures, is significantly better at operating computers and at cybersecurity than its predecessor, GPT-5.6 Sol.

What's new
On September 3, 2026, OpenAI released GPT-6 Astra. The big news is that the model can operate software through the screen and controls, just as a human does, instead of requiring a dedicated technical connection (API integration) for each individual program.
In OpenAI's demonstrations, Astra fills out online forms, updates customer records in CRM systems, edits spreadsheets and does research on the web. It also controls engineering programs such as KiCad and FreeCAD. According to OpenAI, it can produce finished documents, spreadsheets and presentations and carry out multistep workflows.
Astra is also the first OpenAI model rated "Critical" for cybersecurity under the company's safety framework, the Preparedness Framework. OpenAI says the model found two previously unknown security vulnerabilities (so-called zero-days) during testing and reported them to the maintainers. According to Constellation Research, it was trained on more than 100,000 GPUs (powerful computing chips) at OpenAI's Stargate facility in Texas.
What's smart
The smart part is that Astra can work in the programs companies already use, without anyone having to build a connection first. On OSWorld 2.0, a computer-use benchmark, Astra takes about 40 minutes per task versus about 75 minutes for GPT-5.6 Sol. That is roughly 47 percent less time, according to OpenAI.
In the developer tool Codex, Astra can keep notes across context windows (the amount of text a model can hold in mind at one time), so earlier windows remain searchable. The feature is experimental and must be enabled in Codex's configuration file.
OpenAI also highlights safety. According to the company, Astra refused unauthorized tasks in 100 percent of the tested cases versus 48 percent for Sol. Greg Brockman, president of OpenAI, says: "Astra is our most intelligent and, also very important, our most aligned model to date." There is a downside, though: Astra's written reasoning is harder to monitor because it solves problems in fewer written steps. OpenAI calls it serious and says it remains a research priority.
Cheaper or better?
Better, according to OpenAI's own measurements. On an offline subset of OSWorld 2.0, Astra scores 72.6 percent versus 65.7 for Sol. On Terminal-Bench 4.0, the figure is 57.7 versus 37.3 percent. Anthropic's Claude Fable 5.1 is reported at around 55.8 percent, but the sources disagree on the exact figure. On the cybersecurity benchmark ExploitBench, Astra scores 100 percent versus 78.5 for Sol, and on SRE-Bench, which deals with operational failures, it succeeds 88 percent of the time on the first attempt versus 55.9.
DataNorth assesses that Astra leads clearly in computer use, cybersecurity and advanced math, but is on par with the field in general reasoning and everyday coding. All figures are OpenAI's own and have not been independently verified, and the competitors' results were partly run by OpenAI itself with different settings than the competitors' own.
API pricing is $10 per million input tokens and $50 per million output tokens. Tokens are the chunks of text the model reads and writes. Fast mode costs double, meaning $20 and $100, and delivers up to twice the speed according to DataNorth (iClarified states 2.5 times). Pricing for Sol for comparison is not given in the sources, and there is no separate monthly price for ordinary users.
What it's good for
Astra is most clearly aimed at work that spans many programs: office tasks such as forms, spreadsheets and customer systems, engineering work and IT operations. It is also strong at security work, but the offensive cybersecurity capabilities remain in the Daybreak Blue program. The public version has multiple layers of protection.
The rollout starts with a small group of business customers in the Daybreak program. Paid ChatGPT plans (Plus, Pro, Business and Enterprise), the API and Amazon Bedrock follow over the coming days. Administrators must enable the model themselves, as it is off by default. The sources do not say whether or when Astra will be available in Denmark or the EU, and a general availability date has not been given.
Sources
Found via Kunstig Intelligens
Get the week's AI news in your inbox
Choose your level, topics and length. One email a week, unsubscribe at any time.
Subscribe to Promptly Newsletter



