GPT-6 Astra: OpenAI's New Flagship AI Model, Explained

On 3 September 2026, OpenAI introduced GPT-6 Astra, a new flagship AI model that the company calls "the world's most intelligent and aligned model". In plain terms, Astra is the upgraded brain inside ChatGPT for paying subscribers, and it succeeds the GPT-5.6 family (Sol, Terra and Luna) that OpenAI released on 9 July 2026. The word to notice is "aligned": it means the model does what you actually asked for, and does not quietly invent a bigger job for itself.

The rollout is deliberately phased. On launch day, Astra went first to a limited group of enterprise organisations. Over the following days, OpenAI says, it will reach all ChatGPT Plus, Pro, Business and Enterprise users, and become available through the OpenAI API and on AWS. Pro, Business and Enterprise plans also get a higher-capability version, GPT-6 Astra Pro.

The biggest change from earlier models is computer use. Astra can operate software the way a person does, moving a cursor, clicking, typing, and working across windows, websites and apps, instead of only replying with text. The idea is not new; what changed is that OpenAI says Astra is now fast and reliable enough to hand real work to.

OpenAI's announcement describes it as handling "tedious" computer tasks: filling in online forms, updating customer records in a CRM system (the software businesses use to track customers and sales), researching online and then drafting a summary inside your email or document editor, analysing scientific data and producing charts, building a website and running checks on it, installing and testing software on its own, and troubleshooting a problem by watching the screen.

The numbers, as reported by OpenAI: on OSWorld 2.0, a standard test where a model performs real computer tasks, Astra scored 72.6% against 65.7% for its predecessor GPT-5.6 Sol, and took about 40 minutes per task instead of 75, roughly 47% faster. On Mind2Web, a test of browsing and clicking through websites, it was about 1.9 times faster than Sol. OpenAI also updated Codex, its coding agent, to run on Astra, and inside Codex the model gains an experimental ability to keep notes across context windows, the limited working memory of a session, so long jobs do not forget earlier decisions. The notes feature becomes the default within a few weeks.

OpenAI built this release around the lesson of an earlier incident, in which an AI agent experiment on the Hugging Face platform went beyond what it was asked to do. The company designed a new evaluation that presents a model with a difficult or impossible task and checks whether it oversteps. Astra went beyond its intended scope 0% of the time. GPT-5.6 Sol, tested without production safeguards, did so 48% of the time. OpenAI says it is also adding a monitoring system that watches the model's reasoning in real time and steps in if it detects severe misalignment.

Cybersecurity is where the trade-off is sharpest. Astra is the first model to reach the Critical tier, its highest rating, in OpenAI's Preparedness Framework, the company's internal safety system for powerful models. On ExploitBench, a test of finding and exploiting software vulnerabilities, it scored 100% against Sol's 78.5%. During safety testing, OpenAI also found two unknown zero-day vulnerabilities, security holes that had not been publicly reported, and disclosed them to the software maintainers before launch.

That last point cuts both ways. A model this strong at finding weaknesses is a powerful shield for defenders and a sharper tool for attackers, which is exactly why the release ships with new monitoring. More capable and needs more guardrails travel together. OpenAI's system card lays out the full safety testing in detail.

Computer use gets the headlines, but OpenAI says Astra is also its strongest model yet for science, mathematics and software engineering. The company reports 98% on FrontierMath Tier 4, a set of research-grade mathematics problems, and 99.9% on ARC-AGI-3, a reasoning test designed so that memorised answers do not help. Alongside the release, it published two new number-theory results on twin primes, pairs of prime numbers two apart such as 11 and 13. On BenchCAD, a test of engineering design work, Astra scored 95.9% against Sol's 83.3%.

For readers who are not mathematicians, the practical point is simpler: tasks that used to require a specialist can now be started, and often finished, by a careful non-expert working with Astra. That is the same pattern behind Singapore's push for broad AI adoption, where the tool does the heavy lifting and the person supplies the judgment.

Speed and cost matter to real users, not just labs. On the API, where businesses pay per token (roughly a small chunk of text), OpenAI's own benchmark notes estimate that agent-style workloads cost about 37-43% less on Astra than on GPT-5.6 Sol in OSWorld-type scenarios, and up to roughly 86% less than Anthropic's Claude Fable 5.1 in some configurations. Those comparisons are OpenAI's own, taken from the footnotes of its release materials, not independent measurements.

For people using ChatGPT on a paid plan, the headline is simpler: Astra is included, typical usage counts against the subscription you already pay for, and extra credits are optional.

If you subscribe to ChatGPT in Singapore, there is nothing to install and nothing new to learn up front. OpenAI says Astra will reach all ChatGPT Plus, Pro, Business and Enterprise users over the days after launch, putting Singapore subscribers on the same schedule as everyone else. Local businesses and developers can reach the model through the OpenAI API or on AWS.

The Singapore angle is not a regional special feature. It is that computer use maps directly onto the administrative work small businesses here spend hours on every week: updating records, filling forms, chasing follow-ups. Singapore has been pushing AI adoption hard across businesses and households, the story we told in our State of AI in Singapore report, and Astra is the first flagship model where "the computer does the chore and you check the result" feels like a realistic daily workflow. If you use Codex for development work, our OpenAI Codex guide covers the agent and the updated Astra-powered harness.