On September 3, OpenAI unveiled its latest AI model, GPT-6 Astra, initiating a rollout for a select group of organizations through its Trusted Access program. Access will soon be extended to users of ChatGPT Plus, Pro, Business, and Enterprise, as well as to those using OpenAI API and Amazon Bedrock.
— OpenAI (@OpenAI) September 3, 2026
Additionally, Microsoft has started offering the model via its limited access Microsoft Foundry program on Azure, with plans to broaden access for participating clients shortly.
According to OpenAI, GPT-6 Astra outperforms GPT-5.6 Sol across various tests related to computer interaction, programming, scientific tasks, and cybersecurity, although some benchmarks indicate that models from Anthropic yield superior results.
OpenAI clarified that the comparative metrics presented reflect the highest scores achieved under any computational effort level. The tests on GPT models were conducted in research environments or through API, meaning the performance metrics for user-accessible versions may vary.
Enhanced Capabilities
As per the release, GPT-6 Astra can autonomously navigate browsers and applications, enabling it to fill out forms, update CRM records, search for information, edit documents, analyze data, and test software.
In the offline subset OSWorld 2.0, which assesses task execution in graphical interfaces, Astra scored 72.6%, compared to 65.7% for GPT-5.6 Sol. In latency simulations, OpenAI Astra completed tasks approximately 47% faster, taking around 40 minutes per task versus 75 minutes for GPT-5.6 Sol.
Source: OpenAI.In the Agents’ Last Exam, where AI tackles complex professional tasks in real programs, Astra achieved a score of 59.3%, while GPT-5.6 Sol scored 53.6% and Claude Opus 5 reached 55.5%.
Source: OpenAI.In ScreenSpot-Pro, which evaluates understanding of computer interface elements, Astra recorded 92.7% compared to 76.9% for GPT-5.6 Sol.
Source: OpenAI.OpenAI also introduced an experimental context retention mechanism in Codex, allowing Astra to maintain notes across filled context windows and search for requirements or test outcomes from previous messages and results.
This feature currently needs to be activated manually in config.toml, but OpenAI plans to make it standard for Astra in the coming weeks.
Code, Math, and Science
In Terminal-Bench 4.0 for agent programming, GPT-6 Astra scored 57.9%, whereas GPT-5.6 Sol achieved 37.3%, and Claude Fable 5.1 scored 55.8%.
Source: OpenAI.On FrontierCode 1.1 Extended, the new model reached 64.5%, and on the scientific Terminal-Bench Science 0.1, it scored 64.6%, compared to 22.4% for GPT-5.6 Sol and 52.6% for Claude Fable 5.1.
In FrontierMath Tier 4, the model obtained 97.6%. Developers reported that Astra helped improve two results in prime number theory: it reduced the upper limit of the gap between infinite pairs of prime numbers from 240 to 186, and modified one of the terms in the assessment of large gaps between prime numbers, a term that had remained unchanged for over 80 years.
Source: OpenAI.In Humanity’s Last Exam with Astra, the score was 57.2%, falling short of Claude Fable 5.1's 65%.
Critical Cyber Capabilities
GPT-6 Astra has become the first OpenAI model to achieve a Critical level in cyber capabilities within the Preparedness Framework. Without production constraints, it scored 100% on ExploitBench, compared to 78.5% for GPT-5.6 Sol. On ExploitGym, the scores were 42.4% and 30.3%, respectively.
In an internal ExploitBench set (June-August 2026), Astra executed arbitrary code in 39% of cases among 20 recently disclosed high-severity vulnerabilities in V8, covering 13 stable versions of Chrome. In contrast, GPT-5.6 Sol showed an 11.5% success rate.
Source: OpenAI.In expert trials, Astra utilized undisclosed vulnerabilities to execute arbitrary code in secured browsers and create exploits for privilege escalation in secure operating systems.
"Astra's ability to identify and develop exploits for zero-day vulnerabilities could help find and mitigate weaknesses but simultaneously necessitates stricter security measures," the company stated.
The released version includes additional restrictions. Astra will refrain from advanced cyber tasks, including the creation of proof-of-concept exploits. The company plans to gradually extend broader access for protective scenarios through OpenAI Daybreak.
Developers have also implemented extra monitoring for the model's reasoning and actions. A classification system is designed to identify potentially unauthorized behavior and automatically halt relevant actions.
OpenAI acknowledged that Astra's written reasoning chain is less amenable to monitoring than that of GPT-5.6 Sol. This conclusion is primarily based on competitive tests where the model was directly prompted to evade observation.
The company attributes this result to Astra's improved ability to handle written reasoning in simpler tasks, solving them with fewer text steps. However, for complex tasks, the model is reportedly still struggling to conceal the reasoning necessary for resolution.
$1 Billion for Cybersecurity
Alongside the release of Astra, OpenAI announced a $1 billion initiative called Daybreak for Frontline Defenders. This fund includes subsidized access to models and Daybreak products, training, technical support, and partnership programs.
The initiative will initially launch in the U.S. for operators in water supply and electrical grids, state and municipal authorities, regional and local banks, non-profit organizations, and open-source developers with limited cybersecurity resources. The program is expected to expand to partner countries in the coming weeks.
"Defenders have a window of opportunity: it is narrowing, and AI needs to be leveraged to fill security gaps ahead of attackers," OpenAI stated.
Daybreak is already being utilized by professionals from about 2,000 approved organizations and workspaces. Daybreak Blue is designed for standard protective tasks, while Daybreak Red grants vetted organizations access to specialized cyber models for more sensitive scenarios.
Access and Pricing
Users of ChatGPT Pro, Business, and Enterprise will also gain access to a distinct version of GPT-6 Astra Pro. OpenAI did not specify how this version differs from the basic one. Enterprise administrators will need to enable Astra for their workspaces themselves.
For developers, the model will be available in OpenAI API under the identifier gpt-6-astra. According to the documentation, the context window is 1.05 million tokens, with a maximum response volume of 128,000 tokens. The knowledge cutoff date is April 30, 2026.
The standard pricing is set at $10 for 1 million input tokens and $50 for 1 million output tokens. Cached input costs $1 for 1 million tokens, while cache write is $12.5.
For requests exceeding 272,000 input tokens, the input and caching rates double, while the output cost increases by 1.5 times for the entire request. The accelerated mode operates up to twice as fast as the standard one and is priced at double the applicable rate.
It’s worth noting that in August, OpenAI temporarily slowed the scaling of new AI models and paused reinforcement learning training for the latest systems intended for future deployment for two weeks.