OpenAI has announced its next generation flagship large language model, GPT-6 Astra, which it claims to be the world’s most intelligent and aligned AI system to date. The model went into limited preview on September 3, 2026 and began a broader rollout to paid users the following day.
Astra is a big step up from its predecessor family (including the previous flagship Sol), according to OpenAI. The model is state-of-the-art in the use of computers, web browsing, software engineering, cybersecurity, scientific research and professional knowledge work, the company says. It can work through multi-step workflows, stay on task, better understand user intent, and perform complex tasks on a computer or browser — like filling out forms, managing calendars, making appointments, searching for jobs, or even assisting with technical design work — with greater speed, accuracy, and reliability than past systems.
Performance and Capability Benchmark
- 98% on FrontierMath Tier 4 (near-saturation of difficult mathematical problems) The model is already contributing to solutions for long-standing open problems in mathematics.
- 99.9% on ARC-AGI-3, under some test conditions.
- 100% on ExploitBench
- It tops many benchmarks for computer use, agency, coding, and science, including new state-of-the-art results on long-horizon tasks and reverse-engineering problems.
The model has a context window of 1,050,000 tokens and supports a maximum of 128,000 output tokens. The model has a knowledge cutoff of April 30, 2026. OpenAI says it’s made “significant progress” on alignment: the model is better at staying within the bounds of what it’s allowed to do, and has lower rates of going off the rails or acting in ways it shouldn’t, compared to previous versions.
OpenAI president Greg Brockman has a bold vision for the release, saying it could be seen in retrospect as the start of the AGI age, which the company defines as systems that can do economically valuable work at or above human level. The training run was the largest the company has ever done, other executives said, with more than 100,000 GPUs involved.
Pricing and Availability
GPT-6 Astra will first be deployed to a limited number of organizations (including those in OpenAI’s early-access program, Daybreak). Access for ChatGPT Plus, ChatGPT Pro, ChatGPT Business and ChatGPT Enterprise subscribers will follow. It’s also available via the OpenAI API ( gpt-6-astra ), Microsoft Azure and AWS Bedrock. Access is being rolled out in stages, and some of the higher tier plans will get it before others. Advanced cybersecurity capabilities are still more limited.
API Pricing: $10 per million input tokens, $50 per million output tokens (with separate caching pricing). Enterprise and developer tools feature zero-data-retention for qualifying customers.
Context and perspective
It comes as the frontier AI race continues, and after OpenAI’s earlier 2026 models. Along with the capability jump, OpenAI also emphasizes safety and alignment improvements, saying the model was reviewed before a broader release. Astra has demonstrated early results in handling professional workflows, coding projects, scientific analysis and everyday computer tasks with less human supervision.
As the model is rolled out to more users, independent evaluations and real-world adoption will test claims by OpenAI of a generational leap in both intelligence and reliability. Users can refer to OpenAI’s official announcement and model documentation for the most recent information.
