OpenAI Launches GPT-6 Astra; Claims Lead Over Claude Fable 5.1 In Coding & Reasoning

OpenAI has officially launched GPT-6 Astra, its newest flagship model, positioning it as a major leap in computer use, coding, cybersecurity and scientific reasoning, while publishing benchmark comparisons that put it ahead of Anthropic's Claude Fable 5.1 on several fronts.

Rollout and availability

GPT-6 Astra began rolling out to a limited set of organisations, with plans to become available to all ChatGPT Plus, Pro, Business and Enterprise users over the following days, alongside access through the OpenAI API, Microsoft Azure and AWS Bedrock. The company said Astra usage would be included within existing subscription allowances, with the option to purchase additional credits, and that companies enrolled in its application-based cybersecurity programme would get access first.

Pricing

On the API, OpenAI is charging standard rates of $10 per million input tokens and $50 per million output tokens, with separate, lower rates for cached input, and a 'Fast' processing mode priced at roughly double the standard rate for about twice the speed. Independent tracking of the launch pegged this as roughly 2.5 times the promotional pricing of OpenAI's prior model, GPT-5.6 Sol, with a context window of 1,050,000 tokens and a maximum output of 128,000 tokens.

What the model can do

OpenAI describes Astra as its most intelligent and aligned model to date, saying it is state-of-the-art on computer use, browsing, software engineering, cybersecurity, science and professional work. The company said the model saturates several of its toughest internal and third-party benchmarks, including near-perfect scores on tests of advanced mathematics reasoning and simulated computer environments.

On everyday computer tasks like filling forms, updating records, drafting research summaries, or running quality checks on a website; OpenAI said Astra completes work roughly 47 percent faster per task than its predecessor while scoring higher on accuracy in timed simulations. The company also highlighted gains in producing 'immediately usable' business documents, spreadsheets and slide decks that match a user's existing templates and style, and said the model is better at asking clarifying questions only when a decision genuinely depends on the answer, rather than interrupting work unnecessarily.

For coding, OpenAI called Astra its best model yet for software engineering, citing improved agentic coding performance and reduced need for developers to iterate on generated code.

Cybersecurity capability and new restrictions

OpenAI disclosed that Astra represents a significant jump in cyber capability, clearing the 'Critical' threshold for cybersecurity under its internal risk-management framework. The company said that on tests of turning known vulnerabilities into working exploits, Astra scored a perfect result compared with a considerably lower score for its predecessor, and that during testing it independently found two previously unknown security flaws, which OpenAI said it is now disclosing to the affected software maintainers. Because of this capability jump, OpenAI said the publicly released version of Astra refuses more advanced cybersecurity requests, such as building proof-of-concept exploits, with plans to loosen those restrictions gradually for vetted users through its Daybreak programme.

Comparison with Claude Fable 5.1

OpenAI's own published benchmark tables set Astra directly against Anthropic's Claude Fable 5.1 across categories including coding, professional work, science, cybersecurity and abstract reasoning. By OpenAI's figures, Astra outperformed Fable 5.1 on measures such as a terminal-based science-workflow test (64.6 percent versus 52.6 percent) and a 3D-modelling reconstruction benchmark (95.9 percent versus 84.3 percent), while claiming lower estimated API cost per task in several of these categories. On a general coding benchmark, OpenAI's published numbers showed Astra slightly ahead of Fable 5.1, while on a broader intelligence index, Fable 5.1 was shown scoring higher than Astra.

Separately, Al Jazeera reported that OpenAI's own announcement claimed Astra achieved perfect or near-perfect scores in key reasoning benchmarks, which the company said beat both GPT-5.6 Sol and Claude Fable 5.

OpenAI's release comes against the backdrop of a security incident at Hugging Face that had prompted the company to delay this launch to add further safeguards.