GPT-6 Astra Explained: Features, Capabilities, Benchmarks, Pricing

OpenAI introduced GPT-6 Astra on September 3, 2026. The new multimodal large language model is available through the API and to ChatGPT Plus, Pro, Business, and Enterprise users. It was initially released to Daybreak program testers before becoming broadly available on September 4. The model supports text, code, limited image handling, web search, and file search, with an API context window of 1,050,000 tokens.

Compared with its predecessor, GPT-6 Astra performs better on advanced professional, scientific, and cybersecurity tasks. It also offers a much larger context window and uses a new reasoning technique that OpenAI calls "recurrent depth." Benchmark results, including 99.9% on ARC-AGI-3 and 98% on Tier 4 FrontierMath, show gains over GPT-5.6. OpenAI also reports improved speed and reliability for longer workflows, such as document creation and prolonged system use.

For operators, owners, and compliance leads in regulated industries, these improvements could make multi-step automation, document and knowledge management, and professional research more reliable. At the same time, teams considering GPT-6 Astra will need to weigh its capabilities and pricing against its increased safeguards, new safety and monitoring constraints, and partial restrictions on cybersecurity capabilities, as well as how it compares with other available models.

What Is GPT-6 Astra and Who Built It?

GPT-6 Astra is a large language model built by OpenAI, announced on September 3, 2026. Astra is OpenAI’s current flagship model, available for API use and within the ChatGPT suite for Plus, Pro, Business, and Enterprise users. The model supports structured outputs, streaming, function calling, file and image input, prompt caching, web search, and exceptionally large context windows. It was pretrained on more than 100,000 GPUs at OpenAI’s Stargate site in Texas, using a technique called recurrent depth as part of its reasoning methods.

Key Capabilities and Technical Features

GPT-6 Astra improves on both general and specialist tasks compared to prior OpenAI models. It supports: a 1,050,000 token context window (allowing long documents and workflows), up to 128,000 output tokens, and input types including text, code, and images. Key features include streaming outputs, structured data generation, highly scalable function calling, file and web search, and prompt caching.

Performance areas where GPT-6 Astra expands beyond earlier GPT models include enhanced software engineering, cybersecurity, scientific computing, document preparation, and understanding user intent across longer, multi-step tasks.

• Maximum 1,050,000 token context window
• 128,000 output tokens supported
• Text, code, and image input
• Real-time streaming and structured outputs
• Integrated file search and web retrieval
• Extensive function calling and automation hooks
• Prompt caching for repeated workflows
• Pretrained on over 100,000 GPUs with recurrent depth reasoning

Benchmarks and Comparative Performance

On standardized evaluations, OpenAI states that GPT-6 Astra delivers record scores compared to prior models. Astra achieves 98% on FrontierMath Tier 4, 99.9% on the ARC-AGI-3 reasoning task, and 100% on ExploitBench testing for basic exploit detection ability. In OpenAI’s internal use and reported user testing, the model has shown greater reliability for document automation, software engineering, cybersecurity analysis, and extended dialog.

OpenAI reports improved intent understanding, multi-step reasoning, and output quality for presentation and document workflows—especially at large scale.

• FrontierMath Tier 4: 98%
• ARC-AGI-3: 99.9%
• ExploitBench: 100% (basic exploit tasks)
• Improvements in handling long, multi-step tasks and intent recognition

Pricing and Availability

GPT-6 Astra became available to selected organizations and Daybreak testers on September 3, 2026, and to most ChatGPT Plus, Pro, Business, and Enterprise users by September 4, 2026. The model is also available via the OpenAI API and through AWS for integration.

Pricing is set at $10 per 1 million input tokens and $50 per 1 million output tokens through the API. The model’s knowledge is current as of April 30, 2026. Access is subject to OpenAI's terms and may vary for high-volume or enterprise use.

• $10 per 1M input tokens
• $50 per 1M output tokens
• Available via ChatGPT Plus and above, API, and AWS
• Knowledge cutoff: April 30, 2026

Best Use Cases and Industry Fit

Strong candidates for GPT-6 Astra include business automation, document and contract generation, technical analysis, multi-step professional workflows, and regulated-industry support for compliance analysis or reporting. The large context window enables end-to-end workflows that previously required model chaining or external memory.

Sectors with multi-document compliance, audit, or research needs—such as law, finance, insurance, and regulated healthcare—can benefit from Astra’s expanded reasoning ability and reliability at scale. The model’s safety updates and restricted advanced cybersecurity features are especially relevant for organizations subject to strict security requirements.

• Complex document automation (contracts, reports, filings)
• Professional operations research and legal analysis
• Extended business process automation (regulatory, insurance, audit)
• Large-scale code, cybersecurity, or scientific workflows

Limitations, Safety, and Monitorability

Astra brings improvements, but there are real constraints—particularly for compliance-focused users. OpenAI confirms that some advanced cybersecurity functionalities are withheld and that Astra is the first OpenAI model to reach the Critical level under its Preparedness Framework. The model’s recurrent depth reasoning method obscures parts of its decision process, which some researchers have flagged as a challenge for monitorability and explainability.

Operators in regulated sectors should note: Astra is not fully auditable, and advanced cyber capabilities are not user-facing at launch. OpenAI’s safety documentation and system card outline the model’s boundaries and safeguards, and ongoing updates address regulatory, audit, and institutional review needs.

How GPT-6 Astra Compares to Main Alternatives

Compared to earlier OpenAI models, GPT-6 Astra provides larger context, stronger professional task scores, and better reliability for long, multi-step jobs. Against competitors like Anthropic Claude or Google Gemini, Astra’s context size, integrated web and file search, and support for advanced automation are now differentiators, though monitorability and withheld cyber capabilities may matter more in sensitive industries.

Teams looking to choose between GPT-6 Astra and alternatives should weigh price, documented safety controls, output quality for compliance workloads, and ongoing regulatory changes in model access and capabilities. Performance on compliance, data privacy, and explainability remains a leading decision factor.