OpenAI unveiled GPT-6 Astra with a context window of 1.05 million tokens
OpenAI unveiled GPT-6 Astra, a new flagship AI model designed for complex multi-step tasks. The company emphasizes programming, computer interaction, research, and autonomous task execution. OpenAI co-founder Greg Brockman called the launch the beginning of the “AGI” era.
What GPT-6 Astra Can Do
GPT-6 Astra is designed for extended autonomous operation and can independently interact with websites and applications. OpenAI states that the model can fill out forms, work with spreadsheets, search the internet for information, write and test code, create documents, and execute chains of related actions.
One of the main development directions is computer control. Instead of providing individual hints, the model can manage interfaces and complete tasks in full. Among the demonstrated scenarios, OpenAI showed job and housing searches, tax document preparation, game creation, and work with legal materials.
In API, web search, file search, image generation, Code Interpreter, shell, computer use, MCP, and other tools are supported.
Context of over one million tokens
GPT-6 Astra’s context window is 1.05 million tokens, with a maximum response size of 128,000 tokens. The model’s knowledge base is limited to April 30, 2026.
Astra accepts text and images, and can also generate text as output. Five reasoning intensity levels are available: low, medium, high, xhigh, and max.
Test results
According to OpenAI, GPT-6 Astra significantly outperforms GPT-5.6 Sol in several reasoning, programming, and tool-use tests.
In ARC-AGI-3, the model scored 98.6%; in FrontierMath Tier 4 v2, it achieved 97.6%; in GPQA Diamond, it reached 96%; and in BenchCAD, it obtained 95.9%. The result for DeepSWE v1.1 was 74.1%, and for Terminal-Bench Science, it was 64.6%.
The results of ARC-AGI-3 cannot be directly compared with all other models: the outcome depends significantly on the agent environment and testing settings used.
Cybersecurity
GPT-6 Astra became the first OpenAI model to be assigned a Critical level for cybersecurity capabilities under the Preparedness Framework.
According to OpenAI, with the necessary tools, Astra can independently discover previously unknown vulnerabilities and develop methods to exploit them in well-protected systems.
On ExploitBench, the model scored 100% even at the lowest reasoning level tested. On SRE-Bench, with four attempts, Astra solved 99.2% of tasks compared to 68.7% for GPT-5.6 Sol, while using roughly four times fewer output tokens.
Due to these capabilities, the most sensitive tools for vulnerability discovery and exploitation will not be available to all users. OpenAI is implementing additional safeguards and restricted access for cybersecurity tasks.
Price and availability
On September 3, OpenAI began providing GPT-6 Astra to corporate clients as part of a limited access program. In the coming days, the model will be available on API and for ChatGPT Plus, Pro, Business, and Enterprise subscribers.
Cost of API:
- $10 for 1 million input tokens;
- $1 per 1 million cached input tokens;
- $12,50 for 1 million tokens when writing to cache;
- $50 for 1 million output tokens.
For requests containing more than 272,000 input tokens, higher rates apply. OpenAI has not yet announced the availability of GPT-6 Astra for free accounts.
