Performance Benchmarks and Computational Capabilities
OpenAI has released GPT-6 Astra. The model functions as a significant step in AI development, focusing on computer use, software engineering, and scientific research. Internal testing shows Astra reaching a 98% score on FrontierMath Tier 4 and a 99.9% success rate on ARC-AGI-3. These figures represent a clear shift from previous versions like GPT-5.6 Sol.
Latency results from OSWorld 2.0 demonstrate increased efficiency. Astra completes tasks in 47% less time than its predecessor. It achieves this by combining improved training methods with a faster Codex harness. The model handles complex, multi-step workflows such as organizing calendars, updating customer records, and troubleshooting screen errors without requiring constant user intervention.
Cybersecurity and Safety Architecture
Security is a core focus for this release. Astra hit a 100% score on ExploitBench during internal testing. While these capabilities allow the model to assist defenders in patching vulnerabilities, they also present potential risks. OpenAI implemented strict safeguards, including a 0% rate of unauthorized scope expansion in testing environments. This stands in contrast to the 48% rate observed in earlier iterations.
The organization maintains a defensive approach to these high-level cyber capabilities. Astra refuses to generate proof-of-concept exploits for vulnerabilities. Users may experience pauses in ChatGPT or Codex if the system flags a task as high-risk. These safety protocols are designed to prevent misuse while allowing legitimate research and security work. OpenAI plans to introduce less restrictive settings for specific defensive workflows in the near future.
Professional Deployment and Integration
Businesses can access Astra starting today across ChatGPT Plus, Pro, Business, and Enterprise plans. The model is also available through the OpenAI API, Microsoft Azure, and Amazon Bedrock. Enterprise administrators control access via their workspaces. The system is designed to adhere to business templates, ensuring that documents, spreadsheets, and presentations meet specific visual and narrative standards.
Developers using the API will pay a standard rate of $10 per million input tokens and $50 per million output tokens. A fast mode is offered at twice the standard price. The platform also includes support for Zero Data Retention for eligible customers. These deployment choices indicate a push toward integrating the model directly into enterprise IT infrastructure. The focus remains on providing usable, reliable artifacts that align with corporate needs.