Web Desk: OpenAI has introduced GPT-6 Astra, a new frontier artificial intelligence model designed to handle complex professional tasks, operate across computer applications and improve how businesses delegate work to AI systems.
The company said Astra combines advances in pre-training, reinforcement learning and alignment, positioning the model as a major step beyond traditional chat-based AI.
OpenAI said GPT-6 Astra achieved near-saturation results on several demanding benchmarks, including a 98% score on FrontierMath Tier 4 and 99.9% on ARC-AGI-3.
The model also recorded a 100% score on ExploitBench, according to the company. OpenAI said Astra has also contributed to solving previously unresolved mathematical problems.
Beyond benchmark performance, the company said the model delivers improvements in computer operation, web browsing, software engineering, cybersecurity, scientific research and other professional workloads.
OpenAI said the next phase of enterprise AI will increasingly focus on whether models can complete meaningful work rather than simply respond to prompts.
Astra can take an open-ended objective, break it into multiple steps, develop a plan and produce a finished result. It can also reassess its approach when users provide new instructions and use approved tools across different applications.
For businesses, OpenAI said this could allow AI systems to provide decision support, prepare documents and analyses, and carry out multi-stage workflows under human supervision.
Alongside its capability gains, OpenAI said Astra is its most aligned model to date, with improvements in understanding user intentions and staying within defined boundaries.
The company cited a new evaluation designed to test whether an AI system would exceed its authorized scope when confronted with difficult or impossible tasks.
According to OpenAI, GPT-5.6 Sol exceeded its assigned boundaries in 48% of tests when operating without production safeguards, while GPT-6 Astra did so in none of the cases tested.
OpenAI said the result reflects its effort to make increasingly capable AI systems more predictable and trustworthy when they act on behalf of users.
Astra can interact with computer interfaces and applications, including software that does not offer dedicated application programming interfaces, or APIs.
The model can interpret information displayed on a screen and perform approved actions such as updating records, navigating development environments, testing software and compiling information into reports.
However, OpenAI and Microsoft stressed the importance of safeguards because information presented to an AI agent can be incomplete, misleading or deliberately crafted to influence its actions.
Microsoft Foundry provides controls for permissions, approvals, monitoring and human intervention, allowing companies to restrict agents to approved resources and require authorization for sensitive actions.
The companies said early enterprise use cases span several industries and functions.
In software development, Astra can help reproduce complicated bugs, investigate potential causes, suggest fixes and prepare changes for developers to review.
For business intelligence teams, it can assist with building and refining Power BI dashboards while helping analysts examine data and identify important trade-offs.
The model can also generate business documents, spreadsheets and presentations using established templates and organizational standards.
In other workflows, Astra can help process forms, update customer information, test websites and navigate approved software interfaces where direct API connections are unavailable.
Replit said access to Astra through Microsoft Foundry could allow developers to move beyond conventional code generation toward more active software creation.
Microsoft Foundry is designed to provide organizations with infrastructure for identity management, access controls, encryption, private networking, content filtering, monitoring and governance.
OpenAI said customer prompts and outputs are not used to train its models.
The companies nevertheless cautioned that technical safeguards do not eliminate all risks. Businesses remain responsible for determining which controls and oversight mechanisms are appropriate for their specific applications and regulatory requirements.
GPT-6 Astra is initially being offered to a limited group of organizations through Microsoft’s Foundry Limited Access Program.
OpenAI said access will expand in the coming days to ChatGPT Plus, Pro, Business and Enterprise customers, as well as users of the OpenAI API and platforms including Microsoft Azure and AWS Bedrock.
The model will be offered through standard deployment options, including global and U.S. data-zone configurations.
OpenAI listed separate prices depending on context length and deployment type.
For standard global deployments with short context, pricing is set at $10 per million input tokens, $1 for cached input, $12.50 for cached writes and $50 for output tokens.
Long-context global deployments cost $20 per million input tokens, $2 for cached input, $25 for cached writes and $75 for output.
US data-zone deployments are priced higher, ranging from $11 to $22 per million input tokens for input and $55 to $82.50 for output, depending on context length.
OpenAI said actual consumption and costs will depend on individual workloads and configurations.
The launch marks another step toward AI systems that can move from answering questions to planning, reasoning and carrying out complex tasks across the digital environments used by businesses.
Read more: Apple’s cheapest new iPhone could leave buyers stunned
