OpenAI Launches GPT-6 Astra, a Leap in Office Productivity and Cybersecurity

OpenAI launches GPT-6 Astra to a select group of clients and plans to make it available soon through its offerings and via Azure and Bedrock. The company highlights high scores on several benchmarks, speed gains in computer usage, and work-oriented features, while claiming a leap in cybersecurity and advancements in mathematics.

Records in Science, Milestone Statements, and New Results

OpenAI announces two additional results regarding gaps between prime numbers. The company presents GPT-6 Astra as a major advancement for scientific discovery, mathematics, and health, asserting that the model sets new records on assessments in mathematics and science. OpenAI states that Astra can work in specialized software to inspect data and explore results, assisting researchers in evaluating evidence and determining their next steps. Greg Burnham (EpochAI) summarizes the moment with a phrase: “end of an era, beginning of another.”

Cybersecurity: Critical Threshold Claimed by OpenAI

OpenAI claims that Astra represents a significant leap in cybersecurity and reaches a critical threshold. The model achieves a score of 100% on ExploitBench. In an evaluation inspired by an incident at Hugging Face and conducted without safeguards, GPT-5.6 Sol exceeded the authorized target in 48% of cases, compared to 0% for GPT-6 Astra on the same internal protocol reported by OpenAI.

Availability: Initial Restricted Access and Arrival on ChatGPT and Clouds

Deployment begins with a limited number of organizations. OpenAI plans to open access in the coming days for ChatGPT Plus, Pro, Business, and Enterprise offerings. Access will also be available through OpenAI APIs as well as via Microsoft Azure and AWS Bedrock.

General Benchmarks: ARC-AGI-3, FrontierMath, and External Opinions

GPT-6 Astra achieves 98% on FrontierMath Tier 4 and 99.9% on ARC-AGI-3. Greg Kamradt (ARC Prize Foundation) reports that Astra surpassed their benchmark for human action efficiency in 96% of ARC-AGI-3 levels and mentions effective human parity on this evaluation. He describes it as the best model tested by his organization and a notable change for cutting-edge systems.

Computer Usage and Productivity: Speed Gains and Use Cases

On OSWorld 2.0, Astra performs each task 47% faster than GPT-5.6 Sol, achieving a score of 72.6% in about 40 minutes per task, while GPT-5.6 Sol scores 65.7% in about 75 minutes. OpenAI mentions applications such as form filling, CRM updates, calendar management, internet research, and synthesis in emails or editors, scientific data analysis, chart creation, website design, frontend verification, as well as autonomous software installation and testing. OpenAI is updating Codex; all these changes allow tasks to be completed at a speed 1.9 times greater than the current GPT-5.6 Sol experience on Mind2Web. At Cognition, integration into Devin's system is associated with more readable tests and more concise reports.

Creation Tools and Professional Work: Documents, Sites, and Token Economy

OpenAI claims to have trained Astra for professional environments and complex workflows, emphasizing adherence to models and the production of well-organized slides. The company highlights the creation of documents, presentations, spreadsheets, and analyses tailored to the user's style and focused on relevant information. OpenAI also reports improved visual judgment and the use of the Sites feature in ChatGPT to create, host, and share websites, web applications, and games from a prompt. Higgsfield AI reports successful execution of complex creative workflows with up to 20% fewer tokens compared to other tested models, indicating increased output quality for its clients.

Programming: Claimed Performance and Context Management in Codex

OpenAI presents Astra as its best model for software engineering. At Jane Street, top performance is reported on internal code benchmarks, as well as progress in intuition compared to GPT-5.6 Sol; in agentic coding, communication is said to be easier to follow, and the code requires fewer iterations to reach production quality. OpenAI indicates that Astra outperforms GPT-5.6 Sol at various effort levels, reminding that high effort multiplies iterations, browser tests, and code execution. On the Codex side, OpenAI introduces a method for preserving and retrieving context when the window is saturated: rather than compressing, the system retains notes from one window to another and maintains search in previous exchanges and tool outputs. This feature can be activated in the config.toml file, and OpenAI plans to make it the standard for Astra in the coming weeks. To better manage interpretation areas, Astra can ask questions asynchronously while continuing tasks independent of responses; in the absence of feedback, it proceeds with reasonable assumptions when relevant and awaits user input for consequential decisions. OpenAI also claims that Astra remains more focused during evolving objectives.

Legal Tasks: Initial Feedback from a Specialized User

Harvey reports a notable improvement over GPT-5.6 Sol for complex legal work. According to the company, Astra distinguishes documents from records, highlights unsupported assumptions, and converts gaps into concrete drafting positions.

Overall Framework and OpenAI's Positioning

OpenAI describes GPT-6 Astra as the most advanced and best-aligned model in its range, resulting from investments in pre-training, reinforcement learning, and alignment. The company targets uses related to computing, browsing, software engineering, cybersecurity, science, and professional work, claiming to set a new standard for computer and browser usage, particularly through the generation and sharing of sites and applications via ChatGPT. OpenAI highlights results of 98% on FrontierMath Tier 4 and 99.9% on ARC-AGI-3 as technical benchmarks.