On September 3, 2026, amid a wave of next-generation LLM releases, OpenAI officially unveiled its latest frontier model GPT-6 Astra (Astra). The company described it as a "generational leap" in software engineering, professional workflows, scientific discovery and cybersecurity, calling it "the world's most intelligent and most aligned model." OpenAI President Greg Brockman said at the launch that Astra's debut marks the industry stepping into the "general artificial intelligence (AGI) era." The model is currently available as a limited preview to trusted partners, with plans to roll it out in the coming days to ChatGPT Plus, Pro, Business and Enterprise users and via API, Azure and AWS.

Details: Capability Gains and Technical Breakthroughs

Astra integrates years of research across pretraining, reinforcement learning and alignment, and OpenAI described it as one of the largest training runs to date (using more than 100,000 GPUs). Key highlights focus on the following areas:

Computer Use and Professional Workflows

Astra reaches new heights on computer-operation, browser-use and cross-professional workflow benchmarks, including Agents’ Last Exam, AutomationBench and ScreenSpot Pro. The company says it completes tasks about 47% faster than GPT-5.6 Sol, scoring 72.6% on tasks averaging roughly 40 minutes. The model is better at staying focused, respecting task boundaries, understanding user intent, handling tedious steps and completing multi-step workflows, enabling users to delegate tasks with greater confidence.

Software Engineering and Coding

It sets new records on coding and terminal benchmarks such as TerminalBench-4.0, showing stronger long-horizon software engineering capabilities and the ability to autonomously manage complex codebases, debugging and multi-file collaboration.

Science and Mathematics

On FrontierMath Tier 4 it achieved a 98% saturation score and has reportedly helped resolve several long-standing mathematical problems; ARC-AGI-3 reached 99.9% (with new adapters approaching a human-efficiency baseline of about 96%). Science-related benchmarks Terminal-Bench Science 0.1 and HealthBench Pro also lead their categories. OpenAI emphasizes Astra's potential to generate new knowledge, a trait the company views as consistent with AGI characteristics.

Cybersecurity

ExploitBench recorded a 100% score. Given the model's powerful capabilities, advanced security functions are initially available only to vetted defenders through the Daybreak program, using a restricted-access approach similar to Anthropic's Glasswing to manage dual-use risks. In a new evaluation measuring whether the model exceeds authorized boundaries, Astra performed markedly better than its predecessor (0% overstep vs. GPT-5.6 Sol's 48.2%).

Alignment and Safety

Astra is positioned as OpenAI's "most aligned" model, with substantial improvements in understanding user intent and controllability. Its release was delayed after the July Hugging Face incident, during which OpenAI added extra safety safeguards. The model has passed the U.S. government's voluntary review framework.

In specs, Astra features a context window of roughly 1.05 million tokens, a maximum output of 128,000 tokens, and a knowledge cutoff of April 30, 2026. Specific API pricing has not been fully disclosed in public materials but will be rolled out via ChatGPT and API.

Technical Context and Release Pace

Astra's launch follows intensive updates from Anthropic Fable 5.1, Meta Muse Spark 1.3 and others, reflecting heated competition among frontier models in the latter half of 2026. OpenAI positions it as "able to do anything you can do on a computer, and faster," and showed demos from 3D historical reconstruction to parallel-agent puzzle solving.<