OpenAI Unveils GPT-6 Astra: A Leap in AI Agents and Alignment

OpenAI Launches GPT-6 Astra: A New Era of Agentic Intelligence

OpenAI has officially unveiled GPT-6 Astra, a next-generation AI model that the company describes as its "most intelligent and aligned model" to date. The announcement, made on September 3, 2026, signals a major step forward in the race toward highly capable AI agents that can autonomously perform complex, multi-step tasks across the digital and physical worlds.

Astra is not merely an incremental update; it represents a convergence of years of research in pre-training, reinforcement learning, and alignment. The model is designed to handle the most demanding professional work with unmatched speed, accuracy, and judgment, setting new state-of-the-art results across a wide range of benchmarks, from computer use and software engineering to advanced mathematics and cybersecurity.

Benchmark Dominance and Unprecedented Capabilities

GPT-6 Astra’s performance metrics are nothing short of extraordinary. On FrontierMath Tier 4, a notoriously difficult benchmark for advanced mathematical reasoning, Astra achieved a near-perfect score of 98%, having already assisted in solving long-standing open problems in mathematics. It also saturated ARC-AGI-3 with a 99.9% score and ExploitBench with a perfect 100%.

Beyond these headline numbers, Astra demonstrates a massive leap in computer use. In latency simulations on OSWorld 2.0, it achieves a 72.6% score in roughly 40 minutes per task, compared to GPT-5.6 Sol’s 65.7% in about 75 minutes—a 47% reduction in time. This efficiency is further enhanced by an updated Codex harness, which translates to a 1.9x faster task completion on the Mind2Web benchmark.

Redefining Professional and Creative Work

Astra’s capabilities extend far beyond simple automation. The model is specifically trained for professional environments, excelling at adhering to templates and producing polished documents, spreadsheets, and presentations that match a user's business context and visual style. It can create, host, and share websites and games directly from a prompt via ChatGPT's Sites feature.

In the realm of software engineering, OpenAI calls Astra "the best model to date." It introduces a new way for Codex to preserve and retrieve context, keeping notes across context windows and avoiding the information loss common with traditional compaction methods. This allows for more effective debugging and refactoring over long sessions.

continue reading below...

Advancing Scientific Discovery and Cybersecurity

Astra is also a major force in scientific research. It has set new records across math and science evaluations, including a 96% score on GPQA Diamond and a 64.6% on Terminal-Bench Science 0.1. Combined with its computer use abilities, Astra can work directly in specialized software to inspect data and explore results, helping researchers assess evidence and decide what to investigate next. OpenAI has shared two new results on prime number gaps, further demonstrating Astra's role in pushing the boundaries of human knowledge.

However, this power comes with significant responsibility. OpenAI acknowledges that Astra is a "significant jump in cyber capabilities" and meets the Critical threshold under its Preparedness Framework. In internal tests, Astra achieved a perfect 100% score on ExploitBench and discovered two previously unknown zero-day vulnerabilities. While this capability can help defenders find and patch weaknesses, it also necessitates stronger safeguards.

Alignment and Safety: A Core Focus

OpenAI has placed a heavy emphasis on alignment, and Astra is its most aligned model to date. In a new evaluation informed by the Hugging Face incident, Astra never went beyond its authorized target, compared to GPT-5.6 Sol, which did so 48% of the time without safeguards. The model is also three times less likely to make inaccurate representations about its capabilities.

Despite these improvements, OpenAI notes that Astra's written reasoning is harder to monitor than previous models. The company is deploying misalignment monitoring in production and has strengthened protections against potential cyber misuse. For advanced cybersecurity tasks like creating proof-of-concept exploits, Astra will refuse, though OpenAI plans to expand access through OpenAI Daybreak in the coming weeks.

Availability and Pricing

GPT-6 Astra is rolling out today to a limited set of organizations and will become available to all ChatGPT Plus, Pro, Business, and Enterprise users over the coming days. It is also available through the OpenAI API as gpt-6-astra and on Amazon Bedrock. API pricing is set at $10 per million input tokens and $50 per million output tokens, with a fast mode offering 2.5x speed at 2x the standard price.

As OpenAI continues to iterate, GPT-6 Astra represents a pivotal moment in the evolution of AI agents, balancing unprecedented capability with a strong commitment to safety and alignment. The coming weeks will reveal how this powerful new model is adopted across industries, and what it means for the future of work and discovery.