OpenAI unveiled GPT-6 Astra on September 3, 2026, calling it the company's most capable and aligned model yet and marking what President Greg Brockman described as "the start of the AGI era." The release represents the first model in the GPT-6 generation and arrives just two days after competitor Anthropic launched Claude Fable 5.1 on September 1.

Benchmark Dominance

GPT-6 Astra demonstrated state-of-the-art performance across multiple benchmarks. According to OpenAI's official announcement, Astra scored 97.6% on FrontierMath Tier 4 v2 — the toughest tier of the research-grade mathematics benchmark — compared to 87.8% for both Claude Fable 5.1 and its predecessor. The model also achieved 99.9% on ARC-AGI-3 and a perfect 100% on ExploitBench.

On cybersecurity tasks, Astra scored 72.6%, a significant jump from the 65.7% achieved by GPT-5.6 Sol, while reducing the average time per task from approximately 75 minutes to 40 minutes, The New Stack reported.

Cybersecurity Concerns

The launch has drawn scrutiny over the model's advanced cyber capabilities. Reuters reported that OpenAI cautioned about Astra's cybersecurity potential before the rollout, noting that the model is the first to reach what the company internally classifies as a "Critical" cybersecurity threshold.

OpenAI has begun a phased rollout, with companies participating in early access programs receiving the model first. The company published a detailed safety overview alongside the launch, emphasizing alignment work and frontier safeguards.

The AI Arms Race Intensifies

The back-to-back releases from Anthropic and OpenAI signal an accelerating competitive cycle in the frontier AI space. Coursiv's comparison notes that while Astra leads on most comparable benchmarks against Claude Fable 5.1, Claude Fable 5, and Claude Opus 5, the margins vary by category.

OpenAI characterized Astra as representing "a significant increase in cybersecurity capabilities" compared to previous models, describing it as both more token-efficient and more capable across computer use, coding, scientific research, and AI-powered automation. The model can reportedly control user computers directly, a feature that expands its practical utility but also raises novel safety questions about autonomous agent behavior.

What Comes Next

The phased deployment approach suggests OpenAI is proceeding cautiously despite its bold AGI claims. Industry observers will be watching how enterprise customers integrate Astra's computer-use capabilities and whether the model's cybersecurity strengths prompt regulatory attention from governments already grappling with AI oversight frameworks like the European Union's AI Act.

With Anthropic's Claude Fable 5.1 launched just 48 hours earlier and Google's next-generation Gemini models expected later this year, the frontier AI competition shows no signs of slowing.