OpenAI releases GPT-6 Astra, claims system approaches AGI and crosses critical cyber threshold

OpenAI released GPT-6 Astra on September 3, 2026. [4] The model was trained using more than 100,000 GPUs at the company's Stargate facility in Texas. [6] This marks the first time OpenAI used other models in a significant role to supervise a model's training. [6] The system relies on a looped reasoning technique called recurrent depth, which safety researchers have flagged as making the model's chain of thought harder for external observers to monitor. [15]

During a press briefing, OpenAI president Greg Brockman described the release as a generational leap and ended by saying, "Welcome to the AGI era." [6] The system is rolling out initially to a limited set of organizations through OpenAI's Daybreak Access program, with broader availability planned for ChatGPT Plus, Pro, Business, and Enterprise customers. [4][6] OpenAI CEO Sam Altman stated that future models will force companies to move more cautiously as their capabilities increasingly outpace the safeguards designed to control them. [16]

OpenAI's launch materials call Astra the world's best computer use model. [7] The model is designed to navigate software across browsers, desktop applications, and spreadsheets without requiring developers to build dedicated API integrations. [7] A demonstration showed the model using Google Maps to locate a pediatrician, visit the provider's website, and fill out a contact form. [12]

Astra outperforms its predecessor, GPT-5.6 Sol, as well as Anthropic's Fable 5 models in benchmarks. [8] In internal testing, the model scored 98 percent on FrontierMath Tier 4 and 100 percent on ExploitBench, a metric for building working exploits from known software vulnerabilities. [4] The ARC Prize organization confirmed Astra scored 62.7 percent on the ARC-AGI-3 Semi-Private evaluation at a cost of $26,000 per run. [15] For API developers, the standard pricing is set at $10 per million input tokens and $50 per million output tokens. [4]

Astra is the first OpenAI model designated as reaching a critical cybersecurity risk threshold under the company's Preparedness Framework. [6][9] The model can potentially identify previously unknown vulnerabilities and develop exploit chains across protected systems without step-by-step human guidance. [6][14] Because of these capabilities, OpenAI slowed down Astra's release to conduct additional safety testing. [6] The most advanced cybersecurity functions remain restricted to a limited group of enterprise testers. [14]

Alongside the launch, OpenAI announced a $1 billion initiative called Daybreak for Frontline Defenders. [5] This program provides subsidized access to models and technical support for organizations operating critical infrastructure, such as water systems, electricity providers, and local governments. [5][11] As part of the program, OpenAI is starting a pilot with the Multi-State Information Sharing and Analysis Center to train cyber defenders at state and local governments on using its models. [5]

Astra's ability to operate software directly shifts the enterprise AI paradigm from conversational interfaces to agents that execute workflows natively across standard desktop applications. By heavily leveraging other AI models to supervise Astra's training, OpenAI is demonstrating a path to scaling capabilities beyond human oversight limits. Astra's benchmark results specifically target rivals like Anthropic's Fable 5 and Claude Opus 5, signaling an intensifying race around autonomous computer use. The $1 billion Daybreak initiative highlights a growing recognition that frontier models are dual-use technologies that lower the barrier for both cyberattacks and defense. What the sources don't address: Whether the models supervising Astra's training were also subject to the critical cybersecurity risk thresholds imposed on Astra itself.