OpenAI released GPT-6 Astra

OpenAI released GPT-6 Astra on 03/09/2026, introducing a new frontier AI model focused on computer use, complex reasoning, software engineering, cybersecurity, scientific work and multi-step professional tasks. OpenAI described Astra as its most capable and aligned model at launch. The model was initially made available to a limited group of organizations, with access scheduled to expand to ChatGPT Plus, Pro, Business and Enterprise users over the following days, alongside availability through the OpenAI API, Microsoft Azure and Amazon Bedrock. Astra combines reasoning with direct computer interaction, allowing it to carry out workflows involving browsers and professional software rather than only generating instructions for a user. OpenAI demonstrated the model working on tasks involving spreadsheets, Power BI, game development, electrical engineering, web development, legal documents and other computer-based workflows. On OSWorld 2.0, OpenAI reported a score of 72.6% at approximately 40 minutes per task, compared with 65.7% at approximately 75 minutes for GPT-5.6 Sol. The company said this represented about a 47% reduction in task time in its latency simulations. GPT-6 Astra also introduced a context window of 1,050,000 tokens and a maximum output of 128,000 tokens. OpenAI's API documentation lists a knowledge cutoff of 30/04/2026. In OpenAI's long-context evaluation, Astra scored 96.3% on an eight-needle MRCR v2 test covering the 512,000-to-one-million-token range, compared with 73.8% for GPT-5.6 Sol. The model posted high scores on several reasoning and technical evaluations reported by OpenAI. GPT-6 Astra achieved 99.9% on ARC-AGI-3, 98% on FrontierMath Tier 4 and 100% on ExploitBench. These figures differ slightly from some early third-party reports that listed a 97.6% FrontierMath Tier 4 result. OpenAI's official launch material reports the score as 98%. Cybersecurity capability was one of the most closely controlled aspects of the release. OpenAI classified GPT-6 Astra as its first model to reach the "Critical" cybersecurity capability threshold under the company's Preparedness Framework. OpenAI said evaluations showed that, with appropriate tools and access, Astra could discover previously unknown vulnerabilities and develop exploitation methods against well-protected systems without requiring a person to direct every step. During testing, the model discovered and used two previously unknown vulnerabilities as part of an exploit chain, which OpenAI said it was disclosing to the relevant maintainers. Because of those capabilities, the company introduced additional safeguards around Astra's deployment. OpenAI reported strengthened monitoring, internal isolation, checkpoint encryption, alignment evaluations and protections intended to reduce harmful cyber use. The company also said Astra demonstrated greater resistance to jailbreaks and prompt-injection attacks than GPT-5.6 Sol. Results involving Astra's highest cybersecurity capabilities reflected controlled Daybreak Blue access rather than the model's standard production configuration. For developers, OpenAI priced standard GPT-6 Astra API usage at $10 per million input tokens and $50 per million output tokens, with cached input priced at $1 per million tokens. Prompts exceeding 272,000 input tokens are subject to higher rates. OpenAI also stated that Astra supports Zero Data Retention for eligible API customers. GPT-6 Astra Pro was announced for Pro, Business and Enterprise plans as part of the broader rollout. Why This Moment Matters: The release combined several developments that had previously been treated as separate AI capabilities: advanced reasoning, long-context processing and direct interaction with computer environments. It also created a new deployment challenge for OpenAI because Astra became the company's first publicly deployed model classified at the Critical cybersecurity capability level, requiring stronger safeguards alongside its broader computer-use abilities.