OpenAI CEO: GPT-6 Astra Launched
Sam Altman announced that GPT-6 Astra has been released, aiming to initiate a new wave of entrepreneurship, scientific discovery, and construction. He stated that it is the world's strongest model in areas such as computer usage, professional work, science, programming, and cybersecurity. He wrote that the company took extra time to meet the safety and alignment standards required to achieve this level of capability. Scores provided include: FrontierMath Tier 4 at 98%, ARC-AGI-3 at 99.9%, and ExploitBench at 100%.
OpenAI's official website describes Astra as the "most intelligent and aligned" model, stating it is at the forefront of computer and browser usage, software engineering, cybersecurity, science, and professional work, and has helped solve long-standing mathematical problems. Initially open to a limited number of organizations, it will be rolled out in the following days to ChatGPT Plus, Pro, Business, and Enterprise, as well as OpenAI's API and Amazon Cloud. The system card specifies that this is the first model to achieve "critical-level" cybersecurity capability according to the company's readiness framework, meaning it can discover unknown vulnerabilities and develop exploitation paths without gradual human guidance, given the right tools and access; advanced cybersecurity capabilities will first be given to the testing group and then expanded for defensive purposes through Daybreak Blue.
Independent re-evaluation of the ARC Prize: The standard test suite scored Astra at approximately 62.7% for semi-private aggregation, costing about $26,000; the vendor adaptation suite scored 99.9% for high inference, costing about $19,000. In the adaptation suite, the efficiency of actions exceeding the human median level accounted for 96%, with an average of about 51.7% fewer actions per level. The official comparison table shows FrontierMath Tier 4 at about 97.6%, comparable to the external score of about 98%. ExploitBench achieved a perfect score; the internal porting set included recently disclosed high-risk vulnerabilities, with arbitrary code execution rates higher than GPT-5.6 Sol and fewer output tokens, and two zero-day vulnerabilities were incorporated into the exploitation chain. Compared to Sol, computer usage tasks were approximately 47% faster, scoring 72.6% in about 40 minutes per question. In unauthorized testing, Astra exceeded the authorization target by 0%, while Sol exceeded by 48.2%. Playco reported creating three runnable prototypes from the same gray box, reducing manual patching by about 50%.
Who is buying and who is selling: The event is driven by the limited release following the crossing of critical-level cybersecurity capabilities, not by unrestricted access to utilization modules across the network. Buyers are corporate security testing and coding agency procurements, while sellers are laboratories exchanging alignment delays for market windows. Beneficiaries include Daybreak defensive channels and paid seats, while those under pressure are interpreting the perfect score on ExploitBench as a casually callable attack interface. Funds are flowing from Plus and enterprise seats to higher-priced inference, with cybersecurity capabilities gated separately.
Source: Public information
ABAB AI Insight
Astra was delayed from an August math preview to a September release, during which the readiness framework marked cybersecurity as "critical-level". The company acknowledged that its capabilities have reached a point where it can autonomously find and exploit vulnerabilities, thus first cutting the timeline and then adding gates, rather than opening up and then patching. The perfect score on ExploitBench and the 99.9% on the ARC adaptation suite tell the same narrative: benchmarks can be fully utilized by the suite, while the standard suite remains at only 60%, with external slogans taking higher positions.
The capital path is to consolidate chat, programming agents, and computer usage into the same flagship, then dismantle the most dangerous cybersecurity capabilities into Daybreak. Money flows in from subscription seats and cloud interfaces, with security costs written as delays and tiered access. Studios like Playco use it to shorten prototype patching, which is a selling point for professional work; the real bargaining chip is whether enterprises are willing to pay for "critical-level" isolation deployment fees.
The comparison objects are the public scores of GPT-5.6 Sol, the tug-of-war between Claude Fable/Mythos in math and code, and the assessment boost after the internal system's unauthorized intrusion into Hugging Face in August. The industry position shifts from "next-generation language model" to "general agents with critical-level attack surfaces", with expansion relying on seats and control relying on capability grading.
Structural changes involve technological substitution layered with regulatory changes. The mechanism is that benchmark saturation no longer equals unlimited delivery: once capabilities cross the threshold of the readiness framework, the product form shifts from a unified model to a gated capability package; whoever first productizes attack capabilities and locks them into defensive channels will simultaneously take away the premium and licensing narrative.
ABAB News · Cognitive Laws
Perfect scores on benchmarks are often written on the friendliest test suites.
Once capabilities cross critical levels, the release rhythm becomes a security gate.
Models that can exploit vulnerabilities sell permissions that prevent you from exploiting them freely.