OpenAI launches GPT-6 Astra, its first model to cross the 'Critical' cybersecurity threshold; Cognition counters with SWE-2 coding model

OpenAI released GPT-6 Astra on 2026-09-04, its first model rated Critical for cybersecurity risk under its Preparedness Framework, claiming 100% on ExploitBench and discovery of two zero-days while shipping the public version restricted to defensive code…

OpenAI launched GPT-6 Astra on 2026-09-04 as its new flagship model, days after it became the first model to cross the company's Preparedness Framework 'Critical' cybersecurity threshold, a rating that triggers additional deployment restrictions including manual enterprise enablement (per CSO Online). OpenAI claims state-of-the-art computer use, software engineering, and math/science capabilities: 100% on ExploitBench (vs 78.5% for predecessor GPT-5.6 Sol), 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, and 42.4% on ExploitGym (vs 30.3%). On zero-days, sources differ slightly: CSO Online says the model found two previously unknown zero-day vulnerabilities in software released in the three months before launch, while The Hacker News cites demonstrated exploit development on two zero-days disclosed between June and August 2026; The Hacker News adds that the unsafeguarded model can use unknown vulnerabilities for code execution in hardened browsers and OS privilege-escalation exploits. The released version is limited to secure code review and patching, refuses proof-of-concept exploit requests, and its safety checks may interrupt legitimate defensive work pending user review; less restrictive safeguards are planned via OpenAI Daybreak, which includes a $1 billion 'Daybreak for Frontline Defenders' program for critical infrastructure sectors and a pilot with the US MS-ISAC for public sector and water system defenders. The system card reports improved alignment but decreased chain-of-thought monitorability versus Sol, and 0% out-of-scope behavior in a new evaluation (vs 48% for Sol). Pricing is $10 per 1M input tokens and $50 per 1M output tokens, with a $20/$100 fast tier reported only by Latent Space. Rollout starts with limited organizations, then ChatGPT Plus/Pro/Business/Enterprise, the API (model id gpt-6-astra), and AWS/Amazon Bedrock; The Hacker News also lists Azure, which the other reports do not. Independently, Artificial Analysis scored Astra 67 on the Coding Agent Index and 61 on the Intelligence Index, behind Claude Fable 5.1. On 2026-09-10, Cognition launched SWE-2, a proprietary 2.8T-parameter mixture-of-experts model (104B active per token) post-trained from Kimi K3, with vendor-reported scores of 50.0 on FrontierCode 1.1 Main, 73.0 on DeepSWE 1.1, 92.8 on Terminal-Bench 2.1, and 27.3 on Terminal-Bench 4.0. Cognition claims SWE-2 is one point behind Claude Fable 5.1 on FrontierCode at 64% lower cost (one report instead describes it as…

• Launch dates: GPT-6 Astra on 2026-09-04; Cognition SWE-2 on 2026-09-10.
• GPT-6 Astra is the first model to cross OpenAI's Preparedness Framework 'Critical' cybersecurity threshold, triggering additional deployment restrictions such as manual enterprise enablement.
• OpenAI-claimed scores: ExploitBench 100% (vs 78.5% for GPT-5.6 Sol), FrontierMath Tier 4 98%, ARC-AGI-3 99.9%, ExploitGym 42.4% (vs 30.3%).
• Zero-days: CSO Online reports two previously unknown zero-days found in pre-launch testing on software released in the three months before launch; The Hacker News cites exploit development on two zero-days disclosed between June and August…
• Pricing: $10/$50 per 1M input/output tokens; Latent Space additionally reports a $20/$100 fast tier.
• Availability: limited organizations first, then ChatGPT Plus/Pro/Business/Enterprise, the API (gpt-6-astra), and AWS/Amazon Bedrock; The Hacker News also lists Azure, which the other reports omit (source disagreement).
• Released version is limited to secure code review and patching, refuses PoC exploit-generation requests, and safety checks may interrupt legitimate defensive cybersecurity work pending user review; unsafeguarded variant can use unknown…
• System card: improved alignment but decreased chain-of-thought monitorability vs Sol; 0% out-of-scope behavior in a new evaluation vs 48% for Sol; stronger jailbreak robustness, expanded monitoring context, and misalignment containment…