OpenAI’s New Model Aces the Benchmarks and Admits It Is Better at Hiding
OpenAI said the model led many coding and science tests and reached a critical cybersecurity threshold under its Preparedness Framework.
- On Thursday, September 3, 2026, OpenAI unveiled GPT-6 Astra, described as its most capable and aligned model, with benchmark results showing significant performance gains over GPT-5.6 Sol across coding, scientific reasoning, and professional tasks.
- OpenAI delayed the release earlier this year after Astra reached the 'Critical' cybersecurity capability threshold under its Preparedness Framework, prompting strengthened safeguards after the model identified two previously unknown zero-day vulnerabilities on ExploitBench.
- Data confirms a 99.9% score on ARC-AGI-3 and 97.6% on FrontierMath Tier 4, with Greg Kamradt of the ARC Prize Foundation saying Astra "surpassed our human action-efficiency baseline on 96% of levels, effectively reaching human parity on the benchmark."
- Access expands within days to ChatGPT Plus, Pro, Business, and Enterprise users, while the system is available through the OpenAI API, AWS Bedrock, and a separate Astra Pro variant for paid tiers.
- Company officials acknowledged these are internally reported results and internal reasoning is harder to monitor than in Sol; OpenAI added a 20% compute overhead for safety, though Astra does not outperform Claude Fable 5.1 on every evaluation.
31 Articles
31 Articles
OpenAI presents Astra, a model that navigates, programs and manages tasks without constant supervision while setting new benchmarks records.
OpenAI will sell you Astra, but not the system that scored 98.6% on ARC-AGI-3
OpenAI's GPT-6 Astra, released on Thursday, scored higher and cost less with a different harness. It's the latest example of why the software around an AI model deserves as much attention as the model.
OpenAI’s new Astra model can code better. But here’s why its cybersecurity skills matter as much
OpenAI describes Astra as a model built for end-to-end work, rather than only answering prompts. But its release comes at a time when cybersecurity and AI models have made headlines.
OpenAI’s new model aces the benchmarks and admits it is better at hiding
OpenAI has published benchmarks for GPT-6 Astra showing large jumps over GPT-5.6 Sol, a 100% score on an exploit benchmark and two zero-days found in testing, alongside an admission that the model’s reasoning is harder to monitor when it attempts evasion.
Coverage Details
Bias Distribution
- 57% of the sources lean Left
Factuality
To view factuality data please Upgrade to Premium












