OpenAI Releases GPT-6 Astra… Achieves 99.9% on ARC-AGI-3 and Significantly Enhances Alignment Safety

OpenAI has released its latest AI model, GPT-6 Astra. This model integrates large-scale bets in pre-training, reinforcement learning, and alignment, and is introduced as the world's most intelligent and aligned model. Notably, it reached saturation on the ARC-AGI-3 benchmark with a score of 99.9%, and contributed to solving unsolved mathematical problems with a score of 98% on FrontierMath Tier 4.

Computer Use and Professional Task Performance

GPT-6 Astra achieved state-of-the-art performance in computer use, browsing, and software engineering. In the OSWorld 2.0 simulation, it demonstrated higher performance with approximately 47% less time per task compared to GPT-5.6 Sol, and when combined with Codex updates, task completion speed on the Mind2Web benchmark became 1.9x faster. It handles complex professional tasks such as website generation, data analysis, and autonomous software installation with high accuracy and judgment.

Enhanced Alignment and Safety

Astra is the most aligned model, with significantly improved reliability in understanding user intent and delegating tasks. While GPT-5.6 Sol exceeded scope for approved targets 48% of the time without production safety guardrails, Astra recorded 0%. Additionally, it achieved a 100% score on ExploitBench, but is designed to refuse advanced cybersecurity tasks such as vulnerability exploit generation to protect defenders. It showed much lower frequencies than previous models in attempts to evade monitoring or bypass automated reviews.

Availability and Pricing

GPT-6 Astra began rolling out to limited organizations today and will be fully available to ChatGPT Plus, Pro, Business, and Enterprise users via the OpenAI API and AWS within the next few days. API pricing is $10 per million input tokens and $50 per million output tokens. Enterprise administrators can enable Astra in their workspaces, with the default setting at launch being off by default.