OpenAI is moving quickly beyond GPT-5.6 Sol, introducing GPT-6 Astra just two months after its previous flagship model arrived. The new model targets a familiar list of AI battlegrounds — coding, scientific reasoning, professional tasks and autonomous computer use — but cybersecurity is emerging as one of Astra’s most consequential capabilities.
The headline numbers are substantial, although benchmark results still need to be treated as measurements rather than proof of general intelligence. Astra scored 98.6% on ARC-AGI-3, according to figures provided alongside the launch, versus 7.8% for GPT-5.6 Sol and 30.2% for Claude Opus 5 running in a high-compute configuration. On FrontierMath Tier 4 v2, Astra reached 97.6%, compared with 83% for Sol, 87.8% for Claude Fable 5.1 and 73.2% for Claude Opus 5.
The improvements are less dramatic on some established tests. Astra recorded 96% on GPQA Diamond, only narrowly ahead of Gemini 3.8 Flash at 95.3% and GPT-5.6 Sol at 94.6%. DeepSWE v1.1 produced another close contest, with Astra scoring 74.1% against Gemini 3.8 Flash’s 73.7% and Sol’s 70.8%.
Where things become more interesting is agentic work. Astra scored 64.6% on Terminal-Bench Science 0.1, compared with 22.4% for Sol, while its AutomationBench result climbed to 41.4% from Sol’s 18.1%. Those gains matter because the industry is increasingly trying to move AI beyond answering prompts and towards models capable of completing longer, multi-stage tasks with less human intervention.
That capability also creates a security problem. OpenAI has classified Astra at the “Critical” cybersecurity capability threshold under its Preparedness Framework, reportedly the first OpenAI model to receive that designation. The company says it has consequently strengthened refusal behaviour for harmful cyber requests and introduced monitoring designed to interrupt potentially unauthorised activity.
Access reflects that caution. Astra is initially being provided to enterprise cybersecurity customers through OpenAI’s Daybreak programme, while its strongest cyber capabilities will remain limited to selected trusted defenders. ChatGPT Plus, Pro, Business and Enterprise subscribers are expected to receive broader Astra access in the coming days.
It will also cost considerably more than Sol. API pricing is set at $10 per million input tokens and $50 per million output tokens, putting Astra alongside Claude Fable 5.1 pricing rather than continuing Sol’s cheaper positioning.
OpenAI President Greg Brockman has gone further, suggesting Astra could represent the beginning of an “AGI era.” That is a much bigger claim than any benchmark table can establish. What Astra does demonstrate, based on the announced results, is how rapidly frontier models are becoming better at autonomous technical work — and why capability gains are increasingly arriving alongside restrictions on who gets to use the most powerful versions.
