OpenAI has launched GPT-6 Astra, presenting it as its most capable model so far. The release also brings a broader question into focus: when does a steady rise in AI abilities become an “AGI era”? President Greg Brockman says Astra might already qualify under OpenAI’s definition, or be within reach.
A broad capability jump, with benchmarks to match
OpenAI says Astra outperforms its predecessor, GPT-5.6 Sol, across several disciplines. In benchmarks published by the company, it scored 99.9 percent on ARC-AGI-3 for logical reasoning, 97.6 percent on FrontierMath Tier 4 v2, and 74.1 percent on DeepSWE v1.1 for software engineering. The ARC-AGI-3 result was achieved under OpenAI’s own test conditions.
The company also reported scores of 96 percent on GPQA Diamond for expert knowledge, 95.9 percent on BenchCAD, and 100 percent on ExploitBench. These results offer one way to compare capabilities, though the figures come from different tests and measure different tasks.
OpenAI researcher Aidan Clark described Astra’s development as the company’s largest training run ever. The model was pretrained on more than 100,000 GPUs at the Stargate facility in Texas. Clark said the improvement from Sol to Astra was larger than the jump to Sol from earlier models, partly because previous AI models helped monitor training.
Computer use and longer coding sessions
Astra is designed to operate computers, and OpenAI points to OSWorld 2.0 as evidence of progress. Astra scored 72.6 percent on the benchmark, with tasks taking about 40 minutes on average. GPT-5.6 Sol scored 65.7 percent, at roughly 75 minutes per task.
OpenAI is also updating Codex, its coding environment. An experimental feature allows the model to keep notes across multiple context windows during long sessions. Earlier windows remain searchable, letting Astra retrieve requirements or test results that may not have made it into those notes. OpenAI plans to make this feature the default in the coming weeks.
Higher token prices, different costs per task
Through the API, standard mode costs $10 per million input tokens and $50 per million output tokens. Fast mode promises 2.5x speed and doubles the price. Overall, Astra is 2.5x more expensive than GPT-5.6 Sol by token price, putting it in the same price range as Anthropic’s Fable 5.1.
Brockman argues that token prices do not tell the whole story. He says tokens are not directly comparable between competitors, or even across OpenAI’s model families. OpenAI instead points to the cost of completing a task: on DeepSWE v1.1, it estimates Astra’s top configuration costs about 57 percent less per task than Sol.
That distinction matters for customers choosing a model. A higher price for each token can still work out to a lower cost for a particular task, according to OpenAI’s estimate. The result will depend on the use case, so the headline token rates do not by themselves show what a customer’s work will cost.
AGI claims meet cybersecurity risks
Brockman said during the press briefing that there is no clearly defined moment when AI becomes AGI. OpenAI’s team had once expected an obvious threshold that everyone would recognize, he said, but the transition has been more gradual than expected. He closed the briefing with: “Welcome to the AGI era.”
The launch also marks a security milestone for OpenAI. Astra is the first model the company classifies as “critical” under its Preparedness Framework. OpenAI says it can find previously unknown vulnerabilities and build exploit chains across well-defended systems when given suitable tools and access, without step-by-step human guidance. The company also says Astra’s written reasoning is harder to monitor than GPT-5.6 Sol’s.
OpenAI reports that Astra found two previously unknown vulnerabilities during evaluation and that it reported them to the affected vendors. Human experts confirmed the model could identify novel zero-day vulnerabilities in several software categories, including browsers and operating systems. The company says its most advanced cybersecurity capabilities are currently limited to trusted defenders in the Daybreak Blue program, reflecting the dual-use risk: finding a vulnerability can help defenders patch it, but can also help attackers exploit it.
Access is rolling out first to select organizations through OpenAI’s Daybreak program. Broader availability for ChatGPT Plus, Pro, Business, and Enterprise customers is expected in the coming days, alongside access through the API and cloud platforms including AWS Bedrock and Microsoft Azure. Pro, Business, and Enterprise subscribers get GPT-6 Astra Pro, while enterprise workspace admins need to activate it manually.