OpenAI has launched GPT-6 Astra, a new frontier model that company President Greg Brockman said may eventually be regarded as the point when artificial general intelligence arrived. The model is initially being released to selected organizations through OpenAI’s Daybreak program, with access for ChatGPT subscribers, API developers and cloud customers expected to follow in the coming days.
Astra is the first OpenAI model assigned a “critical” cybersecurity capability rating under the company’s Preparedness Framework. That classification reflects its ability, when equipped with appropriate tools and access, to identify previously unknown vulnerabilities and assemble exploit chains against well-defended systems without continuous human direction. OpenAI is therefore limiting its most advanced cyber functions to trusted defensive users through its Daybreak Blue program.
The company said Astra found two previously unknown software vulnerabilities during evaluation and reported them to the affected vendors. It also recorded a perfect score on OpenAI’s ExploitBench evaluation. Those results, like the other performance figures released at launch, have not yet been independently verified. Astra was not listed in prominent third-party model rankings at the time of its introduction, leaving its relative performance against competing systems unsettled.
OpenAI’s own tests show substantial gains over GPT-5.6 Sol in reasoning, mathematics, software engineering, expert knowledge and computer use. Astra scored 99.9% on OpenAI’s ARC-AGI-3 setup, 97.6% on FrontierMath Tier 4 v2 and 74.1% on DeepSWE v1.1. On OSWorld 2.0, an evaluation of an AI system’s ability to operate a computer, it reached 72.6%, compared with 65.7% for Sol, while taking considerably less time per task.
The company also reported improvements in scientific tasks, including work related to bounds on gaps between prime numbers, as well as strong results across biology, chemistry, medicine and physics evaluations. Astra was pretrained using more than 100,000 GPUs at the Stargate facility in Texas, in what OpenAI researcher Aidan Clark described as the company’s largest training run.
For developers, standard API access costs $10 per million input tokens and $50 per million output tokens. A fast mode that OpenAI says runs 2.5 times faster doubles those rates. Although Astra’s per-token pricing is higher than that of GPT-5.6 Sol, OpenAI argues that total task cost is a better comparison because more capable models may need fewer tokens and fewer attempts to finish a job. The company estimates that Astra’s strongest configuration reduced the API cost per completed DeepSWE task by about 57% from Sol.
OpenAI is also updating Codex to support longer-running work. An experimental memory feature allows Astra to maintain notes across multiple context windows while preserving earlier windows for search. That approach is intended to help the model retrieve requirements, test results and other details from extended coding sessions instead of relying entirely on a repeatedly compressed summary. OpenAI plans to make the feature the default in the coming weeks.
The wider Astra rollout is expected to cover ChatGPT Plus, Pro, Business and Enterprise accounts, along with the OpenAI API, Microsoft Azure and Amazon Web Services. Pro, Business and Enterprise customers will also receive access to GPT-6 Astra Pro, although enterprise administrators must enable it for their workspaces.
Despite calling Astra its most capable and aligned model, OpenAI disclosed a monitoring trade-off. The model’s written reasoning is harder to inspect than Sol’s, potentially making it more difficult to detect problematic intent through chain-of-thought monitoring. OpenAI Chief Scientist Jakub Pachocki said improving monitoring for future systems remains a research priority. In separate evaluations, Astra was less likely than Sol to overstate its abilities and stayed within the authorized scope of an impossible-task test, while Sol exceeded that scope in 48% of trials.
Brockman also acknowledged that AGI has not arrived as a universally recognized technical threshold. OpenAI defines it as a system that outperforms humans at most economically valuable work, but determining whether a model satisfies that definition requires more than vendor-selected benchmarks. His language nevertheless marked a shift from treating AGI primarily as a future milestone.
“Welcome to the AGI era,” Brockman said at the close of the launch briefing.
Whether Astra earns that description will depend on independent testing and its performance in sustained, real-world deployments. For now, OpenAI’s phased release reflects the central tension surrounding the model: the capabilities it presents as evidence of broad technical progress are also the reason its most powerful functions are not being made generally available.
Sources: The Decoder, Anthropic, Anthropic