OpenAI launches GPT-6 Astra: inference efficiency breaks the speed-accuracy trade-off and reaches the critical breakthrough threshold

OpenAI announced the launch of its new model, GPT-6 Astra, as the direct successor to GPT-5.6 Sole, describing it as the most intelligent and aligned model in its history and a step toward artificial general intelligence. The model was developed through one of the largest computational training runs to date, using more than 100,000 graphics processing units, delivering core upgrades in reasoning, programming, cybersecurity and agent-based task performance.

Astra offers operational efficiency that enables tasks to be completed with fewer steps and code tokens, alongside a noticeable reduction in real-world errors and hallucinations compared with its predecessor.The most prominent shift in the model’s behavior lies in its ability to continue working and processing data in the background while awaiting the user’s response and clarifications.It surpasses the complete pause that characterized previous generations. Accuracy gains become evident when the model runs in fast inference settings with low latency, allowing developers to execute quickly without sacrificing output quality.

In terms of autonomous computational task performance, OS World 2.0 benchmarks showed Astra completing operations about 47 % faster than the Sole model, with an accuracy of 72.6 % versus 65.7 % for the previous generation. This leap enhances software agents’ ability to manage computing interfaces and execute complex workflows with higher temporal and resource efficiency.

In the digital security domain, the model achieved a perfect 100 % score on the Exploit Bench test and recorded 42.4 % on the Exploit Game test, compared with roughly 30.3 % for the Sole model. These capabilities led the company to label it the first model to reach the “critical capability” threshold in cybersecurity, due to its ability to discover previously unknown vulnerabilities and develop exploit paths without human intervention, prompting the restriction of these advanced features to a limited set of testing partners and the imposition of additional safeguards, especially as the model’s broader control over its idea chain makes it harder to detect during adversarial testing.

For technology firms in the Gulf, Egypt and the Levant, the launch directly impacts the efficiency of building autonomous agents and cloud-computing budgets. Pricing that starts at $10 per million input tokens and $50 per million output tokens, combined with a lower token consumption per task, reduces operational costs for developers of customer-service and financial-analysis platforms. Conversely, as models reach the vulnerability-discovery threshold, cybersecurity teams in the region are compelled to accelerate the integration of automated self-defense agents, since periodic system reviews relying solely on human effort are no longer sufficient to keep pace with attacks driven by advanced reasoning models.

GPT-6 Astra has begun rolling out gradually to a select group of entities, with expectations that it will reach ChatGPT subscribers on the Plus, Pro, Business and Enterprise plans in the coming days, and will also be available through the OpenAI API and the Amazon Web Services cloud platform.