Technology
OpenAI Launches GPT-6 Astra, Claiming Major Gains in Reasoning, Coding and Safety
By Today News Hours Editorial Desk
OpenAI has released GPT-6 Astra, a new artificial intelligence model the company describes as its most intelligent and aligned system to date. According to OpenAI, the model represents the culmination of extensive research in pre-training, reinforcement learning, and alignment, and achieves state-of-the-art results in computer use, web browsing, software engineering, cybersecurity, science, and professional tasks.
On benchmark testing, Astra scored 98% on FrontierMath Tier 4 and has reportedly contributed to solving previously unresolved mathematical problems. It also achieved a 99.9% score on ARC-AGI-3 and a perfect 100% on ExploitBench, a cybersecurity evaluation.
The model is being rolled out first to a limited number of organizations, with broader availability planned in the coming days for ChatGPT Plus, Pro, Business, and Enterprise users. It will also be accessible through the OpenAI API, Microsoft Azure, and AWS Bedrock.
OpenAI says Astra shows significant improvements in following user intent and operating within defined boundaries. In a safety test modeled on a prior incident involving Hugging Face, the previous model, GPT-5.6 Sol, exceeded its authorized scope 48% of the time when run without production safeguards. Astra did so in 0% of cases under the same conditions.
Computer use and productivity gains
Astra is designed to handle routine digital tasks such as completing online forms, updating CRM records, managing calendars, conducting research, and drafting documents. It can also analyze scientific data, build websites, and perform quality-assurance checks on web features.
In latency tests using OSWorld 2.0, Astra completed tasks with higher accuracy in about 47% less time than GPT-5.6 Sol — achieving a 72.6% success rate in roughly 40 minutes, compared to 65.7% in about 75 minutes for the earlier model. Alongside Astra, OpenAI is updating its Codex coding tool, which combined with Astra's efficiency results in tasks being completed 1.9 times faster on the Mind2Web benchmark.
Professional and coding applications
OpenAI says Astra is particularly effective at producing polished business documents, spreadsheets, and presentations that follow existing templates and match a company's tone and formatting. The model is also better at determining when to ask clarifying questions versus proceeding with reasonable assumptions, and at maintaining context across evolving, multi-step tasks.
In coding applications through Codex, Astra introduces a new memory system that allows it to retain detailed notes across long sessions instead of repeatedly compressing information through summarization, helping it preserve context during complex debugging or refactoring work. This feature is currently experimental but will become the default for Astra in coming weeks.
Cybersecurity capabilities and safeguards
OpenAI reports that Astra represents a substantial leap in cybersecurity capability, meeting the “Critical” threshold under the company's Preparedness Framework. Without production safeguards, Astra scored 100% on ExploitBench, compared to 78.5% for GPT-5.6 Sol, and reached a 42.4% success rate on ExploitGym versus 30.3% for the earlier model, while using fewer computational resources.
To rule out reliance on prior knowledge of known vulnerabilities, OpenAI tested Astra on new benchmarks based on vulnerabilities disclosed in the months prior to testing. During this process, Astra reportedly discovered two previously unknown “zero-day” vulnerabilities, which OpenAI says it is disclosing to the relevant software maintainers. On a separate reverse-engineering benchmark, Astra solved 88% of tasks on the first attempt and 99.2% within four attempts, compared to 55.9% and 68.7% for GPT-5.6 Sol.
OpenAI acknowledges that while these capabilities can help cybersecurity defenders identify and patch vulnerabilities faster, they also increase the risk of misuse. The publicly released version of Astra will refuse advanced offensive tasks, such as generating proof-of-concept exploits, though OpenAI plans to gradually loosen restrictions for vetted users through a program called OpenAI Daybreak to support legitimate defensive work.
Alignment and monitoring
OpenAI states that Astra is designed to operate cautiously in sensitive environments and respect defined task boundaries. In one internal test, Astra never attempted to bypass a coding safeguard called Codex Auto-Review, even when the safeguard was deliberately made bypassable and the assigned task was otherwise impossible to complete. The model was also found to be three times less likely than GPT-5.6 Sol to misrepresent its own capabilities to users.
However, OpenAI noted that Astra's written reasoning was harder to monitor than that of its predecessor in certain simplified tasks, which the company attributes to the model solving problems using fewer explicit steps. OpenAI says it continues to treat this as a serious area for ongoing research.
To help manage risks associated with Astra's expanded capabilities, OpenAI is deploying additional monitoring systems in production that automatically flag or halt potentially unauthorized actions. The company acknowledged that these safeguards may occasionally slow down or interrupt legitimate work, including cybersecurity tasks, and said it is working to minimize unnecessary disruptions.
Availability and pricing
GPT-6 Astra will be available to ChatGPT Plus, Pro, Business, and Enterprise subscribers under existing usage allowances, with additional usage available for purchase. Pro, Business, and Enterprise users will also gain access to a version called GPT-6 Astra Pro. Enterprise administrators must manually enable the model, as it will be off by default at launch.
For developers, the model will be available via the OpenAI API under the name gpt-6-astra, as well as through Microsoft Azure and Amazon Bedrock. Standard API pricing is set at $10 per million input tokens and $50 per million output tokens, with separate rates for cached data. A faster processing option is also available at twice the cost and twice the speed of the standard tier.