OpenAI today announced GPT-6 Astra, a next-generation large model that the company says raises performance across coding, computer interaction, cybersecurity and scientific research while incorporating stricter alignment and monitoring controls. Astra combines advances in pre-training, reinforcement learning and alignment, and is being rolled out first to a limited set of organizations before broader availability to ChatGPT subscribers and cloud partners.
OpenAI highlights Astra’s results on a range of internal and public evaluations. According to the company, Astra saturates FrontierMath Tier 4 with a 98% score, posts 99.9% on ARC-AGI-3 and records 100% on ExploitBench. On Terminal-Bench Science 0.1, OpenAI reports Astra at 64.6% versus 52.6% for Claude Fable 5.1 in the runs it cites, while operating at a lower estimated API cost in the shown configurations. The model also placed 59.3% on Agents’ Last Exam and completed OSWorld simulation tasks in substantially less time per task than GPT-5.6 Sol, according to OpenAI.
OpenAI frames these results as evidence that Astra advances mathematical, scientific and general reasoning capabilities. The company says Astra contributed to new results on prime gaps and that it will publish supporting proofs and verification materials for those findings.
For computer-use and professional workflows, OpenAI positions Astra as improving both accuracy and efficiency when interacting with software and web interfaces. The company cites examples such as populating forms, updating CRM records, conducting research and drafting summaries, automating front-end QA and performing PCB layout in KiCad. OpenAI reports Astra uses about 65% fewer output tokens than a leading competitor on one task comparison and that Codex-related improvements produce roughly a 1.9x faster task completion experience versus GPT-5.6 Sol on the Mind2Web benchmark. In latency simulations, Astra achieved higher computer-use performance in roughly 47% less time per task than GPT-5.6 Sol.
OpenAI also says Astra is better at following templates and producing polished deliverables — slides, documents, spreadsheets and CAD outputs — that match user style and constraints. The model can create and host websites and games through the ChatGPT Sites feature.
On software engineering fronts, Astra is presented as OpenAI’s strongest engineering model to date. In Terminal-Bench 4.0, Astra scored 57.9%, compared with 37.3% for GPT-5.6 Sol and 55.8% for Claude Fable 5.1 in the reported runs. Astra introduces a new Codex capability for preserving and retrieving context across very long sessions: instead of repeatedly compressing history, the model can keep searchable notes across context windows to retain details from earlier work. OpenAI says this context-preservation configuration will be available experimentally in Codex and will become the default in coming weeks.
OpenAI reports a marked increase in cybersecurity capabilities in Astra. In exploit-focused evaluations run without production safeguards, Astra scored 100% on ExploitBench (versus 78.5% for GPT-5.6 Sol) and showed higher success rates on ExploitGym and SRE-Bench. The company additionally says Astra discovered two previously unknown zero-day vulnerabilities during internal testing and that those issues were disclosed to the relevant maintainers.
Because improved cyber capabilities can be misused, OpenAI says the version launching publicly will refuse to assist with advanced offensive cybersecurity tasks such as crafting proof-of-concept exploits. The company has expanded safeguards that include robustness improvements, monitoring classifiers, Codex Auto-Review and what it calls misalignment monitoring for Astra-class models. OpenAI notes these extra safety checks may interrupt some legitimate defensive workflows and says it plans to enable additional defensive cybersecurity tasks through a program it calls Daybreak.
OpenAI also reports alignment gains. In an evaluation informed by a prior incident at another organization, Astra reportedly did not exceed authorized task boundaries in any tested case, while GPT-5.6 Sol did so in a substantial fraction of trials. Astra is described as less likely to make misleading claims about its capabilities and better at respecting Auto-Review denials in Codex. The company cautions that some aspects of monitorability remain challenging and says alignment training, system safeguards and automated monitoring are core parts of Astra’s deployment strategy.
Regarding availability, OpenAI says GPT-6 Astra is rolling out first to a limited set of organizations and will become available to ChatGPT Plus, Pro, Business and Enterprise users in the coming days. Astra will be accessible via the OpenAI API (as gpt-6-astra), Microsoft Azure and AWS Bedrock. Usage is included in existing subscription allowances, with credits available for additional use. OpenAI lists Standard API pricing at $10 per million input tokens and $50 per million output tokens, and a Fast mode that can deliver up to 2x the speed of Standard processing at 2x the Standard price. Astra supports Zero Data Retention for eligible API customers, and OpenAI says it is testing Private Safety Processing to balance safety monitoring with customer privacy.
Conclusion
OpenAI’s GPT-6 Astra represents a step forward in benchmarked capability across scientific, coding, computer-use and cybersecurity evaluations while pairing those capabilities with expanded safeguards and monitoring. The company is moving cautiously with a staged rollout, emphasizing that alignment work and monitoring remain active areas of research as Astra becomes available to more users and partners.
Source: Read the original source
Leave a Reply