What happened
OpenAI announced GPT-6 Astra, a new model it describes as its most intelligent and aligned. The company says Astra can perform multistep computer tasks, browse the web, write software, analyze scientific data, create websites and work with professional documents, spreadsheets and presentations. OpenAI also says Astra can design printed circuit boards in KiCad, including component placement and copper routing, but the source provides a demonstration rather than independent validation of manufacturability. OpenAI reports strong results on its own and third-party evaluations, including 72.6% on OSWorld 2.0, 57.9% on Terminal-Bench 4.0, 95.9% on BenchCAD and 96.0% on GPQA Diamond. It says Astra completed OSWorld tasks in about 40 minutes on average in its latency simulation, compared with about 75 minutes for GPT-5.6 Sol. The company also reports a 100% score on ExploitBench and says Astra discovered two previously unknown vulnerabilities during testing, which it disclosed to maintainers. The rollout began with a limited set of organizations. OpenAI says Astra will become available over the following days to ChatGPT Plus, Pro, Business and Enterprise users, and through the OpenAI API, Microsoft Azure and Amazon Bedrock. Enterprise administrators must enable Astra, and access is off by default at launch. API standard pricing is listed as $10 per million input tokens and $50 per million output tokens, with separate cache rates. Fast mode costs twice the standard price and is described as delivering up to twice the speed.
OpenAI says GPT-6 Astra combines new work in pre-training, reinforcement learning and alignment. It describes the model as capable of filling forms, updating CRM records, organizing calendars, conducting online research, drafting documents and email, installing and testing software, and performing frontend quality checks. OpenAI also says an updated Codex harness makes computer-use tasks 1.9 times faster than the current GPT-5.6 Sol experience on the Mind2Web benchmark.
The company reports a 98% score on FrontierMath Tier 4 and a 99.9% score on ARC-AGI-3. It says Astra helped establish a new bound for short prime gaps and improve a bound concerning large prime gaps, with proofs and verification materials linked from the announcement. These are claims by OpenAI; the source does not provide independent confirmation of the mathematical results.
OpenAI says Astra meets its Preparedness Framework’s Critical threshold for cybersecurity. In tests without production safeguards, it reports a 100% ExploitBench score, a 42.4% ExploitGym success rate and an 88.0% single-attempt score on SRE-Bench. The company says the launch version will refuse more advanced tasks such as creating proof-of-concept exploits, while a planned Daybreak effort may expand access to defensive validation, malware analysis and detection engineering.
Why it matters
Astra matters because OpenAI is positioning one model as a general-purpose operator across software, office work, scientific tools and cybersecurity, rather than as a chatbot limited to text generation. If the company’s reported results transfer to real deployments, the model could reduce the time required for routine computer-based work and make advanced technical assistance more accessible. The cybersecurity claims also raise the stakes: the same capabilities that help defenders identify weaknesses may help attackers develop exploits. The evidence remains primarily OpenAI’s own testing, so the announcement establishes what OpenAI claims, not an independently confirmed performance record.
The product’s significance is its breadth: Astra is presented as an agent able to interpret instructions, use software, maintain context across long tasks and produce finished work artifacts. That combination could affect knowledge work more directly than improvements limited to response quality.
The source reports that Astra exceeded a human action-efficiency baseline on 96% of ARC-AGI-3 levels and made fewer boundary-crossing errors than GPT-5.6 Sol in OpenAI’s evaluations. Those findings could be relevant to deployments where models act through tools, but they were conducted under stated harnesses and configurations that may differ from what users experience.
The cybersecurity results create a dual-use concern. OpenAI says Astra can identify and develop zero-day exploits in testing, while also claiming improved alignment and stronger monitoring. The practical outcome will depend on safeguards, access controls, incident response and the accuracy of automated intervention systems.
What to watch next
The key questions are whether Astra’s reported benchmark gains hold up in independent testing and ordinary production environments, how often its monitoring systems interrupt legitimate work, and whether its cyber safeguards prevent misuse without blocking defensive research. The source does not identify the organizations receiving the initial rollout, specify regional availability or provide separate Azure and Bedrock pricing. It also does not establish that the PCB demonstration produces fabrication-ready designs without engineering review.
OpenAI says access will expand over the coming days, but the source does not give a precise completion date, list the initial organizations or state whether all advertised channels will launch simultaneously.
The company says Astra usage is included within existing subscription allowances and that additional credits can be purchased. It gives API token prices, but does not document ChatGPT plan prices, allowance changes, cache rates, Azure pricing or Bedrock pricing here.
OpenAI acknowledges that Astra’s written reasoning was harder to monitor in tests designed to elicit evasion. It says production monitoring can pause or stop legitimate work, including defensive cybersecurity. Independent evaluations should examine both failure rates and unnecessary interruptions.
The PCB, scientific-software and professional-work examples are demonstrations and company-selected evaluations. The source does not establish how Astra performs across different hardware, software versions, proprietary data, manufacturing constraints or high-consequence workflows.