GPT-6 Astra: OpenAI's New Benchmark Ceiling in Computer Use, Cybersecurity, and Science
GPT-6 Astra arrives with scores that push past what most frontier models have posted on reasoning, autonomous computer operation, and exploit discovery, while OpenAI keeps the rollout deliberately narrow and the cybersecurity guardrails deliberately tight.
GPT-6 Astra arrives with scores that push past what most frontier models have posted on reasoning, autonomous computer operation, and exploit discovery, while OpenAI keeps the rollout deliberately narrow and the cybersecurity guardrails deliberately tight.
A model that treats computer use as a solved-enough problem
OpenAI is positioning GPT-6 Astra as state-of-the-art across computer use, browsing, software engineering, cybersecurity, science, and professional work. The numbers back that framing in specific ways rather than vague superlatives. On OSWorld 2.0, a benchmark that simulates real desktop and browser tasks end to end, GPT-6 Astra scored 72.6%, with latency simulations showing roughly 40 minutes per task, a figure that matters because it tells you the model is being measured on sustained multi-step operation rather than single-shot answers. On Mind2Web, a separate web-task benchmark, GPT-6 Astra completed tasks 1.9 times faster than the current GPT-5.6 Sol experience, a direct speed comparison against OpenAI's own prior deployed model rather than an abstract claim. Terminal-Bench 4.0, which tests command-line and systems tasks, produced a 57.9% score, and DeepSWE v1.1, focused on software engineering, landed at 74.1%. None of these numbers say the model works unsupervised in production; they say the trajectory on autonomous task completion has moved, and by how much.
Reasoning and science scores that raise the ceiling, not the floor
On FrontierMath Tier 4 v2, a benchmark built specifically to be resistant to memorization and pattern-matching on hard mathematics, GPT-6 Astra scored 97.6%. ARC-AGI-3, designed to test novel abstract reasoning rather than trained-in patterns, came in at 99.9%. GPQA Diamond, a graduate-level science question set, produced a 96.0% score. These are ceiling-testing benchmarks, meaning they are constructed to be hard for exactly the kind of shortcut behavior that large models sometimes fall back on, so a high score is meaningful evidence about reasoning depth. That said, HealthBench Professional, which scores health-related answers with length-adjusted scoring to penalize models for padding responses, produced a 63.4% result for GPT-6 Astra. That gap between near-ceiling scores on abstract math and reasoning versus a mid-60s score on professional health questions is worth sitting with, because it tells you the model's strengths are not evenly distributed across domains, and professional health use cases sit closer to the model's current limits than its abstract reasoning does.
The cybersecurity numbers come with a matching set of restrictions
GPT-6 Astra scored 100.0% on ExploitBench and 100.0% on Impossible ExploitGym, but only 42.4% on the standard ExploitGym success rate, a split that suggests the model handles known, well-structured exploit patterns far more reliably than genuinely open-ended exploit development. On SRE-Bench, which measures site reliability and operations tasks, it solved 88.0% of tasks in a single attempt and 99.2% within four attempts. During evaluation, GPT-6 Astra discovered two previously unknown zero-day vulnerabilities, a result serious enough that OpenAI classified the model as meeting the Critical threshold under its Preparedness Framework. That classification is precisely why the launch version refuses advanced tasks such as creating proof-of-concept exploits, and why it scored 0.00% on OpenAI's internal circumvention benchmark, meaning testers could not get the model to bypass those refusals during evaluation. The capability and the restriction are two halves of the same fact: a model good enough to find real vulnerabilities is a model OpenAI is choosing to constrain at launch.
Who can actually reach it, and what it costs
Access is not open by default. GPT-6 Astra is initially rolling out to a limited set of organizations, with a planned expansion to ChatGPT Plus, Pro, Business, and Enterprise users. Outside of ChatGPT, it is reachable through the OpenAI API under the identifier gpt-6-astra, as well as through Microsoft Azure and Amazon Bedrock, giving teams already committed to either cloud a path that does not require a direct OpenAI relationship. Standard API pricing runs $10 per million input tokens and $50 per million output tokens, with a Fast mode available at up to 2x Standard processing speed for teams that need lower latency on interactive tasks. Eligible API customers can enable Zero Data Retention, which OpenAI supports as an option for that eligible group specifically, not as a default behavior applied to everyone. For teams evaluating whether to build against it now versus waiting for broader ChatGPT availability, the practical decision point is whether current API access or cloud-platform access through Azure or Bedrock already covers their organization, since the limited initial rollout means direct ChatGPT access may not arrive on the same timeline.
Specs
- API model identifier: gpt-6-astra
- OSWorld 2.0 score: 72.6%
- FrontierMath Tier 4 v2 score: 97.6%
- ARC-AGI-3 score: 99.9%
- ExploitBench score: 100.0%
- ExploitGym success rate: 42.4%
- SRE-Bench one-attempt score: 88.0%
- Standard input-token price: $10 per million tokens
- Standard output-token price: $50 per million tokens
- Fast mode speed: Up to 2x Standard processing speed
Sources
- OpenAI GPT-6 Astra announcement — OpenAI