OpenAI's GPT-6 Astra Brings Computer Control—and Cyber Risk
OpenAI’s new flagship model can operate software and posts major benchmark gains, but its critical cyber capabilities and reduced monitorability complicate the company’s AGI-era pitch.
Path to Astra: critical capabilities and frontier safeguards
OpenAI's GPT-6 Astra on ARC-AGI-3
'Welcome to the AGI era': OpenAI launches GPT-6 Astra
OpenAI launches GPT-6 Astra, its most powerful model yet, and touts its ability to use your computer
OpenAI’s GPT-6 Astra: Benchmarks, Cyber Risks, and Market Impact
From chatbot to computer operator
OpenAI has begun releasing its newest flagship model to selected organizations, with broader access planned over the coming days for paying ChatGPT customers, developers and users of Microsoft Azure and Amazon Bedrock.venturebeat The company is pitching the system less as a better answer engine than as an operator that can navigate browsers, spreadsheets and desktop applications to complete multistep work.venturebeat +1
OpenAI says the model scored 72.6% on an offline subset of OSWorld 2.0 while taking about 40 minutes per task, versus 65.7% and roughly 75 minutes for its GPT-5.6 Sol predecessor.venturebeat Standard API pricing is $10 per million input tokens and $50 per million output tokens, putting the launch at the expensive end of the commercial model market.futurumgroup
Headline benchmarks need context
The most dramatic reasoning result comes from ARC-AGI-3, a test built around unfamiliar interactive environments. ARC Prize reported a 62.7% score using its provider-neutral standard harness, but 99.9% with an OpenAI-specific adapter that preserves opaque reasoning state and manages longer conversations.arcprize Both were state of the art, yet the gap shows how strongly memory and surrounding software can shape a model’s apparent intelligence.
ARC Prize said the system used fewer actions than the median human participant on 96% of levels under the adapted setup. The benchmark’s creators nevertheless said saturation is not proof of artificial general intelligence because its tasks remain bounded, deterministic and closed-ended.arcprize That caution contrasts with OpenAI president Greg Brockman’s launch-day declaration that it is reasonable to view the industry as entering an “AGI era.”venturebeat +1
Cyber power raises the stakes
The release also marks the first time OpenAI has classified one of its models at the “Critical” cybersecurity level. The company says that, with suitable tools and access, it can discover unknown flaws and build exploits against hardened systems without step-by-step human direction; internal testing included a perfect ExploitBench score and the discovery of two previously unknown vulnerabilities.openai
OpenAI says it delayed parts of development and release while adding stricter isolation, monitoring and misuse protections. General users will initially be blocked from advanced offensive tasks, while vetted defenders are expected to receive wider capabilities through the Daybreak program.openai +1 The unresolved issue is observability: OpenAI’s own material says the model’s reasoning is harder to monitor than its predecessor’s, even as its alignment scores improved.futurumgroup For businesses, the launch therefore pairs a potentially useful computer operator with a governance burden that grows in step with its autonomy.