OpenAI Releases GPT-6 Astra: 1.05M Context AI Model
OpenAI launches GPT-6 Astra, featuring a 1.05-million token context window, computer-use capabilities, and restricted access.
Stock photo for illustration only, not from the actual event
- OpenAI introduces closed-hosted GPT-6 Astra with a 1.05M-token context window supporting text and images.
- Astra achieves 72.6% on OSWorld V2-Offline and 99.9% on ARC-AGI-3, with 74.1% in coding benchmarks.
- It is OpenAI's first model to cross the Critical cybersecurity threshold in testing.
- Pricing is set at $10 per million input tokens and $50 per million output tokens under strict access rules.
OpenAI has officially unveiled its latest closed-hosted artificial intelligence model, GPT-6 Astra, which comes without released weights, meaning self-hosting is currently not an option. The model is rolling out exclusively for organizations enrolled in OpenAI’s Trusted Access and Daybreak programs.
The most significant shift for developers lies in context handling. While the previous Codex model relied on compaction—summarizing earlier conversation turns and discarding details an agent might need later—Astra retains notes across context windows and searches back into earlier messages and tool outputs. This experimental feature is configurable via config.toml and will become the default for Codex in the coming weeks.
Stock photo for illustration only, not from the actual event
Astra also introduces the ability to ask users a question while continuing background work that does not depend on the answer, removing a common bottleneck where unresolved decisions stall entire jobs. The official model page lists a 1,050,000-token context window, a 128,000 max output token limit, and an April 30, 2026 knowledge cutoff, alongside two new reasoning effort levels above high: xhigh and max.
Regarding performance metrics, OpenAI reports that Astra scores 72.6% on OSWorld V2-Offline compared to 65.7% for GPT-5.6 Sol, reducing average task completion time from roughly 75 minutes down to 40 minutes. On the ARC-AGI-3 benchmark, it records a 99.9% score using a Responses API harness. In coding tasks, Astra logs 74.1% on DeepSWE v1.1, keeping pace with current industry standards.
"Astra is the first model OpenAI has designated as reaching the Critical cybersecurity threshold in its Preparedness Framework."
OpenAI
The designation of GPT-6 Astra reaching the Critical cybersecurity threshold marks a major milestone in AI safety governance. As autonomous agent models gain sophisticated computer-use capabilities and the ability to discover novel vulnerabilities, strict deployment gates and real-time safety checkpoints become essential tools for mitigating catastrophic risks.
Astra stands out as the first model designated by OpenAI as reaching the Critical cybersecurity threshold under its Preparedness Framework. During testing, the model successfully developed exploits for hardened browsers and operating systems, while also uncovering two previously unknown V8 vulnerabilities that OpenAI is disclosing to maintainers.
Priced at $10 per million input tokens and $50 per million output tokens—with cached inputs at $1.00—Astra applies strict safety interventions. Standard access refuses advanced cybersecurity work, and API safety checks halt non-compliant tasks outright rather than pausing for approval.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment