OpenAI Releases GPT-6 Astra: A 1.05M-Context Computer-Use Model Gated Behind a ‘Critical’ Cyber Threshold
MarkTechPost Asif Razzaq ● Covered by 4 sources
OpenAI shipped GPT-6 Astra, a hosted model that works inside apps, not just in chat. It’s gated for trusted users because OpenAI says its cyber skills hit a “Critical” threshold.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI has launched GPT-6 Astra, and the company is making a very specific pitch: this is a computer-use system first, a chat model second. The idea is that it can work across browsers, spreadsheets, desktop apps and terminals the way a person would, then finish the job instead of explaining how to do it. That’s the headline. The catch is equally clear. Astra is closed, hosted and unreleased as weights, so there is no self-hosting route for anyone who wanted to run it on their own hardware.
Access is tight for now. The model is live only for organizations in OpenAI’s Trusted Access and Daybreak programs, and the broader rollout is not immediate. OpenAI says the model’s newer memory approach matters most for long agent runs: Codex used to compact earlier turns into summaries when context filled up, which could erase the exact details a later step needs. Astra instead keeps notes across context windows and searches back through earlier messages and tool output. The feature starts as experimental behind a config.toml setting, then becomes the Codex default in the coming weeks.
The model page also gives a sense of scale. OpenAI lists a 1,050,000-token context window, 128,000 max output tokens and an April 30, 2026 knowledge cutoff. Input is text and image; output is text only. reasoning.effort adds two levels above high, called xhigh and max. Tool support includes computer use, hosted shell, apply patch, skills, MCP and tool search. Fine-tuning is not supported.
On benchmarks, Astra looks strongest when the task resembles operating software. OpenAI reports 72.6% on OSWorld V2-Offline, up from 65.7% for GPT-5.6 Sol, while average task time drops from roughly 75 minutes to 40. The coding picture is much less dramatic. Astra scores 74.1% on DeepSWE v1.1, just ahead of Sol’s 72.7%, and the source itself notes those gaps amount to only a couple of tasks on a 113-task benchmark.
The part that changes the access story is cyber. OpenAI says Astra is the first model it has designated as reaching the Critical cybersecurity threshold in its Preparedness Framework. In testing, it developed exploits for hardened browsers and operating systems, and it found two previously unknown V8 vulnerabilities that OpenAI says it is disclosing to maintainers. Standard access blocks advanced cybersecurity work such as exploit discovery, and for API developers a safety check can stop a task outright instead of pausing for approval. OpenAI also says users outside trusted-access programs may run into slowdowns, pauses or blocks, even on unrelated work.
My take — AI-written commentary, not fact-checked reporting
This is the familiar OpenAI move: build something genuinely useful, then wrap it in a velvet rope and a safety speech. The closed model part matters as much as the benchmark number, because the industry keeps calling this progress while making sure nobody else can inspect, host or really stress-test it. Very efficient, very modern, and just a little too tidy.
Read more about this at: MarkTechPost
