A flagship model built around work
OpenAI has launched GPT-6 Astra across ChatGPT Work, Codex and its developer API. The company describes Astra as its most capable model for professional work, emphasizing computer use, browsing, software engineering, cybersecurity, science and document-heavy business tasks.
The launch moves the center of competition beyond chat quality. Astra is designed to work through applications people already use, including interfaces without an API, so the practical question is whether it can complete a whole task reliably rather than produce one strong answer.
Computer use becomes a core capability
OpenAI says Astra can write code and operate applications inside ChatGPT Work and Codex. That opens workflows in spreadsheets, presentations, internal tools and browser-based systems, but it also expands the consequences of a mistaken action.
The company is pairing the model with confirmation policies, automated review and new enterprise controls for approved websites, desktop applications, uploads, downloads and browsing history. Organizations should begin with narrow permissions and require approval for external communication, deletion, access changes and other consequential steps.
Pricing and availability
API pricing begins at $10 per million input tokens and $50 per million output tokens. OpenAI argues that Astra completes tasks with fewer tokens and retries, but buyers should evaluate total cost per successful task rather than comparing token prices alone.
Enterprise administrators can enable Astra under their applicable agreement and rate card. Access is off by default at launch, which gives teams an opportunity to establish evaluation datasets, permission boundaries and rollback procedures before broad deployment.
Early benchmarks need careful interpretation
OpenAI reports a 57.9% result on Terminal-Bench 4.0, compared with 37.3% for GPT-5.6 Sol and 55.8% for Claude Fable 5.1. It also cites customer evaluations in document analysis, coding and long-running agent workflows.
Those figures are useful signals, not universal guarantees. Harness design, tool permissions, retry policies and task selection can materially change results. Teams should reproduce representative workflows using held-out data and score accuracy, completion rate, intervention time, latency and cost together.
A higher cybersecurity capability tier
OpenAI says Astra is the first of its models to reach the Critical cybersecurity capability threshold under its Preparedness Framework. The company says it has strengthened boundary-respecting training, jailbreak resistance and automated checks intended to block harmful responses and unauthorized actions.
That designation makes deployment governance part of the product story. Security teams should treat Astra as a privileged operator when it receives tools or credentials, isolate environments, minimize secrets, retain audit trails and test failure behavior before granting production access.
What teams should test first
The best initial evaluation is a bounded workflow with a clear finish line: reconcile a spreadsheet, investigate a known code defect or assemble a cited briefing from approved sources. Review not only the final artifact but also the model's intermediate decisions and tool calls.
Readers can follow the AINewsInu homepage, our AI Models hub and AI Coding coverage for verified updates on Astra availability, independent evaluations and the safeguards required for computer-using systems.
Sources & further reading
Social-media activity is treated as a signal of attention, not proof. Product claims are attributed to the linked publisher or announcement.