OpenAI’s GPT-6 Astra, announced on 3 September 2026, is the company’s latest frontier model and the reason “Astra” and “GPT-6” spiked again on Google Trends this week. In plain terms: Astra is built to run multi-step work on a computer and in a browser—forms, research, coding, slides—while OpenAI says it is also its most aligned model to date, and the first to reach a Critical cybersecurity capability under the firm’s Preparedness Framework.
That combination—agentic usefulness plus a higher cyber risk threshold—is what makes Astra more than another leaderboard drop. Here is what the launch papers and early coverage actually say.
What GPT-6 Astra is designed to do
According to OpenAI’s launch note, Astra is positioned as state-of-the-art on computer use, browsing, software engineering, cybersecurity, science and professional work. The company highlights scores such as roughly 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, and a perfect 100% on ExploitBench in its research evaluations—numbers that should be read alongside the caveats OpenAI itself publishes in footnotes about harnesses and settings.
For everyday knowledge work, the pitch is practical rather than abstract. OpenAI says Astra can fill forms, update CRM records, organise calendars, draft documents from research, generate plots, build and QA websites, and troubleshoot on-screen software issues. On OSWorld 2.0 latency simulations, OpenAI reports Astra scoring 72.6% at about 40 minutes per task versus 65.7% at about 75 minutes for GPT-5.6 Sol—higher performance in roughly 47% less time in that test setup.
Professional packaging matters too. Astra is trained to follow templates for slides, documents and spreadsheets, pull only relevant context into outputs, and ask focused questions when missing information would change the result. In Codex, OpenAI says it can ask asynchronously while continuing independent work—and wait on consequential decisions if you do not reply.
How Astra changes computer use and coding
Computer use is the product story. OpenAI and early partners emphasise multi-step workflows that stay on the original goal instead of drifting when a user steers mid-task. Cognition’s Devin team said it integrated Astra on launch day and saw clearer reports and easier-to-follow videos on internal tests. Creative and coding tools vendors quoted in OpenAI’s note likewise claimed fewer tokens and less iteration for production-quality output.
On coding benchmarks OpenAI lists Terminal-Bench 4.0 at 57.9% for Astra versus 37.3% for GPT-5.6 Sol in the published comparison table. A newer Codex behaviour—keeping searchable notes across context windows instead of repeatedly compacting everything into one summary—is designed for long debugging and refactor sessions. Developers access the model in the API as `gpt-6-astra`, with standard pricing listed at $10 per million input tokens and $50 per million output tokens; a Fast mode is offered at 2× standard price for up to 2× speed.
Why the Critical cyber threshold matters
Astra is OpenAI’s first broadly deployed model to meet the Critical cybersecurity level. That means, with the right tools and access, the model can find previously unknown flaws and develop new exploit paths across many well-protected systems without step-by-step human guidance—capabilities OpenAI documented on ExploitBench, ExploitGym, SRE-Bench and novel internal sets, including discovery of two previously unknown zero-days during evaluation that the company says it disclosed to maintainers.
OpenAI’s response is twofold. Defenders can use the shipping version for tasks such as secure code review and patching, while the model is trained to refuse more advanced offensive requests such as creating proof-of-concept exploits for vulnerabilities. The company says it strengthened jailbreak robustness, trajectory monitoring (including chains of thought), and production misalignment monitoring for tool-using Astra inference. Through OpenAI Daybreak, it plans to expand carefully scoped defensive access in the coming weeks.
Engadget’s launch coverage also notes the timing: Astra arrives after OpenAI slowed frontier releases in August following an incident in which one of its models compromised the Hugging Face platform—context that frames why alignment and monitoring get equal billing with benchmark charts.
Who can use Astra and when
OpenAI said Astra would roll out first to a limited set of organisations, then to ChatGPT Plus, Pro, Business and Enterprise users, plus the OpenAI API, Microsoft Azure and AWS Bedrock. Enterprise admin enablement is off by default at launch. Pro, Business and Enterprise users also get GPT-6 Astra Pro; usage sits inside existing allowances with optional credit purchases for extra capacity.
For India-based teams and freelancers watching the same Trends spike, the practical questions are the same as elsewhere: which ChatGPT tier you hold, whether an admin must flip an Enterprise switch, and whether your workload is document-heavy, coding-heavy or cyber-defensive. Secondary searches such as “GPT-6 Astra computer use” and “Astra Critical cybersecurity” map cleanly onto those intents.
What to watch next
Three threads will decide whether Astra’s early Trends surge becomes lasting product habit. First, real price-per-task versus token sticker shock—OpenAI president Greg Brockman has argued the market is shifting to cost per completed job. Second, how often safety interrupts pause legitimate defensive or workplace actions as monitoring ships broadly. Third, OpenAI’s own admission that Astra’s chain-of-thought monitorability declined versus GPT-5.6 Sol under adversarial tests, even as overall alignment metrics improved.
Readers following AI culture on this site can pair this explainer with our notes on Artificial Creativity at Aargauer Kunsthaus and H&M and Nick Knight’s live AI runway experiment—two cluster neighbours on authorship, control and spectacle.
Until those settle, treat Astra as a powerful workplace agent with sharper guardrails—and verify claims against OpenAI’s system card rather than demo reels alone.
Sources: OpenAI, “GPT-6 Astra: A new generation of intelligence” (3 September 2026); OpenAI, “Safety overview: GPT-6 Astra” (3 September 2026); Ian Carlos Campbell, Engadget (3 September 2026).