GPT-6 Astra
OpenAI's flagship model, released on 3 September 2026, and the first the company says meets its Critical cybersecurity risk threshold.
Astra was announced in August 2026 as OpenAI’s next model family and shipped on 3 September as GPT-6 Astra, four weeks later than planned after the company said it had slowed the release over cyber risk. It takes a context window of roughly 1.1 million tokens, produces up to 128,000 tokens per reply, and costs $10 per million input tokens and $50 per million output. OpenAI reports 72.6 percent on OSWorld 2.0, the benchmark for driving a computer desktop, ahead of GPT-5.6 Sol and Claude Opus 5.
The detail that got the most attention was not a benchmark. OpenAI says Astra is the first model to reach the Critical tier of its own cyber risk scale, meaning it can find unknown software flaws and write working exploits without step-by-step human guidance. It ships with alignment training plus production safeguards including automated code review and misalignment monitoring.
-
An AI Reads and Rewrites Another AI's Weights, With Mixed Evidence
-
Gemini 4 Carbon: Google's Strongest Model Goes to Cyber Defenders First
-
50 AI Agents Played Minecraft Together and Invented Bureaucracy
-
AI Agents Face a Real Tax Exam and Pass About One Assignment in Nine
-
OpenAI's Decisions API Answers Quick Yes/No Questions for $0.10 per Million Tokens
-
Google Cloud Now Sells OpenAI Models Next to Gemini and Claude
-
OpenAI's Coding Agents Switched Off Their Own Safety Monitoring, and It Can't Say Why
-
Mercury 3 Reviewed: A Chatbot That Writes Its Answer in Chunks, Not Word by Word
-
OpenAI Says GPT-6 Has Got Much Better at COBOL and Other Old Programming Languages
-
ChatGPT Pro Can Now Run Several Agents on One Task at the Same Time
-
What Happens When an AI Assistant Cannot Get Your Refund?
-
Nvidia Opens Nemotron 3.5 Super VL and a Local Coding Model That Fits in 24 GB
-
ChatGPT Now Answers With Buttons, Charts and Mini Apps
-
A Model 28 Times Cheaper Found Three Quarters of the Bugs, and Missed Half the Security Ones
-
Sakana's New Model Is Not a Model, It's a Dispatcher for Other Models
-
Two Benchmark Labs Measured GPT-6 Astra and Reached Opposite Conclusions
-
OpenAI Ships GPT-6 Astra, and Calls It Its First Critical Cyber Model
-
OpenAI Is Cutting Cursor Off From Its Models After the SpaceX Takeover
-
Watching Its Own Model Now Costs OpenAI 20 Percent Extra, and It Paused a Training Run to Do It
-
OpenAI Froze Its Own Training for Two Weeks, and Now Watches Its Models Like a Hawk
-
Anthropic Nudged Its Own Risk Rating Up and Says It Is Keeping a Model in the Drawer
-
GLM-5.3 Got Much Better at Breaking Software, So Z.ai Is Holding the Weights Back
-
A Fields Medalist Who Wrote a Paper on AI Extinction Risk Just Joined OpenAI to Work on Safety
-
OpenAI Says Its Next Model Might Hit the Highest Cyber Risk Level, and Paused Part of the Work
-
OpenAI named its next model family Astra and introduced it with ten solved math problems