OpenAI launches GPT-6 Astra focused on computer use and Critical cybersecurity capability
On Sep 4 OpenAI released flagship model GPT-6 Astra at $10/M input and $50/M output. Astra is the first model to cross the Preparedness Framework Critical cybersecurity threshold, paired with a $1B Daybreak defenders program.
On September 4 OpenAI announced on its official site and via President Greg Brockman's social channels the formal release of its new flagship base model GPT-6 Astra, calling it the world's most intelligent and most aligned model. The API model name is gpt-6-astra, with standard pricing of $10 per million input tokens and $50 per million output tokens, approximately 2.5x the previous GPT-5.6 Sol and aligned with Anthropic's Claude Fable 5.1 pricing. OpenAI opened access immediately to a small set of Trusted Access and Daybreak member organizations, with Plus, Pro, Business and Enterprise subscribers and API users scheduled to receive access over the following days. AWS Bedrock is also providing access.
Astra's central change is extending the model boundary from generating content to executing complete workflows. The official page cites the original framing: anything you can do on a computer, Astra can do for you, fast. The model plans steps autonomously, invokes a browser, reads and writes local files, writes and runs code, operates software interfaces and self-debugges errors in a closed loop. Official demos include KiCad PCB design, Excel and PowerBI data analysis and one-shot generation of runnable games and web pages.
On training scale, OpenAI research VP Aidan Clark said Astra was the first product pretrained on more than 100,000 GPUs at the Texas Stargate site, with the data-center network, inference kernel and model architecture all designed from scratch for that scale. Astra is also the first OpenAI model in which previous-generation models played a deep role in training the next one. Benchmark scores include 97.6% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, 100% on ExploitBench, and 0% override rate on impossible-task alignment tests, versus 48.2% for GPT-5.6 Sol.
Given those capabilities, Astra is the first OpenAI model formally classified as Critical under the Preparedness Framework cybersecurity capability threshold, defined as the ability to discover and exploit unknown software vulnerabilities with limited human supervision. Internal testing reportedly exploited two zero-days. Alongside the launch OpenAI committed $1 billion in subsidized access, training and support via the Daybreak for Frontline Defenders program, initially targeting US water systems, power grids, local governments and community banks. On the evening of September September 4 Sam Altman publicly apologized for the staged rollout and promised compensation in the form of a one-time banked reset credit for every paid subscriber who hit their limit.