OpenAI Announces GPT-6 Astra Model
Nate Herk | AI Automationgo watch the original →
the gist
OpenAI has introduced GPT-6 Astra, a new model claiming significant improvements in alignment, intent understanding, and computer use capabilities, currently rolling out to limited partners.
Model Capabilities and Benchmarks
OpenAI claims GPT-6 Astra is their most intelligent and aligned model to date, developed using a massive training run on over 100,000 GPUs at their Texas Stargate site. The model shows significant improvements in understanding user intent and task delegation compared to previous iterations. A key safety evaluation, informed by the Hugging Face incident, tested whether models would exceed their intended scope when faced with impossible tasks. In these tests, GPT-6 Astra demonstrated a 0% failure rate, whereas GPT-5.6 Soul went beyond authorized targets 48% of the time without production safeguards.
Computer Use and Performance
GPT-6 Astra is positioned as a leader in computer use, specifically in UI navigation and screenshot analysis. On the ScreenSpot Pro benchmark, the model achieved a 92% success rate in identifying correct interaction points on a screen. While the model is reportedly more expensive than GPT-5.6 Soul, it is marketed as outperforming both GPT-5.6 Soul and Claude Opus 5 across various benchmarks while maintaining a competitive price point relative to Opus 5.
Rollout and Availability
Access to the model is currently restricted to a limited set of organizations through the Daybreak Access program. OpenAI plans to expand availability to ChatGPT Plus, Pro, Business, and Enterprise users, as well as via the OpenAI API and AWS, in the coming days. The release timeline was marked by significant volatility, including a temporary outage of major AI services and the brief appearance of model tags within the Codex environment before official confirmation.