GPT-6 Astra - OpenAI's most capable model for end-to-end work
by•
GPT-6 Astra is OpenAI's most capable model for complex reasoning, software engineering, computer use, science, and professional work. Async tool calling and mid-turn steering help it run multistep agent workflows. API: model id gpt-6-astra, $10/$50 per 1M tokens (short context). Rolling out first via Trusted Access/Daybreak, then Plus, Pro, Business, Enterprise, and the API in the coming days. First OpenAI model at the Critical cybersecurity capability threshold (advanced cyber access limited).
Replies
I’ve been using ChatGPT since the beginning, and recently I’ve been using the Codex tool with different models for my coding work. Honestly, it’s one of the best coding tools I’ve used so far. That said, I still really like Claude.
Airuncode
can u fix the arch version of codex please we dont have the desktop oficial app
Dial
the mid-turn steering part is what I'm most curious about. for tool calls you can undo, steering mid-turn is just a nicer UX. but once a step in the workflow is something external and already committed (a payment call, an email sent, a phone call already ringing), can you still steer the model away from a plan it already started executing on, or does steering only work for the parts that haven't left the sandbox yet
Dial
the "Critical cybersecurity capability threshold" line is the part I keep coming back to. gating advanced cyber access sounds responsible on paper, but it also means whoever gets into that access tier ends up with a real offense/defense edge over everyone still on the standard release. genuinely curious how the eligibility bar is set for that tier and whether there's any independent check on it, or if it's OpenAI's own call end to end.