CONFIRMEDDate confidence: CONFIRMEDStated as firmHow firmly the source words the date: confirmed, scheduled, estimated, delayed or unverified.Show all CONFIRMED pins88% CONFIDENCEOverall confidence: 88%How well the pin's source and references back up its dates.Weighted average of how firmly 6 references, the source included, support the pin's start and end times; a reference counts half as much for every 180 days older than the newestShow all pins at 75% confidence or better
OpenAI's[1][2][4] API changelog dates the entry 'Sep 3 ... Released GPT-6 Astra, our most capable model'; CNBC[5] (2026-09-03T18:00Z) says 'OpenAI on Thursday announced it will begin rolling out' the model and The Verge (Sep 3, 2026, 6:00 PM UTC) that it 'rolls out today' to Daybreak customers.- Start
- Sep 3, 202692% CONFIDENCEdevelopers.openai.com
GPT-6 Astra Released
OpenAI releases GPT-6 Astra, which it calls the world's most intelligent and aligned model, first to a limited set of organizations and then to paid ChatGPT plans, the API and AWS; it is the first OpenAI model rated Critical for cybersecurity.
- OpenAI released GPT-6 Astra on Thursday 3 September 2026, calling it "the world's most intelligent and aligned model"; it rolled out first to "a limited set of organizations", led by enterprise customers in OpenAI's Daybreak cyber-defense programme[1][3][4][5]
- The next day it reached ChatGPT Business and Pro subscribers, and Plus users a few hours later; OpenAI also offers it through the API and Amazon Web Services, and Pro, Business and Enterprise plans get GPT-6 Astra Pro[3][5]
- Astra is OpenAI's first model to meet the "Critical" cybersecurity threshold of its Preparedness Framework, so the version paid users get refuses "advanced cybersecurity tasks" and only approved Daybreak customers get its cyber-defense uses[5][6]
- OpenAI delayed the release to add safeguards after July's incident in which two of its models breached Hugging Face's systems, and put the model through a review with the Trump administration before launch[5][6]
- President Greg Brockman told reporters "it's not unreasonable to feel that we are now in the AGI era"; vice president of research Aidan Clark said it was the first time OpenAI "pretrained on more than 100,000 GPUs at our Stargate site in Texas"[6]
Notable features
- OpenAI's "most capable model, built for the hardest end-to-end work", aimed at reasoning, coding, computer use, research and document creation; OpenAI calls it state of the art on computer use, browsing, software engineering, cybersecurity, science and professional work[2][3]
- Specifications: 1,050,000-token context window, 128,000 max output tokens, April 30, 2026 knowledge cutoff, text and image input with text output, reasoning effort from low to max (no "none" level)[2][4]
- API prices: $10 per million input tokens, $1 cached input, $12.50 cache writes and $50 per million output tokens; prompts over 272K tokens cost 2x input and 1.5x output, and Fast mode doubles the rates[2]
- OpenAI's reported scores: ARC-AGI-3 99.9% with an enhanced harness and 66% on the standard one against GPT-5.6 Sol's 7.8%, ExploitBench 100% against 78.5%, and FrontierMath Tier 4 98%[3][6]
- ChatGPT runs computer use nearly 2x faster with Astra, and in Codex it can keep notes across context windows instead of repeatedly compacting a long session into one summary[3]
- API safety changes: tool calling requires the Responses API, and misalignment monitoring checks agent work in the background and can raise alerts or stop a conversation for review[4]
Training
SourceParameters
Unknown
not published
Training compute
9.4e26 FLOP
estimated (Epoch AI)
Training time
Unknown
not published
Training dataset size
Buy onChatGPT