The wait is over. OpenAI released GPT-6 Astra on September 3, 2026, and the “Astra” codename that circulated for months turned out to be GPT-6 itself rather than a separate model.
It is available now on Overchat AI, in ChatGPT's paid tiers, and through the OpenAI API.
Below: the confirmed release date and specs, how Astra actually benchmarks against GPT-5.6 Sol and Claude Fable 5.1, what it costs, and whether it's worth switching to for your work.
Yes. OpenAI released GPT-6 on September 3, 2026, under the name GPT-6 Astra. It went to approved users first and reached general availability the next day, September 4.
For most of 2025 and the first half of 2026, this page tracked a moving target: OpenAI had confirmed a 2026 release but never a date, and Sam Altman would only say the gap would be shorter than the 28 months between GPT-4 and GPT-5. That turned out to be right. The model landed in the second half of 2026, roughly in line with when the Abilene hardware came online.
One piece of long-running confusion is also settled. “Astra” had circulated for months as the codename for OpenAI’s next major model, and many assumed it was a separate release sitting somewhere between GPT-5.6 and GPT-6. It was not. Astra is GPT-6.
GPT-6 Astra is OpenAI’s new flagship model and its most capable release to date. It is a closed-weights model built for long-running, complex work — controlling software, running multi-step research, writing and refactoring code, and producing full document sets — rather than for cheap, high-volume chat.
Here are the confirmed specifications:
OpenAI has not published Astra’s architecture, parameter count, expert count, vocabulary size, or exact training compute. The pre-release estimates of 20–50 trillion parameters that circulated widely were never confirmed and remain unverified.
Astra’s headline results come from OpenAI’s own launch evaluations, with independent numbers from Artificial Analysis where available.
Speed improved alongside accuracy. In the OSWorld simulation Astra finished tasks in roughly 40 minutes against 75 minutes for Sol — about 47% faster.
Not across the board, and this is the part the launch numbers can obscure. On the independent Artificial Analysis Intelligence Index, Astra scores 61 — exactly the same as GPT-5.6 Sol, and five points behind Claude Fable 5.1 at 66. The improvements are real but localized to math, tool use, and long-horizon reasoning.
In several evaluations Astra scores below Sol:
The practical read: Astra earns its price on large refactors, complex analysis, agentic programming, tool and application control, very large contexts, and defensive-security work where a mistake is costly. For simple question answering, classification, data extraction, routing, and routine support, it is overkill — GPT-5.6 Terra or Luna will do the same job for far less.
Astra costs about 2.5× more per token than GPT-5.6 Sol:
Two pricing rules matter for budgeting. When input exceeds 272,000 tokens, the whole request is billed at 2× the standard rate for input and cache and 1.5× for output. Batch and Flex modes cost 50% of standard, while API Fast mode costs 2× for up to double the speed.
GPT-6 Astra is available on Overchat AI now, alongside the other frontier models on the platform, so you can compare it directly against Claude and Gemini on your own prompts without managing separate subscriptions.
Astra is rolling out across ChatGPT’s paid tiers — Plus, Pro, Business, and Enterprise. Message limits differ by plan and are subject to change: at the time of writing, Pro $200 includes 200 messages per week, Pro $100 shares a 50-messages-per-week limit with GPT-5.6 Sol Pro, Business Premium includes 50 per week, and Business Standard includes 15 per month.
Astra is available through the OpenAI API, Codex, Microsoft Azure, and AWS Bedrock. For tool-using applications OpenAI recommends the Responses API rather than Chat Completions. Note that Astra does not support fine-tuning, and some sampling controls from earlier models have been removed.
Astra is OpenAI’s first broadly released model carrying a Critical cybersecurity capability rating, which is why it ships with unusually tight access controls and monitoring.
The evaluations behind that rating are striking. Astra scored 100% on ExploitBench even at the lowest verified reasoning effort, reached 42.4% on ExploitGym against 30.3% for Sol, and solved 88.0% of SRE-Bench tasks on the first attempt. In laboratory testing it found and exploited two previously unknown zero-day vulnerabilities created after its knowledge cutoff, and built a browser exploit chain with sandbox escape in 41 hours.
There are limits. In an external Irregular evaluation Astra solved 86 of 226 FrontierCyber tasks against 34 for Sol, but solved none of the seven Elite tasks and did not successfully attack fully hardened targets.
On the defensive side, alignment improved markedly. Astra’s resistance to indirect prompt injection rose from 96.23% to 99.79%, and across more than 54,000 simulated internal Codex tasks it produced roughly half as many serious misalignment flags as Sol.
Astra was trained from scratch. OpenAI says the training corpus contained more than 10 trillion text tokens, over 1 billion images, and more than 1 million hours each of audio and video, drawn from publicly available web data including Common Crawl, partner and private datasets, synthetic data, code, and scientific and technical documents. GPTBot gathered accessible web content from roughly March 2013 through May 2026.
The compute came from the infrastructure that made a model this size possible in the first place.
Stargate is OpenAI’s infrastructure program, backed by OpenAI, SoftBank, Oracle, and Abu Dhabi’s MGX. It is built to supply the power and hardware frontier-scale models need, with nearly 10 GW of planned compute capacity and more than 2 million chips — one of the largest private infrastructure projects ever attempted in the United States.
Abilene is OpenAI’s flagship campus in Texas and the site where much of Astra’s training took place. It hosts over 100,000 NVIDIA chips and roughly 200 MW of active power across its first two buildings, with construction underway to add a further 1.2 GW.
This is the bottleneck Stargate was built to break. OpenAI executives said repeatedly through 2025 that GPU shortages, not ideas, were what kept better models from shipping — the company had stronger systems internally than it could afford to serve. Astra is the first flagship model trained after that constraint eased.
OpenAI has not announced a successor, and Altman has said the team rarely sets targets more than six months ahead. What the Astra release does tell us is that the company has moved to a faster, more iterative cadence — GPT-5, GPT-5.1, GPT-5.6 Sol, and now GPT-6 Astra arrived in far quicker succession than the 28-month GPT-4 to GPT-5 gap. Expect point releases refining Astra before a GPT-7.
The wait is over: GPT-6 arrived on September 3, 2026 as GPT-6 Astra, and Astra turned out to be GPT-6 rather than a separate model. It is a genuine step forward for agentic coding, computer use, advanced mathematics, and multi-step professional work, with a 1,050,000-token context window to match.
But it is a specialized step forward, not a uniform one. Astra ties GPT-5.6 Sol on the independent Artificial Analysis Intelligence Index, trails Claude Fable 5.1, and costs 2.5× more per token than Sol — so whether it is worth switching to depends entirely on what you run on it. High-value, long-horizon work justifies the price; routine chat and extraction do not.
You can try GPT-6 Astra today on Overchat AI and compare it side by side with other frontier models before committing to it.
