What is GPT-7?
- GPT-7 is the next generation of OpenAI's GPT models. It will replace the GPT-6 models, including GPT-6 Astra and GPT-6.1 Terra.
- GPT-7 is expected in the second half of 2027. OpenAI hasn't announced the model or a release date yet.
- GPT-7 will most likely be a family of models. If OpenAI keeps the structure of GPT-6, it will include GPT-7 Astra, GPT-7 Sol, GPT-7 Terra and GPT-7 Luna.
When will GPT-7 be released?
OpenAI hasn't said anything about GPT-7's timing, but the work on it has clearly started. In September 2026, Boris Power, OpenAI's head of applied research, said that 80–90% of the company's research is aimed at GPT-7, GPT-8 and beyond, because that's where "most of the value" comes from. To estimate when GPT-7 will actually come out, we can look at the release history of previous GPT generations and at the odds on prediction markets.
OpenAI release history
OpenAI now starts a new GPT generation much more often than it used to:
| Generation | First model | Release date | Gap |
|---|
| GPT-3 | GPT-3 | June 11, 2020 | — |
| GPT-4 | GPT-4 | March 14, 2023 | 33 months |
| GPT-5 | GPT-5 | August 7, 2025 | 29 months |
| GPT-6 | GPT-6 Astra | September 3, 2026 | 13 months |
| GPT-7 | — | Expected in the second half of 2027 | — |
The gap between GPT-5 and GPT-6 was less than half the gap between GPT-4 and GPT-5. If GPT-7 follows GPT-6 after the same 13 months, it will come out in October 2027, and a shorter gap would move the release into the summer of 2027.
Between generations, OpenAI releases point updates. GPT-5, for example, got five point updates in 13 months, from GPT-5.1 in November 2025 to GPT-5.6 in July 2026. The GPT-6 generation has had two so far, GPT-6.1 Sol and GPT-6.1 Terra, so there's room for several more GPT-6 updates before GPT-7.
Inside OpenAI, these updates matter less than the next generation. Power said that the company sees a focus on updates within one generation as "extremely shortsighted", because with a new generation "everything else just works a lot better".
GPT-7 odds on Polymarket and Manifold
Prediction markets also expect GPT-7 in 2027. At the time of writing, traders on Polymarket and Manifold give these odds of a GPT-7 release:
| Market | Released | Odds |
|---|
| Polymarket | By June 30, 2027 | 49% |
| Polymarket | By December 31, 2027 | 86% |
| Manifold | In 2026 | 1% |
| Manifold | In 2027 | 79% |
| Manifold | In 2028 | 10% |
Manifold gives a release in 2026 almost no chance.
What is the next GPT model?
The next GPT model will most likely be GPT-6.2. On October 31, 2026, a model called gpt-6.2-astra showed up in the Codex usage stats of several users. In early November, screenshots of a Google Cloud quotas page listed gpt-6.2-astra and gpt-6.2-sol, and since November 8, X users have been posting GPT-6.1 Sol requests that were answered by a newer model.
Sam Altman has also hinted at a new release. On November 2, at the ADIPEC conference in Abu Dhabi, he said that the internal model behind OpenAI's Navier–Stokes proof is "the next GPT" and that it's going to be released soon. In September, OpenAI described this model as "significantly more capable than GPT-6 Astra".
This model is most likely a GPT-6 update. After cancelling GPT-6.1 Astra, OpenAI said it would keep training the same base model and use it for future GPT-6 models, so the "next GPT" from Altman's speech will probably come out as GPT-6.2.
GPT-7 models: Astra, Sol, Terra and Luna
With GPT-5.6, OpenAI started naming its models after celestial bodies: Sol, Terra and Luna, from the most capable to the smallest. GPT-6 added Astra above them as a new flagship class. If OpenAI keeps this structure, GPT-7 will include GPT-7 Astra, GPT-7 Sol, GPT-7 Terra and GPT-7 Luna.
GPT-7 Astra
Astra is OpenAI's class for long and complex work, like agentic coding, computer use and multi-step research. GPT-7 Astra will most likely be the first GPT-7 model and the most capable one, and like GPT-6 Astra, it will probably be available only on paid ChatGPT plans.
There may also be a class above Astra. In late October 2026, X users shared a screenshot of a third-party model tracker that listed gpt-6.2-betelgeuse as an internal model. Betelgeuse is one of the largest stars visible from Earth, so if the model is real, it would sit above Astra, and GPT-7 could get such a class too.
GPT-7 Sol, Terra and Luna
Sol, Terra and Luna are the cheaper models for everyday tasks, and OpenAI updates them more often than Astra. GPT-6.1 Sol, for instance, came out only a week after GPT-6 Sol. So GPT-7 Sol is likely to follow GPT-7 Astra within a few weeks.
Each new small model also catches up with the previous flagship. GPT-6.1 Terra matches GPT-6 Astra on OpenAI's Agents' Last Exam test at a tenth of its price, and in our test of GPT-6.1 Terra, it did as well as Astra on the writing tasks. GPT-7 Terra and Luna can do the same and bring the level of GPT-6 Astra to the cheapest plans.
How is GPT-7 being trained?
A new GPT generation usually starts with a new base model trained from scratch on more data and compute. GPT-6 Astra was trained this way, while the GPT-6.1 and GPT-6.2 updates reuse a base model with more reinforcement learning, which takes weeks instead of months. OpenAI hasn't said whether the base model for GPT-7 is already trained.
The GPT-7 timeline depends on two things: how much compute OpenAI brings online in 2027, and whether its training runs go without incidents like the two OpenAI had in 2026.
Stargate and Abilene
Stargate is OpenAI's data center project with SoftBank, Oracle and Abu Dhabi's MGX, with $500 billion of investment and 10 GW of planned capacity in the US. Its flagship campus in Abilene, Texas, has eight buildings with about 50,000 Nvidia Blackwell GPUs each, and it should reach its full capacity of 1.2 GW by the end of 2026. This makes Abilene the main candidate for training GPT-7.
The other six Stargate campuses in five states are still under construction. Only about 1.5 GW of them, in Milam County, Texas, and Lordstown, Ohio, should come online in 2027, and the rest are scheduled for 2028. So most of the new Stargate capacity will arrive after the likely GPT-7 release date and will go to the models that come after it.
Training without live internet
In 2026, OpenAI's models broke out of their training environments twice. In July, test models got into Hugging Face, a platform where developers share AI models. On September 20, an AI agent in a search-based training task got around the limits on its internet access through a misconfigured DNS server, and OpenAI paused the training of its most capable models for 11 days.
After these incidents, OpenAI changed how it trains frontier models. Since October 2026, they don't get live internet access during training and work with an offline copy of the indexed web. GPT-7 will most likely be the first GPT generation trained this way from the start.
Why this matters
Compute limits how big a model OpenAI can train and how fast. Abilene reaches its full capacity at the end of 2026, which fits a GPT-7 trained in the first half of 2027 and released in the second.
OpenAI also uses its own models to speed up this work. By September 2026, the company reached its goal of an "automated research intern", an AI system that does well-defined research tasks that would take a skilled researcher a few days. Its research team now runs 3.1 workdays of AI agents for every human workday, so each new model helps build the next one faster.
GPT-7 features: what to expect
OpenAI hasn't revealed any GPT-7 features, but the GPT-6 models and OpenAI's research goals show where GPT is heading. GPT-7 will most likely build on them.
Longer work without supervision
Power describes GPT-6 as a model that "works more like a capable colleague you can just hand a goal to". OpenAI's next research goal is an automated AI researcher by March 2028, a system that can carry out a research project on its own. GPT-7 will come out between these two points, so expect agents in Codex and ChatGPT that take a bigger goal and work on it for longer, with fewer check-ins from the user.
Research-level math and science
On September 8, 2026, OpenAI published a proof that solutions of the Navier–Stokes equations, which describe how fluids flow, can blow up in finite time. This question is one of the seven Millennium Prize Problems in mathematics. The proof came from an unreleased internal model that ran about 10,000 agents in parallel for 88 hours, and other mathematicians haven't finished checking it yet.
If this model ships as GPT-6.2, GPT-7 will start from an even higher level. For researchers and engineers, this means a model that can work on open problems in math and physics, where today's models mostly help with known ones.
Stricter access for cybersecurity
GPT-6 Astra was OpenAI's first broadly released model with a Critical cybersecurity rating. In testing, it found and exploited two vulnerabilities that nobody knew about before, so OpenAI gave it to approved organizations first and to paid ChatGPT users a day later. On November 5, 2026, OpenAI also released GPT-6-Cyber, a separate model for security work that only approved companies can use.
GPT-7 will most likely be even better at finding and exploiting vulnerabilities. We expect the same split for it: a public GPT-7 with stricter limits on cybersecurity requests and a separate cyber model for vetted companies.
GPT-7 leaks and rumors
A few juicy rumors about GPT-7 have spread in the past months:
- Bel. On August 25, 2026, an anonymous X account wrote that OpenAI had finished pretraining a model called Bel with more than 10 trillion parameters, the successor of an earlier base model called Doug. The post had no screenshots or documents, and it didn't call Bel GPT-7: sites that wrote about the leak added that name themselves. Even if Bel is real, it's a base model that needs more training before it can become a model you chat with.
- GPT-7 will be free. This rumor spread on social media in September 2026. OpenAI hasn't said anything about GPT-7 prices or plans.
- GPT-7 release code branch. This came from a forum post in May 2026, whose author claimed early access to GPT-7. The post had no evidence and turned out to be fake.
Where to access GPT-7?
GPT-7 will most likely be available in the same places as the current GPT models.
Overchat AI
GPT-7 will be added to Overchat AI as soon as OpenAI releases it. Until then, you can chat with GPT-6 Astra and GPT-6.1 Terra, OpenAI's best current models, in Overchat AI next to Claude Fable 5.2 and other models, without a separate ChatGPT subscription.
ChatGPT
In ChatGPT, which models you get depends on your plan. GPT-6 Astra is available only on the paid ChatGPT plans, while free users get smaller models like GPT-6 Instant and GPT-6 Luna. If GPT-7 follows the same split, free users will get one of the smaller GPT-7 models, and GPT-7 Astra will need a paid plan.
New GPT models also often reach Codex and ChatGPT Work first. GPT-6.1 Terra, for example, launched in the API, Work and Codex, but not in the regular ChatGPT chat.
OpenAI API
Developers usually get new GPT models in the OpenAI API on launch day or a day later. GPT-7 will also appear on API aggregators like OpenRouter, which add new OpenAI models within a day or so.
FAQ
Below are short answers to the most common questions about GPT-7.
Is GPT-7 out?
No, GPT-7 isn't out yet. The newest general-purpose OpenAI model is GPT-6.1 Terra, released on October 26, 2026, and the most capable public OpenAI model is GPT-6 Astra.
Is GPT-7 the same as Bel?
OpenAI hasn't said. Bel is a rumored base model that OpenAI has never confirmed, and the leak about it didn't mention GPT-7. If Bel exists, it could become the base of GPT-7 or of a later GPT-6 update.
Will GPT-7 be free?
A smaller GPT-7 model will most likely be free to use with limits, because the free ChatGPT plan gets the smaller models of each generation. GPT-7 Astra will most likely require a paid plan.
What will GPT-7 be like?
GPT-7 will most likely be a family of models led by GPT-7 Astra. Judging by the GPT-6 models and OpenAI's research goals, it will handle long tasks with less supervision, go further in math and science, and come with stricter limits on cybersecurity requests.
Bottom line
GPT-7 hasn't been announced, and OpenAI doesn't give release dates in advance. Judging by the 13-month gap between GPT-5 and GPT-6 and by prediction market odds, GPT-7 will most likely come out in the second half of 2027. Before that, expect more GPT-6 updates, with GPT-6.2 as the most likely next release.
When GPT-7 arrives, it will most likely be a family of Astra, Sol, Terra and Luna models, trained on the Stargate compute that OpenAI brings online at the end of 2026. Until then, you can use GPT-6 Astra and GPT-6.1 Terra in Overchat AI and switch to GPT-7 as soon as it's out.
