GPT-6 Astra had barely finished rolling out to paying ChatGPT users before the internet moved on to the next thing. Within 48 hours of Astra's launch, three other labs dropped their own flagship models, and a codename nobody had heard of a month earlier — "Bel" — was suddenly trending as "GPT-7." That's the pace of this industry right now: you don't even get a weekend to sit with a new model before the next rumor cycle starts.
The problem is that most of what's circulating about GPT-7 isn't reporting. It's a mix of leaked X posts, screenshot chains, and SEO pages that convert a single anonymous tweet into a confident-sounding spec sheet. If you want to actually understand where OpenAI is headed next, you need to separate what's confirmed, what's a credible leak, what's pure speculation, and what's already been debunked. Here's the breakdown.
What did OpenAI actually confirm with GPT-6 Astra?
Start with what's real, because it's genuinely wild on its own. OpenAI released GPT-6 Astra on September 3, 2026, and OpenAI president Greg Brockman called it a "generational leap" and said it could eventually be seen as the arrival of artificial general intelligence, or AGI. He didn't hedge much either — "I think it might be about this model," Brockman said in a briefing with reporters about whether Astra could mark the arrival of AGI, before ending the briefing by saying: "Welcome to the AGI era."
The technical details back up the hype, at least on paper. OpenAI said that Astra was built on its largest-ever training run, using more than 100,000 GPUs at its Stargate site in Texas, and this is its first model to use other models in a significant role in supervising Astra's training. On benchmarks, OpenAI says Astra scored 99.9% on ARC-AGI-3 and 97.6% on FrontierMath Tier 4, and Astra tops benchmarks in math, coding, and cybersecurity and is the first model OpenAI rates as "critical" under its safety framework, having independently found two previously unknown zero-day vulnerabilities during testing.
It's not a clean sweep, though. Not every result suggests the age of AGI is here — Astra's 57.2% on Humanity's Last Exam and its Artificial Analysis Intelligence Index result trailed several Claude models. And pricing jumped accordingly: GPT-6 Astra costs $10 per million input tokens and $50 per million output tokens in standard mode, making it 2.5x more expensive than GPT-5.6 Sol and putting it in the same price range as Anthropic's Claude Fable 5.1.
If you want to see the model yourself, the official rollout details are on OpenAI's website and in the ChatGPT release notes, which track exactly which plans get access to Astra and Astra Pro.
What's the "Bel" leak, and is it actually GPT-7?
Here's where things get murky. In late August, an anonymous account posted a "scoop" claiming OpenAI had finished a new pretraining run. Specifically, "Bel" is said to be intended as a post-GPT-6 base model that could potentially underpin an AGI-threshold system after further RL, with the scoop describing it as a giant pretrain with more than 10 trillion total parameters, similar in size to GPT-4.5.
The important nuance almost every headline skips: Bel is not GPT-7. It's a base model, not a shipped product. As one outlet put it plainly, Bel isn't GPT-7 — it's supposedly the raw foundation for building the generation after GPT-6, a base block still needing refinement, not a product ready for ChatGPT. Another report framed it the same way: despite the GPT-6 comparisons, Bel is reportedly not being positioned as GPT-7 itself — instead, the leaked information suggests Bel could serve as a foundational model or pretraining layer for whatever comes after GPT-6.
And the provenance is thin. The primary source is a post published on X by a single account on August 25, 2026, later reshared by other industry accounts and picked up by trade press across multiple outlets — but none of these sources has seen an OpenAI document, and none has a confirmation from the company.
That doesn't mean it's wrong. It means it's a leak, not a fact. Treat it the way you'd treat any unconfirmed industry rumor: interesting, plausible given OpenAI's pace, but not something to build a roadmap around.
Has OpenAI confirmed anything about GPT-7 directly?
No. Flatly, no. One tracker that watches this exact question for a living put it about as clearly as possible: as of the time of writing, OpenAI has not announced, previewed, teased, or named a model called GPT-7. And it's not buried somewhere obscure either — no GPT-7 has appeared in any of OpenAI's official model documentation, the company's Deployment Safety Hub, its SEC-filed prospectus, or any official statements.
If you want a live source for this instead of trusting a snapshot, check OpenAI's official models page directly — it's the one place that updates the moment something actually ships.
Why should you be skeptical of the naming pattern too?
This is the part worth internalizing, because it explains why so much "GPT-7 leak" content ends up wrong within weeks. OpenAI has a documented habit of letting rumor-driven codenames balloon into bigger version numbers than what actually ships.
Look at what happened with "Spud." Back in March, pretraining was completed on March 24, 2026, Sam Altman confirmed it was only "a few weeks" away from release, and Greg Brockman described it as the result of "two years of research" with a "big model feel". The internet ran with that and assumed a major jump — plenty of coverage speculated it would land as GPT-6. Instead, GPT-5.5, also known by its codename "Spud," was released by OpenAI on April 23, 2026. A point release, not the generational leap people had priced in.
The same pattern repeated with the next rumor wave. One tracker summarized it bluntly: Spud turned out to be GPT-5.5; GPT-5.6 shipped 9 July. Even the codenames themselves were contested — one set of sources said ember, beacon, and iris while another said Kindle, Kepler, Levi, and Spud, and if outlets had a real backend source, the codenames would converge — since they don't, at least one set is guesswork dressed up as a leak.
Prediction markets are useful here, but only if you read them correctly. A Polymarket contract priced a launch before June 30, 2026 at roughly 80-89% probability — a genuinely useful number, but a crowd's probability estimate, not insider information. High market odds mean "likely soon." They don't mean confirmed. If you're tracking this stuff, Polymarket's AI model category is worth a bookmark, but treat every contract as a sentiment gauge, not a leaked memo.
What does the Sora shutdown actually tell you about the roadmap?
This is the most underrated signal in the entire GPT-7 conversation, and it has nothing to do with a leaked spec sheet. It's about where OpenAI is choosing to spend its GPUs.
Earlier this year, OpenAI shut down Sora as a standalone product. The official reasoning focused on consolidation, but the resource math is what actually mattered: the company itself acknowledges that computing costs are a determining factor in its decisions, since video generation is one of the most infrastructure-intensive tasks, requiring powerful GPUs, large data centers, and ongoing energy and maintenance costs. Coverage at the time connected the dots even more directly, noting the company was desperately looking to free up computing resources to power its coding and enterprise products, with the strategic logic being that every GPU running a consumer video clip was a GPU not training the next enterprise model.
Here's why that matters for GPT-7 speculation specifically: you don't shut down a video product for GPU reasons and then simultaneously hand a fully native, unified text-image-video "Omni" architecture to your very next flagship model. Those two moves contradict each other. If compute was tight enough to kill a whole product line, native video generation folded into the next frontier model isn't "two weeks away" — it's a much longer bet, if it happens at all. When a leak shows you a detailed Omni spec table for GPT-7, that resource conflict is worth remembering before you take it seriously.
How should you actually rate GPT-7 claims you see online?
Use a simple four-tier filter whenever you see a GPT-7 headline:
- Confirmed — Something on OpenAI's own site, help center, or model documentation. Right now, that list stops at GPT-6 Astra.
- Credible leak — Sourced to a named or trackable account, cross-referenced by multiple outlets, internally consistent. Bel qualifies here, with the caveat that it's a base model rumor, not a product rumor.
- Speculation — Spec tables, pricing guesses, and release dates built by extrapolating from past model pricing trends rather than any actual leak. Most "GPT-7 will cost $X per million tokens" claims fall here.
- Debunk — Claims contradicted by a company's own actions, like a native-video flagship model landing right after the video team got its resources cut.
Apply that filter and most of the GPT-7 hype collapses into categories two through four. That's not a knock on the people writing about it — it's just the nature of covering a lab that trains multiple generations in parallel and rarely confirms anything before launch day.
What should you actually do with this information?
Don't build workflows, budgets, or product roadmaps around a GPT-7 that hasn't been named, let alone shipped. Astra is the model that exists today, and it's worth learning well before chasing whatever comes next — the ChatGPT release notes are the fastest way to see exactly what Astra can do for your plan tier right now.
If you want to track the next generation without getting jerked around by every leak, bookmark OpenAI's models documentation and check it weekly instead of trusting whatever's trending on social media that day. Pair that with a prediction market like Polymarket for sentiment, not certainty.
The pattern is consistent enough now to bet on: the codename hype always runs ahead of the actual release, and the actual release is usually smaller, later, and stranger than the rumor. Bel might be real. It might even be the base for something extraordinary. But "GPT-7" as a named, shipped product with a confirmed spec sheet? That doesn't exist yet — and the smartest move is to stop treating leaks like release notes until it does.