GPT-7: Release date, confirmed details, and specs

This page tracks OpenAI's next major model, and will update as the facts come in and convert to a full launch breakdown the day GPT-7 is officially announced. 

Is GPT-7 out yet?

No. As of the time of writing, OpenAI has not announced, previewed, teased, or named a model called GPT-7. The newest OpenAI model is GPT-6 Astra, released to approved organizations on September 3, 2026 and to paying ChatGPT users the following day. No GPT-7 has appeared in any of OpenAI’s official model documentation, the company’s Deployment Safety Hub, its SEC-filled prospectus, or any official statements.

Currently, any claims about GPT-7 are linked to confusion about GPT-6’s Astra launch, the unconfirmed “Bel” pretraining leak (See the "What is Bel?" section), or something similar.

What is OpenAI's latest model right now?

GPT-6 Astra, launched September 3, 2026. It is OpenAI's current flagship and the direct successor to GPT-5.6 Sol. 

GPT-7 release date: what's known and what's estimated

OpenAI has given no date for when GPT-7 will be released. However, we can estimate generally when it will happen based on past model release trends. Currently, the cadence points to GPT-7 being released in the second half of 2027.

Here’s the previous release data used for that estimate. For whole-number releases, OpenAI has been bringing these out in shorter and shorter timeframes:

Model	Release	Gap from last release
GPT-3	June 2020	—
GPT-4	March 14, 2023	~33 months
GPT-5	August 7, 2025	~29 months
GPT-6 Astra	September 3, 2026	~13 months

It’s unlikely that OpenAI will use a whole number for a new model release before at least six months have passed, which means the soonest you’d see something named GPT-7 is Q2 2027 (and even that’s fairly optimistic). And with the current pace of competition, they won’t want to skip an entire year and look like they’re behind, hence the H2 2027 estimate.

That said, you’re going to wind up with a lot of point-releases before then — things like ‘GPT-6.1’ and ‘GPT-6.2’. To date, OpenAI has been doing this roughly every six to ten weeks. There’s good reasoning behind this: development is incremental, it keeps interest levels up, and it helps project a sense of competition compared to offerings like Anthropic’s Claude.

Rumors and leaks: What's circulating, and what's verified.

Every item below is dated and rated. Confirmed means OpenAI said it. Reported means a named outlet with sources. Unverified means social media or anonymous.

Date	Claim	Origin	Status
September 2026	"GPT-7 will be free!"	Reddit threads resharing a post from Sam Altman on X from December 24, 2023.	While Sam Altman did post in 2023 that "Someday you'll get $0 GPT-7," it was in the context of a paid GPT-8 already being out. There are no indications for Sam Altman or OpenAI that GPT-7 will be free on release.
August 2026	OpenAI finished pretraining "Bel," a 10T+ parameter successor to the "Doug" base model, intended for post-GPT-6 models.	X account @synthwavedd; amplified by Wccftech, 36Kr, others	Unverified. Single anonymous source. No OpenAI document seen by any outlet. No OpenAI confirmation.
August 2026	Bel "could arrive by end of 2026"	X reposts of the above	Unverified. Even the original leak frames Bel as a base, not a shippable product.
August 2026	Bel can "work continuously for days" and "coordinate hundreds of subagents"	36Kr and derivative coverage	Unverified. These are Astra's marketed capabilities restated with bigger numbers.
May 2026	Forum post claiming early access to a "GPT-7 release code branch" with screenshots	Forum post claiming early access to a "GPT-7 release code branch" with screenshots	Debunked. No verifiable repository hashes; flagged by moderators. Similar hoaxes accompany every OpenAI launch.
Ongoing	"GPT-7 is out"	Social media confusion	False. GPT-6 Astra is out.

GPT-7 specs at a glance

Nothing about GPT-7 has been confirmed. The table below shwos the current flagship as the baseline that any GPT-7 model should be measured against, barring any recent point releases that supersede GPT-6 Astra. 

Spec	GPT-6 Astra (Confirmed)	GPT-
Release date	Sept 3, 2026	Not announced
API model ID	gpt-6-astra	Not announced
Context window	~1.05M tokens	Not announced
Input	Text, image	Not announced
Output	Text	Not annoucned
Knowledge cutoff	April 2026	Not announced
Base model (Reported)	"Doug"	Possibly "Bel" (unverified)
Preparedness level	Critical (Cyber)	Not assessed
Training compute	100,000+ GPUs, Stargate Texas	Not announced

Will the next OpenAI model be called GPT-7?

While nothing is certain until official release, this is extremely likely. However, OpenAI is now using two naming labels at once: the version number (GPT-5.6, GPT-6) and a set of codenames for model classes (Astra, Sol, Luna, etc.) This is pretty common in the industry, given you’ve got things like Anthropic’s Claude (Opus, Haiku, Sonnet, Mistral) and Google’s Gemini (Flash, Pro) also using codenames.

GPT-7 pricing

GPT-7’s pricing has not been announced yet. The best indicator of how much it will cost is what OpenAI has priced its previous models at, listed below. 

Model	Input / M	Output / M	Change
GPT-5.6 Sol (Launch rate in July 2026)	$5	$30	-
GPT-6 Astra (Sept 2026)	$10	$50	~2x input, ~1.7x output
GPT-7	Not announced	Not announced	-

Given flagship prices are going up, not down, it’s a safe bet that GPT-7 will cost $10 or more for inputs and $50 or more for outputs. This is also fairly consistent with AI companies using low costs to get market share then upping the costs over the long term to turn a profit and pay back investors.

The above figures also don’t tell the whole story, because your per-task cost is going to matter more than your per-token cost. Namely, you can pay less for tokens but lose out on the volume you need to actually get something done. Conversely, you might wind up paying more per-token for GPT-7 but end up saving more with per-task costs than if you’d used GPT-6.

Here were the GPT subscription tiers at Astra’s launch for reference: Plus ($20/mo), Pro ($200/mo), Business ($100/mo), Enterprise (Variable). No new tier was introduced for Astra.

GPT-7 benchmarks

No GPT-7 benchmarks currently exist (Note: Are you sensing a theme in this article?) Below are the Astra results that GPT-7 will likely be compared against. 

Vendor-reported (The numbers that OpenAI have released themselves):

Benchmark	GPT-6 Astra	GPT-5.6 Sol
ARC-AGI-3 (custom harness)	99.9%	-
ARC-AGI-3 (standard harness)	62.7%	-
FrontierMath Tier 4	97.6%	-
OSWorld 2.0 (computer use)	72.6%, ~40 min/task	65.7%, ~75 min/task
DeepSWE v1.1	74.1%	72.7%
Terminal-Bench 4.0	57.9%	37.3%
Internal Database Migration	63.9%	42.7%
ExploitBench	100%	78.5%
ExploitGym	42.4%	30.3%

Third-party (Artificial Analysis, single harness):

Index	GPT-6 Astra	GPT-5.6 Sol	Claude Fable 5.1
Intelligence Index	61	61	66
Coding Agent Index	67	~65	70 (67.2 in some runs)

GPT-7 vs GPT-6 Astra

Because GPT-7 has yet to be released, direct comparison isn't possible yet. This section will be updated as more reports come in.

How to get access to GPT-7

GPT-7 has not been released anywhere yet, and so there is no way to access it. What we do know is how OpenAI released Astra, which acts as a signal for how they will likely release GPT-7: access through a gated program first, then paid ChatGPT tiers top-down, then API, with the model's most sensitive capabilities held back. Free-tier access to a flagship at launch has not happened since GPT-5.

GPT-7 in the API

No GPT-7 API endpoint, model string, or rate limits exist. Developer-relevant facts about the current flagship:

• Model ID: gpt-6-astra
• Context: ~1.05M tokens combined prompt and retained conversation
• Pricing: $10 / $1 (cached) / $50 per million tokens
• Modalities: text + image in, text out
• Codex: Astra introduced a context-preservation mechanism that replaces repeated summarization with searchable notes across context windows, enabled via config.toml and slated to become default

What is "Bel"?

Bel is an alleged OpenAI pretraining run, leaked August 25, 2026 by an anonymous X account, described as a 10T+ parameter successor to the "Doug" base model that underlies Astra. This claim is currently unconfirmed by OpenAI or with any outlet that has concrete documents. So, take anything Bel-related with a huge grain of salt.

Is Bel the same as GPT-7?

No, even by the leak's own framing. Bel is described as a base model for post-GPT-6 products, not a product itself.

Why Bel could be a hoax

The name "Bel" may be evidence that the leak isn't real because even as the name for a base model, it's fairly on the nose. Bel is a title that means 'Lord' or 'Master', and applies to various Mesopotamian gods: it's where you get the Biblical "Ba'al" and the derivative "Beelzebub" from.  

On its surface, matching occult-style names to AI models is very much the realm of an internet conspiracy theory. Then again, in 1996 Reebok decided to name one of its sneakers the "Incubus", so bad naming like this isn't completely unheard of.