ChatGPT 6 Release Date: It Launched on September 3, 2026

Update, September 4, 2026: ChatGPT 6 exists. OpenAI launched GPT-6 Astra on September 3, 2026, ending a year of speculation and settling the naming question: Astra shipped as GPT-6, not as a GPT-5 point release such as GPT-5.7. It is the first model OpenAI has ever designated Critical for cyber capability under its Preparedness Framework, which is why the August pause happened and why the rollout looks nothing like a normal flagship launch. Access starts with approved defenders in OpenAI’s Daybreak program, with ChatGPT Plus, Pro, Business and Enterprise, the API, AWS and Azure following over “the coming days”, no date attached. API pricing is $10 per million input tokens and $50 per million output, roughly 2.5 times GPT-5.6 Sol. Note also that there is no product called “ChatGPT 6”; the model is GPT-6 Astra, inside ChatGPT. We take the launch apart number by number in GPT-6 Astra is here: benchmarks, price, and what the numbers don’t say. The sections below are updated to that reality; the pre-launch rumor history is preserved as a record of what the field expected.

For months, the AI world held its breath for OpenAI’s next frontier model, codenamed “Spud.” That wait ended on April 23, 2026, when OpenAI released it as GPT-5.5, not GPT-6, and the company followed with GPT-5.6 on July 9, 2026. Both were point releases inside the GPT-5 family, not the generational leap the “weeks away” leaks promised. The leap came on September 3, 2026, when OpenAI launched GPT-6 Astra, five weeks after it first put the Astra name in public. The question is no longer whether ChatGPT 6 is coming. It is who can use it, what it costs, and which of the launch numbers survive a close reading.

If you are tired of chasing clickbait and vague roadmaps, you are in the right place. We have tracked every official statement, analyzed OpenAI’s infrastructure build-outs, and filtered out the baseless hype to give you a clear, honest picture of where the technology actually stands. This guide cuts through the confusion to bring you only verified updates, confirmed timelines, and credible expectations.

Here’s what we’ll cover.

• When it could realistically arrive
• What’s confirmed vs rumor
• Whether it’s worth waiting

The Key Takeaways

• GPT-6 shipped as Astra on September 3, 2026. OpenAI named the model on August 1, then launched it five weeks later as GPT-6 Astra, with the API id gpt-6-astra. The GPT-6 versus GPT-5.7 naming question is closed.
• It is the first model OpenAI has designated Critical for cyber. The August 7, 2026 finding that OpenAI could not rule out critical cyber capabilities was not a false alarm. Astra shipped carrying that designation, where every earlier model, GPT-5.6-Sol included, sat at the lower High threshold.
• You probably cannot use it yet. Approved defenders in OpenAI’s Daybreak program get it first, with Plus, Pro, Business, Enterprise, the API, AWS and Azure following over “the coming days”. OpenAI gave no dates, and free-tier access has not been announced.
• The April 14 rumor passed quietly. An unverified leak pegged April 14, 2026 as launch day, tied to a “super app” combining ChatGPT, Codex, and the Atlas browser. It came and went without an OpenAI tweet, blog post, or drop. Polymarket traders trimmed “GPT-6 by April 30” from 78% to roughly 72%.
• It costs about 2.5 times GPT-5.6. GPT-6 Astra runs $10 per million input tokens and $50 per million output, level with Anthropic’s Fable 5.1, against $4 / $20 for GPT-5.6 Sol. Fast mode doubles both, Batch and Flex halve them. Until the rollout reaches your plan, GPT-5.6 is still what ChatGPT actually serves you.
• The headline benchmark is not what it looks like. The 98.6% on ARC-AGI-3 measures action efficiency against a human baseline, not a solve rate, and OpenAI has itself published how enabling two harness settings tripled that score. Read the launch numbers with that in mind.
• Don’t rebuild around it yet. Until access reaches your plan, build on the models you can actually run today and keep the workflow portable, then switch when the rollout arrives.

ChatGPT 6 Release Date Status

ChatGPT 6 has a release date now: September 3, 2026. What is still moving is availability, and for most people that is where the practical answer lives.

Current Status

As of September 4, 2026, GPT-6 exists and is called GPT-6 Astra. Getting there took an unusually crowded road. The “Spud” model everyone was waiting for shipped on April 23, 2026 as GPT-5.5, OpenAI followed with GPT-5.6 on July 9, 2026, and only after seven GPT-5 models, from GPT-5.0 through GPT-5.6, did the numbered jump arrive. The “a few weeks” timeline Sam Altman gave on March 24 resolved into GPT-5.5; the cadence was right, the GPT-6 naming prediction was five months early.

The one thing that did change is the name. On August 1, 2026 OpenAI introduced Astra as its next major model, and it did so through research output rather than a product page. An internal version of Astra solved ten open problems in mathematics and theoretical computer science, every one of them untouched for at least a decade, and OpenAI published a Lean 4 certificate for each result in the public openai/ten-proofs repository under an Apache 2.0 license. The headline result is the first explicit construction of a non-sofic group, a question open since Mikhail Gromov introduced soficity in 1999. OpenAI says the tokens behind all ten solutions would have cost roughly $2,000 at Sol API rates.

Read that carefully, because at the time it was not a launch. Astra had no release date, no pricing, no context window, no model card and no ChatGPT availability, and OpenAI had not said whether it would become GPT-6 or land as another point release inside the GPT-5 line. That gap closed on September 3, 2026. We went through the launch result by result in our full breakdown of what OpenAI actually shipped as GPT-6 Astra.

Then it stalled. On August 7, 2026 OpenAI published a security note saying its latest internal evaluations of Astra showed “significant advancements in agentic coding and cybersecurity”, and that those results meant it could not rule out critical cyber capabilities under its Preparedness Framework. The Critical threshold is defined as a model that can identify and develop working zero-day exploits in hardened real-world systems without human intervention. OpenAI noted that every previous model it had evaluated, GPT-5.6-Sol included, sat at the lower High threshold, and it said Astra was not involved in the Hugging Face incident. Its response was to impose isolated testing environments, restricted network and tool access, stronger model weight protections and sandboxed execution, and to pause internal activities involving Astra that did not yet meet those controls.

Eleven days later the scope became clearer. OpenAI told Axios on August 18, 2026 that it had paused two weeks of deployment-focused reinforcement-learning training, that its largest planned frontier RL run remained on hold, that a significant number of Astra and cyber-related research workloads were still paused, and that it was rewriting the Preparedness Framework itself, most of which dates to 2023. Chief scientist Jakob Pachocki framed the changes as broader than any single incident. At the time that read as a timeline signal pointing later. It was not: Astra launched sixteen days after that briefing, on September 3, carrying the Critical designation that caused the pause in the first place. What the pause bought was security controls, not delay. It also sits beside what OpenAI told TIME the same month, that it expects an internal system it would call AGI before the year is out, a claim we weigh in what Altman actually promised about AGI in 2026.

Against that, an August 25, 2026 post on X from the account @synthwavedd claimed OpenAI had finished a pretrain codenamed “Bel”, successor to “Doug”, at more than 10 trillion total parameters, with Doug the base for Astra and GPT-6 and Bel intended as a post-GPT-6 base. It travelled fast, past 660,000 views and into several tech sites within a day. Treat it as a rumor, and note that the GPT-6 launch did not settle it. Every republication we could find cites that one post, OpenAI has said nothing, and several write-ups garbled it, swapping Bel and Doug so that Bel becomes the GPT-6 base.

An unverified source pegged April 14, 2026 as the launch date, alongside a rumored “super app” combining ChatGPT, Codex, and the Atlas browser. That date came and went without so much as an OpenAI tweet, and the model that did arrive nine days later was GPT-5.5, not GPT-6. The real answer landed four and a half months after that: September 3, 2026. No leaked date got it right.

That rapid release cadence was deliberate. OpenAI spent a year gathering real-world data from each GPT-5 generation to tune the agentic behaviours that now define GPT-6.

So… when could it actually land?

It landed on September 3, 2026, later than the leaks predicted and sooner than the August pause suggested. Altman’s “few weeks” framing from March described GPT-5.5. The April 14 rumor missed by nearly five months. Our own reading in late August put late 2026 at the optimistic end, with 2027 realistic, and that was wrong too, because the pause turned out to be about security controls rather than the schedule: OpenAI halted internal Astra activity that did not meet strengthened controls, built the controls, and shipped. The one question still without a date is availability, which we cover below. Frontier releases are often shaped by safety work like red teaming, and OpenAI describes red teaming as a structured way to probe models for failures before release. In this case that work changed how the model ships, not whether it ships.

What would count as real confirmation?

We treated GPT-6 timing and features as “confirmed” only when at least one of these appeared. All of it now has.

Those bars are cleared. On September 3, 2026 OpenAI announced the launch, published model documentation and pricing under the API id gpt-6-astra, and company president Greg Brockman put the GPT-6 name on the record. The ledger below keeps the pre-launch entries so you can see which claims held.

1. An OpenAI blog post or release notes update.
2. A published model card or system card.
3. Official API docs explicitly mentioning GPT-6.
4. A statement attributed to OpenAI in reputable reporting, with clear wording.

Additionally, infrastructure plays a massive role. OpenAI’s Stargate Project signals serious infrastructure investment, but OpenAI hasn’t linked it to a GPT-6 release date. Treat it as context about scale, not a countdown.

What We Know About ChatGPT 6

It is easy to get confused by social media hype, especially when screenshots of “new” models circulate on X (formerly Twitter) or Reddit. Let’s look at the ChatGPT 6 rumors versus what has been officially stated by leadership or confirmed by reliable tech reporting. We call this our Verification Ledger.

Confirmed facts vs rumors

The table below tracks the most popular claims currently circulating. We only mark something as “Confirmed” if it comes directly from an official OpenAI announcement or a primary news source like a verified press release.

The verification ledger

Always check the source of ChatGPT 6 leaks. If a screenshot shows a “release date” prompt inside ChatGPT, it is usually a “hallucination” (the AI making things up) rather than a verified leak. Large Language Models (LLMs) are predictive engines; if you ask them “When are you released?”, they often predict a date based on internet text rather than accessing a hidden internal database.

Furthermore, true leaks rarely appear as simple text outputs. Legitimate leaks historically come in the form of API documentation code strings, updated developer terms of service, or reliable reports from journalists who have spoken to insiders. If you see a “GPT-6” dropdown menu in a screenshot, remember that inspecting and editing HTML elements in a browser takes seconds. Mockups are easy to create, so screenshots alone aren’t evidence.

Security Tip: If you see “early access” offers, treat them as suspicious; enable MFA (Settings, then Security) and only use official OpenAI sign-up flows.

Expected ChatGPT 6 Features

We now have specs rather than direction. GPT-6 Astra carries a 1,050,000-token context window, with a maximum input of 922,000 tokens and 128,000 tokens of output, takes text and images in and returns text, and has a knowledge cutoff of April 30, 2026. The sections below keep the pre-launch expectations, with what actually shipped against each one.

Memory and personalization

Sam Altman has described current AI memory as being in the “GPT-2 era” relative to what’s coming, a striking claim that suggests this is the area where GPT-6 will make its biggest leap. Users expect GPT-6 memory to be vastly improved. Currently, most models have a “context window” that limits how much they remember in a single session, and they struggle to remember things between different sessions unless specifically instructed.

• Context. The AI might remember details from a conversation you had months ago without you needing to remind it.
• Adaptability. It could learn your writing style, coding preferences, or business tone without needing new instructions every time.

This “persistent memory” would fundamentally change the user experience. Instead of starting from scratch with every “New Chat,” you would be picking up a relationship where you left off. Imagine an AI that knows you prefer Python over JavaScript, or that you always want your marketing emails to have a professional but friendly tone. This reduces the “prompt engineering” burden on the user significantly.

Agents and autonomy

The buzzword you will hear often is “Agents.” OpenAI is investing heavily in agentic capabilities, as evidenced by the OpenAI Agents SDK. It’s reasonable to expect future frontier models to push this further, but GPT-6 specifics are unconfirmed.

• Example. Instead of just writing an email draft, an agent might draft it, find the recipient’s address in your contacts, and queue it for sending.
• Autonomy. This moves the AI from a tool you talk to, to a tool that works for you.

In an autonomous workflow, the model might break a complex goal, like “Plan a travel itinerary,” into sub-tasks such as checking flights, verifying hotel availability, comparing prices, and presenting a final package. GPT-5.5 already took a significant step here, scoring 78.7% on OSWorld-Verified, narrowly ahead of Claude Opus 4.8, and operating software interfaces directly via native computer use. This is the expectation GPT-6 Astra was built to meet: OpenAI describes it as state of the art on computer use, browser use, software engineering, cybersecurity, science and professional work, and long-horizon agentic work is the capability the launch leads with.

OpenAI hardware and interface redesign

One area that rarely gets covered in GPT-6 speculation is hardware. Altman has confirmed that OpenAI is developing “a small family of devices” optimised for AI-first interactions, a deliberate move away from traditional form factors like the smartphone keyboard. This suggests GPT-6 won’t just be a smarter chatbot; it may debut alongside new hardware designed to take full advantage of its agentic capabilities.

Altman has also said that “bolting AI onto the existing way of doing things” won’t work as well as redesigning products from an “AI-first” perspective. ChatGPT’s interface is expected to change substantially beyond its current chat-based design in the GPT-6 era.

A new training paradigm

GPT-6 won’t just be a bigger version of GPT-5. Altman has indicated that the GPT-5 and GPT-6 generation uses reinforcement learning as its core training approach, rather than the pre-training paradigms used for earlier models. This is significant: reinforcement learning allows the model to discover patterns and solve problems through trial and error, rather than just absorbing existing text. OpenAI describes this as enabling the AI to potentially contribute to “new algorithms, physics, and biology,” meaning early scientific discovery, not just retrieval.

Multimodal improvements

We also expected significant strides in GPT-6 multimodal video and audio capabilities. That is the one prediction the launch did not deliver. GPT-6 Astra takes text and images in and returns text, so real-time video understanding and audio composition are not part of this release. The multimodal leap, if it comes, belongs to a later model.

GPT 6 vs GPT 5 Comparison

The comparison is no longer hypothetical on one side. The table below sets what the field expected of GPT-6 against what GPT-6 Astra actually shipped with on September 3, 2026.

Models in the GPT-5 class are strong at creative writing, coding and general knowledge, and still make logic errors. Astra is OpenAI’s answer to that, though several of its headline scores need reading carefully before you treat them as settled. For a broader look at how this stacks up against the rest of the industry, see our Best AI Models hub, which we update as each new frontier model ships.

Reasoning vs Agency

The shift is subtle but important.

• Current Models. Good at writing, coding snippets, and summarizing. They act like a very smart encyclopaedia.
• GPT-6 Astra. Built for multi-step work that runs long, with computer use and software engineering as the headline capabilities. It acts more like a smart intern.

Note: The GPT-6 context window is now a published figure rather than a guess: 1,050,000 tokens, with a maximum input of 922,000 and 128,000 tokens of output. That is the same envelope as the GPT-5.6 family, and well short of the “10 million tokens” charts that circulated before launch. The more important metric is still “effective context,” meaning how accurately the model retrieves a specific fact from a mountain of data, rather than how much data it can hold.

Access and Pricing Availability

How do you actually get your hands on it? This is the part that is still moving, and it does not follow OpenAI’s usual pattern.

Normally a new flagship goes to paid ChatGPT users first. GPT-6 Astra does not. Because it is the first model OpenAI has designated Critical for cyber capability, access opens with Daybreak, OpenAI’s application-based program for vetted cybersecurity defenders, including a Daybreak Blue track for organisations protecting critical infrastructure. Everything else follows behind it.

1. Daybreak first. Approved defenders get GPT-6 Astra ahead of everyone else. This is a vetted, application-based program, not a queue you can join by paying.
2. Then ChatGPT Plus ($20/mo), Pro ($200/mo), Business and Enterprise. OpenAI says these follow over “the coming days” and has attached no date to any of them.
3. Free and Go are not announced. ChatGPT Go ($8/mo, localized in some markets) and the free tier have no stated GPT-6 access. GPT-5.6 Luna remains the default there.

For developers, the model id is gpt-6-astra, priced at $10 per million input tokens and $50 per million output, with cached input at $1 and cache writes at $12.50. Batch and Flex run at half price, Fast mode at double. AWS and Azure are listed in the same “coming days” window. Expect tiering and rate limits early on: as of today Tier 5 requires $1,000 paid and 30+ days since the first successful payment, and tier rules can change, see platform.openai.com.

Should You Wait For GPT 6

GPT-6 is here, so the question changes shape. It is no longer whether to wait for the model, it is whether to wait for access. The answer is still no. Build on what you can run today: the principles carry across generations, GPT-5.6 is what most paid plans still serve, and at $10 / $50 per million tokens the cheaper models will stay the right default for plenty of work even after the rollout completes.

How to prepare for GPT 6

The best way to prepare for GPT-6 is to build good habits now. If your data is messy or your prompts are vague, a smarter model won’t fix your fundamental problems.

• Clean your data. AI is only as good as the information you give it. Organize your internal documentation; using tools to chat with PDFs and files helps here, so that when the new model arrives, it has a high-quality knowledge base to learn from.
• Refine prompts. Learn how to instruct AI clearly. The skill of breaking complex tasks into clear steps will be even more valuable when directing “agents.”
• Use Multi-Model Tools. Don’t lock yourself into one vendor. The AI landscape is competitive.

The “Use It Now” Strategy

Tip: You can use platforms like Fello AI today to access the best AI model currently available (whether that is GPT, Claude, or Gemini). This keeps your workflow flexible so you are ready the moment GPT-6 Astra reaches your plan. By building on a platform that aggregates models, you insulate yourself from rollout queues. If Astra access is slow to arrive but Claude Opus 5 or Gemini 3.1 Pro ships a breakthrough feature, you can switch instantly without rebuilding your entire stack.

Conclusion

The ChatGPT 6 release date turned out to be September 3, 2026, and not one leaked date got it right. The April 14 rumor missed by nearly five months, the “Bel” leak still has one source, and the August pause that looked like a delay was a security-controls pause the model shipped through. That is the case for ignoring leaked dates, not for chasing the next one.

What is left to watch is access rather than existence. Until the rollout clears Daybreak and reaches ordinary paid plans, the practical advice has not changed: organize your data, keep your prompts sharp, and use tools that let you swap models without rebuilding around any one of them.

Next Step: Check the current top-performing models on Fello AI and keep your workflow portable, so GPT-6 Astra is a switch rather than a migration.

FAQ