ChatGPT-6 Leaks: Release Date, New Features and What Actually Shipped

Update, September 4, 2026: ChatGPT 6 is no longer a leak. OpenAI launched GPT-6 Astra on September 3, 2026, and the naming question that ran through every rumour on this page is finally closed: Astra shipped as GPT-6, not as another GPT-5 point release. It is the first model OpenAI has ever designated Critical for cyber capability under its Preparedness Framework, which is why the rollout looks nothing like a normal flagship launch. Approved defenders in OpenAI’s Daybreak program get it first, 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 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. Earlier updates: OpenAI shipped the “Spud” model on April 23, 2026 as GPT-5.5, then GPT-5.6 on July 9, 2026. Both were point releases. The rumour history below is preserved as a record of what the field expected, with what actually shipped marked against it.

The GPT-6 rumour cycle is over. OpenAI launched GPT-6 Astra on September 3, 2026, close to a year after the first “GPT-6 is weeks away” posts started circulating. Along the way, the model everyone was actually tracking, codenamed “Spud,” finished pretraining on March 24, 2026 and shipped a month later as GPT-5.5, not GPT-6. Greg Brockman had described it as “two years of research” with a “big model feel.” This page keeps the leaks that drove that year of speculation and marks each one against what OpenAI eventually put on the record.

It is worth seeing how the predictions scored. Polymarket traders put 78% probability on a release by April 30 and over 95% by June 30, 2026, and something did ship inside that window, just not GPT-6. The benchmark leak that had Spud scoring in the high 70s on SWE-bench Pro missed badly, with the real number landing at 58.6%. And every leaked launch date pointed at April 2026, nearly five months early. If you have been reading GPT-6 leaks for the last year, that hit rate is the most useful thing on this page.

The Key Takeaways

• Pretraining Done, Then Shipped: Sam Altman confirmed that OpenAI’s next frontier model (codenamed “Spud”) finished pretraining on March 24, 2026 and was “a few weeks” from release. It arrived on April 23, 2026 as GPT-5.5.
• Naming Resolved, Twice: The April model shipped as GPT-5.5, in three variants (standard, Thinking, and Pro). The real GPT-6 came five months later, on September 3, 2026, as GPT-6 Astra, API id gpt-6-astra.
• Unified Super App, Still Not Real: Rumours pointed to a single product combining ChatGPT, Codex, and an “Atlas” browser launching alongside the model. Neither GPT-5.5 in April nor GPT-6 Astra in September arrived with it.
• Hardware is Real: OpenAI has confirmed multi-billion dollar deals with AMD (6 GW) and Broadcom (10 GW) for next-generation AI infrastructure, with Spud trained at the Stargate facility in Abilene, Texas on over 100,000 H100 GPUs.

What We Officially Know About GPT-6

In the fast-moving world of artificial intelligence, it can be tough to separate solid facts from speculative hype. With GPT-6 that separation is finally easy, because the model exists: GPT-6 Astra, launched September 3, 2026. What follows is the confirmed record, starting with what OpenAI’s leadership said on the way there and ending with what shipped.

Here is a breakdown of what is actually confirmed as of September 2026:

• Pretraining Complete: Sam Altman confirmed on March 24, 2026 that pretraining for the next frontier model was finished. He described the timeline to release as “a few weeks” and called it “a model that could meaningfully accelerate the economy.” That model shipped as GPT-5.5.
• Two Years in the Making: Greg Brockman provided additional context, describing Spud as representing “two years of research” with a “big model feel with no incremental framing.” This signals OpenAI views this as a generational leap, not a minor update.
• A Promise of a Major Leap: In a widely cited interview with WIRED, OpenAI CEO Sam Altman made a direct promise: “GPT-6 will be significantly better than GPT-5… and GPT-7 will be significantly better than GPT-6.”
• Scientific Ambitions: Altman also hinted that “by 6 or 7 we’ll see more meaningful scientific capability,” pointing to AI moving beyond conversation and content creation into complex problem-solving and research. OpenAI now describes GPT-6 Astra as state of the art on science, reporting 64.6% on Terminal-Bench Science against 52.6% for Anthropic’s Fable 5.1.
• Sora Sacrificed for Spud: OpenAI discontinued its Sora video generation tool to redirect GPU resources toward Spud’s training. The Sora team was reassigned to world simulation research for robotics applications.

Those details described a model OpenAI internally viewed as a landmark release rather than a routine upgrade, and the eventual GPT-6 launch bore that reading out. The infrastructure investment, the two-year development cycle, and the willingness to shut down a high-profile product all pointed to something genuinely significant.

Which AI is the best right now?

ChatGPT 6 Will Have Long-Term Memory and Personalization

The most persistent rumour surrounding GPT-6 features was a massive upgrade in its ability to remember. This was never just about recalling the last few lines of a conversation. The talk was of true long-term memory, a system that would let the AI retain key information about you, your preferences, and your projects across all your conversations, potentially for weeks or months.

The goal of this would be a revolutionary level of personalization. Imagine an AI that remembers you are a software developer who prefers Python, that you’re planning a vacation to Italy next spring, or that you prefer a formal tone in business emails. It would stop asking repetitive questions and start anticipating your needs, making every interaction feel uniquely tailored to you.

GPT-5.4 already introduced improved memory features, and the leaks predicted GPT-6 would take this much further, with an unverified claim of a 2 million token context window. The published figure came in lower. GPT-6 Astra carries a 1,050,000-token context window, with a maximum input of 922,000 tokens, 128,000 tokens of output, and a knowledge cutoff of April 30, 2026. That is the same envelope as the GPT-5.6 family, so the memory leap the rumours promised is not what this release delivered.

Will ChatGPT 6 Be “Self-Learning”?

Perhaps the most groundbreaking of all the GPT-6 leaks was the theory that the next model would be capable of continuous learning. This idea, which spread rapidly online, was largely fuelled by a real research paper from MIT titled Self-Adapting Language Models (SEAL).

The concept suggests a radical departure from how current AIs work. Instead of being a static tool that is trained and then released, a self-adapting model could theoretically learn and improve on its own after deployment, getting progressively smarter through its daily interactions. So, what would this actually look like in practice?

The core ideas behind these self-adapting LLMs

• Learning from Experience: The AI would be able to analyze its own performance on tasks. When it generates a suboptimal or incorrect answer, it could identify the error and create its own training data to correct the mistake in the future.
• Making Permanent Changes: This is the most critical part. Through a process of generating self-edits, the model could use reinforcement learning to make persistent weight updates. In simple terms, this means it wouldn’t just learn for a single session. It would permanently modify its own internal neural network to embed the new knowledge.
• Adapting to New Information: A self-learning model could theoretically keep up with a changing world. If a new scientific discovery is made, it could integrate this new information without needing a full-scale retraining by its developers.

It is crucial to understand that OpenAI never confirmed a link between the SEAL paper and GPT-6. The connection was pure speculation, largely driven by the fact that one of the paper’s authors now works at OpenAI. OpenAI employees did hint at “a capability that is very different from what we’ve seen before” in Spud, which only added fuel to the theory. The launch settled it: GPT-6 Astra ships with a fixed knowledge cutoff of April 30, 2026, and nothing in the launch materials describes a model that updates its own weights after deployment.

GPT-6 Agents and Agentic AI

Building on improved memory and reasoning, many experts expected GPT-6 to significantly advance the field of agentic AI. That is the prediction that landed. GPT-5.4 had already introduced computer use, scoring 75% on the OSWorld benchmark for desktop tasks. OpenAI now describes GPT-6 Astra as state of the art on computer use and browser use, and reports 72.6% on OSWorld V2-Offline in about 40 minutes per task, against 65.7% in about 75 minutes for GPT-5.6 Sol. Halving the time per task is arguably the more useful half of that result.

Imagine telling your AI to “find the best flights for a business trip and book the one with the best balance of cost and convenience.” Long-horizon agentic work is exactly what OpenAI led the GPT-6 launch with, so this is closer to routine than experimental now, and the tooling around OpenAI’s agent stack has grown up alongside the models. Two things still hold it back. The rumoured “Atlas” browser did not arrive with the model, and Anthropic reports a higher OSWorld figure for Fable 5.1 while noting it used a different release of the benchmark, so the autonomous-web-agent crown is not cleanly settled.

The Massive AI Infrastructure for GPT-6, 7 & 8

The software features are settled now, and the hardware being built to power what comes after is very real. Spud was trained at OpenAI’s Stargate facility in Abilene, Texas, using over 100,000 H100 GPUs. But the infrastructure story doesn’t end there. OpenAI has confirmed massive partnerships to create next-generation capacity for the models that follow Astra.

The OpenAI AMD Deal

One of the cornerstones of this infrastructure is a major partnership with chipmaker AMD. OpenAI has signed a deal to acquire a massive fleet of new GPUs, committing to an enormous 6 GW of computing power. The rollout for these new AMD Instinct GPUs is confirmed to begin in the second half of 2026.

Custom AI Chips with Broadcom

Going beyond standard hardware, OpenAI is also co-developing custom AI chips with Broadcom. This long-term partnership will create highly specialized data center accelerators designed specifically for OpenAI’s unique workloads. The scale of this project is even larger than the AMD deal, targeting an eventual 10 GW of accelerator capacity, with deployments scheduled to run from the second half of 2026 through the end of 2029.

The Timeline for the GPT-6 Release Date

The GPT-6 release date is settled: September 3, 2026. Getting there took two false summits, and every leaked date on this page pointed at the first one. Here is how the timeline actually ran, with each signal scored against it.

Here’s how each timing signal held up:

• Altman’s Own Timeline: On the day pretraining completed, Altman described the launch as “a few weeks” away. He was right about the cadence and wrong about the model: what arrived weeks later, on April 23, 2026, was GPT-5.5.
• Prediction Markets: Polymarket traders assigned 78% probability of release by April 30 and over 95% by June 30, 2026. A model did ship inside that window, but the market was pricing OpenAI’s next frontier release, not the GPT-6 name, which was still five months away.
• Unverified Leak: An anonymous source claimed April 14-16, 2026 as the launch date, alongside a unified “super app” combining ChatGPT, Codex, and a new “Atlas” browser. Those dates passed without an announcement, and no version of that super app has shipped since.
• Naming Decision (Resolved): OpenAI chose GPT-5.5 in April, with a SWE-bench Pro score of 58.6% against GPT-5.4’s 57.70%, well short of the “high 70s” leaks and a fair explanation for keeping the GPT-5 branding. The GPT-6 name was held back until September 3, 2026, when Astra took it.
• Rollout Order (Not the Usual One): Previous launches went paid tiers first, then the free tier, then Enterprise API. GPT-6 Astra does not follow that pattern. Approved defenders in OpenAI’s Daybreak program get access first, with Plus, Pro, Business, Enterprise, the API, AWS and Azure following over “the coming days”, no date given. Free-tier access has not been announced at all.

The reason for that unusual rollout is the safety designation. GPT-6 Astra is the first model OpenAI has ever classed as Critical for cyber capability under its Preparedness Framework, and that finding is also why OpenAI paused Astra work in August 2026, then shipped the model sixteen days after briefing Axios about the pause. The price reflects the tier it sits in: $10 per million input tokens and $50 per million output, roughly 2.5 times GPT-5.6 Sol’s $4 / $20, with fast mode at double that again.

Overview

The GPT-6 picture is no longer a collection of vague promises and fan theories. Pretraining on “Spud” finished in March 2026, that model shipped as GPT-5.5 in April, GPT-5.6 followed in July, and the generational jump the leaks kept promising arrived on September 3, 2026 as GPT-6 Astra. The infrastructure at Stargate is built, the AMD and Broadcom deals are running, and OpenAI now calls its own model state of the art on computer use, software engineering and science.

Scoring the leaks on this page: the release-date rumours were all wrong, every one of them pointing at April 2026 for a model that landed in September. The “high 70s on SWE-bench Pro” leak was wrong. The 2-million-token context leak was wrong, with the real figure at 1,050,000. The self-learning theory built on the SEAL paper did not materialise, and neither did the unified super app or the Atlas browser. What the leaks did get right was the shape of the thing: a long-horizon agentic model, built on enormous new compute, that OpenAI treats as a generational step rather than a point release. Two caveats travel with the launch numbers, though. The headline 98.6% on ARC-AGI-3 measures action efficiency against a human baseline rather than a solve rate, and OpenAI left GDPval, its own benchmark for economically valuable work, out of the launch materials entirely. For the full breakdown, read GPT-6 Astra is here: benchmarks, price, and what the numbers don’t say.