The GPT-6 release date is currently pegged at 83% odds before December 31st on Poly Market — but that headline barely captures how much the landscape has shifted while everyone was waiting. In the last 60 days, OpenAI shipped GPT-5.5 instead of GPT-6, lost its default slot inside GitHub Copilot, got replaced by Google inside Apple's entire ecosystem, and watched Claude take the top spot on every benchmark that matters. GPT-6 is no longer a victory lap. It's a rescue mission for a trillion-dollar IPO.

Here is everything we actually know, ranked by how confident the signals are — and what you should do about it before the model even ships.

What Is the GPT-6 Release Date and What Do We Actually Know?

The clearest signal right now comes from a Codex routing log entry dated around May 13th, which references two internal code names: Ember Alpha and Beacon Alpha, apparently tied to a GPT-5.6 or GPT-6 candidate. Poly Market currently prices a ship date before June 30th at 88%, and before December 31st at 83%. Those numbers move, but they've been sticky in the high-80s range for weeks.

Beyond that single routing entry, here is how every major rumor stacks up:

  • Context window of 1.5 million tokens or more — Multiple sources mention this figure, but no leaked configuration files back it up yet. Possible, not verified.
  • Redesigned reward audit pipeline — Several OpenAI researchers have hinted at this on technical podcasts following the Goblin incident. Credible based on internal signal.
  • Full AGI researcher autonomy by March 2028 — This was Sam Altman's own claim. He has since walked it back publicly, calling fully automated AI development "unfulfilling and dangerous." Treat this one as officially shelved.
  • Persistent memory (Dreaming V3) — Already real. It shipped on June 4th inside GPT-5.5, not GPT-6. Factual recall jumped from 41.5% to 82.8%. Altman originally promised this for GPT-6 and delivered it early.

The reason GPT-6 is late has nothing to do with engineering slowdowns. OpenAI's Stargate infrastructure build has already secured commitments for over 10 gigawatts of compute — roughly 1% of the entire US electrical grid, dedicated entirely to running AI models. The Apollo program at its peak consumed less than one gigawatt. You do not deploy that overnight. GPT-6 is a compute project on the scale of a national power grid, and it will ship when the infrastructure catches up.

Why Did Apple Drop OpenAI and Partner With Google Instead?

This one genuinely came out of nowhere. At Apple WWDC on June 8th — notably Tim Cook's final WWDC as CEO before John Turnus takes over on September 1st — Apple announced a strategic AI partnership with Google, not OpenAI. Gemini now powers the new Apple Foundation models and the next generation of Siri. ChatGPT was quietly sidelined inside iOS 27.

The deal is reportedly worth around $1 billion per year to Google. For OpenAI, it is not just the lost revenue. Apple represented one of the most visible consumer distribution channels in the world. Every iPhone user who interacted with an AI assistant was a potential ChatGPT convert. That pipeline is now pointing at a competitor.

This was the third of three major blows that landed inside three weeks. Microsoft moved first on June 2nd at Microsoft Build, announcing Project Polaris. Google moved second at Google IO on May 19th with Gemini 3.5 Flash. Apple closed it out at WWDC on June 8th. When Microsoft, Google, and Apple all pivot away from OpenAI inside a single month, that is not a coincidence. That is a signal.

Is Claude Now Better Than GPT-5.5 on Every Benchmark?

On the benchmarks that matter most right now, yes. GPT-5.5 launched on April 23rd and briefly held the top spot on the Artificial Analysis Intelligence Index at 60. Then Anthropic shipped Claude Opus 4.8 on May 28th, which immediately took number one at 61.4. Then on June 9th, Anthropic dropped Claude Fable 5 — described as the first publicly available Mythos-class model — and it now sits at 65 on the same index.

GPT-5.5 is at 60. That is a five-point gap that opened in under six weeks across two model releases from a single competitor.

The head-to-head results are even more direct. Tom's Guide ran GPT-5.5 against Claude Opus 4.7 on everyday tasks across seven categories — writing, reasoning, coding, image analysis, and more. Claude won every single round. GPT-5.5 did not take one category. That is the kind of result that makes people quietly switch their daily driver without making a big announcement about it.

There is also the SWE-Bench story. On SWE-Bench, GPT-5.5 scored 88.7%, which looks strong. But on SWE-Bench Pro — the version that actually maps to real engineering work — it dropped to 58.6%. Leaked internal targets had been sitting in the high 70s. That is a clean miss against OpenAI's own benchmarks.

What Is Microsoft Project Polaris and What Happens to Copilot?

At Microsoft Build on June 2nd, Microsoft announced Project Polaris, an in-house coding model they have been quietly training for over a year. Starting August 2026, Polaris becomes the default engine inside GitHub Copilot for every subscriber. That directly replaces GPT-4 Turbo, which has been powering Copilot under the hood for years.

Nearly 140,000 organizations run on Copilot right now. Most of them had no idea they were running on OpenAI infrastructure. Starting in August, they simply will not be. Microsoft also announced seven MAI models built entirely in-house. Mustafa Suleyman made the timeline explicit: the exclusivity ends in August.

For anyone selling into enterprise, this is the real market signal to watch. August is when enterprise demand starts moving off OpenAI at volume. A multi-model procurement strategy is no longer optional — it is the baseline expectation.

What Was the OpenAI Goblin Incident and What Did It Reveal?

Somewhere inside GPT-5.5's reward model training, a behavioral artifact slipped through. The model began inserting references to goblins and gremlins into completions, completely unprompted. Reddit surfaced the screenshots and they went viral within hours. Then someone leaked the Codex system prompt, and instruction number 140 literally told the model to stop talking about goblins. OpenAI had patched it from the front end, which only confirmed the artifact was real.

OpenAI published an official postmortem on April 30th titled Where the Goblins Came From. They traced the issue to a retired personality profile that had contaminated the reward model across multiple versions. The lesson is straightforward: when you sprint a reward model to ship faster, weird artifacts slip in and you do not catch them until users do. That is the dark side of training at speed, and it explains why the redesigned reward audit pipeline rumor for GPT-6 carries credible internal signal.

What Is OpenAI Dreaming V3 and Why Does It Matter More Than GPT-6?

While everyone debates benchmark points, Dreaming V3 — OpenAI's persistent memory system — already shipped on June 4th inside GPT-5.5. Factual recall jumped from 41.5% to 82.8%. You can turn it on right now.

Here is what 82.8% recall means in practice. Tell the model in March that you are building a React project with Tailwind. In June, open a new chat and ask it to build a component. It already knows your stack. If you are a marketer who has shared brand voice guidelines three times over six months, open a new chat in July and it writes in your tone without being asked.

This is strategically important beyond the feature itself. Every AI assistant today resets to zero when you close the tab. Dreaming V3 breaks that pattern. Once your profile is built, switching to a competitor means starting over from scratch. That is not just a product improvement — it is a retention mechanism dressed up as a user benefit. OpenAI knows exactly what they are building here, and it may matter more to long-term adoption than any benchmark GPT-6 posts at launch.

Is OpenAI Going Public? What the IPO Filing Really Means

On June 8th — the same day as Apple WWDC — OpenAI confidentially filed an S-1 for an IPO. That filing came exactly one week after Anthropic filed its own S-1 at a reported $965 billion valuation. OpenAI's current valuation sits around $852 billion. They are reportedly targeting $1 trillion or more by September, with Goldman Sachs and Morgan Stanley leading the offering.

The math underneath this is uncomfortable. OpenAI is on track for a $14 billion operating loss this year alone. Profitability is not expected before 2030. They need GPT-6 to be genuinely impressive — not for users, not for benchmarks, but to justify a trillion-dollar valuation in front of public markets. That is a different kind of pressure than shipping a good product.

Also worth noting: on the same day Altman filed for a trillion-dollar IPO, he published a post calling for an international organization with the power to slow frontier AI development and describing full automation as "unfulfilling and dangerous." A CEO calling for slower AI on IPO filing day is a tension worth sitting with.

What Is Deepseek V4 and Is It a Real Threat to OpenAI?

Deepseek V4 shipped on April 24th with open weights under an MIT license. Pricing landed at approximately $0.87 per million output tokens — roughly 28 times cheaper than Claude Opus 4.8. US government testing found it performs closer to GPT-5 era than GPT-5.5, but at that price delta, the capability gap barely matters for most production use cases.

The frontier is now splitting in two simultaneous directions. Claude Fable 5 is pushing the ceiling higher and charging $50 per million tokens for the privilege. Deepseek is collapsing the floor and giving the weights away for free. GPT-6 is walking into a market with a five-point benchmark gap to close, two fresh Anthropic models above it, and a sub-dollar open-weight competitor eating the bottom. That is the field. Plan accordingly.