OpenAI released the model codenamed "Spud" on April 23, 2026, but branded it GPT-5.5 rather than GPT-6 because the performance gains, while significant in specific areas, didn't justify a full generational jump in naming. The decision reflects a broader shift in how AI labs approach model releases: incremental improvements shipped frequently rather than waiting for dramatic leaps. GPT-6 remains unannounced and likely won't arrive until late 2026 or 2027.

What Actually Happened to the Model Everyone Was Waiting For?

For months, "GPT-6 release date" dominated AI search trends. Prediction markets offered odds on its arrival. Tech blogs built entire tracking pages around it. Then, on April 23, 2026, OpenAI shipped the model everyone assumed would carry that name, except it didn't.

The model finished pre-training on March 24, 2026, at OpenAI's Stargate data center in Abilene, Texas, using a training cluster of more than 100,000 H100 GPUs (graphics processing units, the specialized chips used to train AI models). Sam Altman's comment that day describing the launch timeline as "a few weeks" out triggered two months of increasingly specific rumors. One unverified source claimed an April 14 launch alongside a "super app" combining ChatGPT, Codex, and OpenAI's Atlas browser, with a 40% performance jump and a 2-million-token context window (the amount of text a model can process at once). April 14 came and went with no announcement. Nine days later, the real model arrived as GPT-5.5 in three variants: standard, GPT-5.5 Thinking, and GPT-5.5 Pro.

Why Didn't OpenAI Call It GPT-6?

The honest answer is that the performance jump wasn't big enough to justify the bigger number. OpenAI's own benchmark data makes this case clearly. GPT-5.5 scored 58.6% on SWE-Bench Pro, a test of software engineering ability, compared to GPT-5.4's 57.7%. That's a 0.9 percentage point improvement, well short of the "high 70s" some leaks had predicted.

However, the gains weren't evenly distributed. In specific areas, GPT-5.5 delivered genuine step changes. Long-context reliability jumped from 36.6% to 74.0% on the MRCR v2 benchmark at 512K-1M token contexts (a measure of how accurately the model recalls information across very long documents). The model also produced roughly 60% fewer hallucinations, or false information, than its predecessor, and scored 82.7% on Terminal-Bench 2.0, about 13 points ahead of Claude Opus 4.7 at the time.

The real story behind the hype involved memory and personalization features. Altman had spent the prior year describing memory as the most important feature of OpenAI's next-generation model, the ability to recall preferences, ongoing projects, and past conversations across weeks, not just within a single chat. Those features landed in GPT-5.5, which likely explains why so many people assumed it had to be the "real" next-generation model.

How to Decide Whether to Switch to GPT-5.5

• For ChatGPT App Users: You're already on GPT-5.5 by default as of May 5, 2026. There's no separate "GPT-6" product you're missing out on by not switching anything. The upgrade happened automatically.
• For Developers Using the API: GPT-5.5 doubled its predecessor's API cost, from $2.50/$15 per million tokens (input/output) to $5.00/$30. The long-context and hallucination-reduction gains justify the move for workloads involving large documents or extended back-and-forth conversations, but cost-sensitive use cases that don't rely on long context may stick with GPT-5.4 pricing.
• For Teams Building on Long Documents: The jump in long-context reliability from 36.6% to 74.0% represents a genuine capability shift, not a minor bump. If your work involves processing extended documents or maintaining accuracy across very long conversations, the upgrade is worth the cost increase.

One detail puts OpenAI's resource priorities in sharp focus: Sora, OpenAI's video generation app, was shut down on the exact same day Spud finished pre-training, March 24, 2026. That timing wasn't coincidental. Sora was reportedly burning around $1 million a day in compute costs at its peak, while its total lifetime in-app revenue came to just $2.1 million. Monthly downloads had fallen roughly 66% between November 2025 and February 2026. Redirecting that compute toward the model that became GPT-5.5 signals a clear priority: OpenAI is betting on the core ChatGPT model and its agentic capabilities over consumer side-products that aren't pulling their financial weight.

When Will GPT-6 Actually Arrive?

Genuinely uncommitted, and probably further out than most people assume. Based on OpenAI's pattern of frequent point releases, the company shipped six separate GPT-5 family models (5.0 through 5.5) in under eight months, a true generational jump to "GPT-6" branding now looks like a Q3-or-later 2026 story at the earliest, with real probability mass extending into 2027.

That's consistent with how Altman has talked about roadmaps generally. He's said OpenAI rarely sets high-confidence targets more than six months out, and the renaming of Spud to GPT-5.5 supports the idea that OpenAI is leaning toward incremental, frequent releases rather than another big symbolic leap.

The more immediate thing to watch isn't GPT-6; it's GPT-5.6. A brief reference to "gpt-5.6" surfaced in OpenAI's developer tools before disappearing, which is the strongest non-speculative evidence it exists in some form. Recent reporting points to a possible release around June 25, 2026, with rumored specs including a 2-million-token context window and pricing roughly five times cheaper than Anthropic's Claude Fable 5. None of that is confirmed by OpenAI directly as of this writing, so treat it as a credible rumor, not a fact, until an official announcement lands.

If the timing holds, it would put GPT-5.6 in the same general window as two other major frontier releases people are actively tracking: Anthropic's Claude Mythos and Google's Gemini 3.5 Pro. Three major labs converging on roughly the same few weeks isn't a coincidence; it's competitive pressure, each one racing not to be the last to ship before the others define the news cycle.

The practical takeaway is simple: model names and model capability have become two separate stories. A bigger number doesn't necessarily mean a bigger leap anymore, and a smaller-sounding point release can quietly ship the exact features everyone was waiting for under a bigger name. Whether you're a casual user or building a product on top of these models, stop anchoring decisions to a specific name or number, and pay attention to what a release actually does on launch day.