GPT-6 release date: what we actually know right now

What OpenAI has actually said about GPT-6

Not much. That's the honest answer. But the bits Sam Altman has dropped in interviews paint a rough picture.

In an August 2025 interview with CNBC, Altman said "people want memory" and described it as the thing he's most excited about for the next generation. He talked about a model that remembers your tone, your past projects, how you work. Not just within a conversation, but across sessions, across weeks.

He also acknowledged that the GPT-5 launch was, in his words, "totally screwed up." The rollout was messy, the naming got confusing with sub-versions (5.2, 5.3-Codex, now 5.4), and a lot of users felt the jump from GPT-4 didn't match the hype. That seems to be informing how they think about GPT-6.

In a separate conversation reported by The Neuron, Altman said something interesting: "The main thing consumers want right now is not more IQ." Enterprise customers still want smarter models, but regular users want better experiences, faster responses, and a model that actually knows them. That's a different product direction than the benchmark-chasing of GPT-4 and GPT-5.

The release timeline math

If you look at how OpenAI has shipped major versions, there's a pattern. Not a strict one, but enough to make a reasonable guess.

The gap between GPT-4 (March 2023) and GPT-5 (August 2025) was about 29 months. Altman said GPT-6 will come faster than that. He also said his team "rarely sets targets more than six months ahead," which tells you they don't have a firm date locked in either.

OpenAI officially confirmed in October 2025 that GPT-6 would not ship that year. With GPT-5.4 just landing on March 5, 2026, the earliest realistic window for GPT-6 looks like mid-to-late 2026. Maybe Q3, maybe Q4. But honestly, nobody outside OpenAI knows for sure, and based on Altman's own admission, they might not either.

One thing worth watching: OpenAI has been shipping sub-versions fast. GPT-5 to 5.2 to 5.3-Codex to 5.4, all within seven months. It's possible they keep doing incremental releases (5.5? 5.6?) while GPT-6 trains in the background. The "GPT-6" label might land later than people expect simply because OpenAI keeps extending the 5.x line.

What GPT-6 will probably look like

Based on what Altman and OpenAI have actually talked about, a few themes keep coming up. I'm not going to list speculative parameter counts or made-up benchmark predictions because nobody has those numbers. But the direction is pretty clear.

The infrastructure behind it

One thing that is concrete: OpenAI has been stacking hardware. The Stargate project with Microsoft involves data centers running on over 5 gigawatts of power with more than 2 million chips. Their Abilene, Texas facility already has 100,000+ Nvidia GPUs running. They took delivery of DGX B200 systems in late 2024 and started getting GB200 racks from Oracle in mid-2025.

This doesn't tell you when GPT-6 ships, but it tells you they're investing at a scale that only makes sense if something significantly bigger than GPT-5 is in the pipeline. Training a model on this kind of infrastructure takes months, which lines up with a second-half-of-2026 release.

Where GPT-5.4 stands right now

For context on where things are today: GPT-5.4 launched March 5, 2026. It's the current best OpenAI has. A million-token context window, 128K output, native computer use that beats human performance on desktop navigation (75% on OSWorld vs. human 72.4%), and 83% on GDPval across 44 professional occupations.

We did a full breakdown of GPT-5.4 pricing and benchmarks and a head-to-head comparison with Claude Opus 4.6 and Gemini 3.1 Pro if you want the full picture of where things stand today.

GPT-5.4 is good. Genuinely good. You can see exactly where it stands against the competition on our LLM leaderboard. But it's also clearly an iteration on 5.2 rather than a generational leap. GPT-6 is where OpenAI seems to be aiming for that leap, especially on memory and autonomy.

What you should ignore

A lot of blogs are already publishing "GPT-6 benchmarks" and "GPT-6 pricing" articles. Those numbers don't exist. The model doesn't exist yet. Anyone giving you specific parameter counts ("20 trillion parameters!") or benchmark scores is guessing or making things up for clicks.

Same goes for "leaked screenshots" of GPT-6 interfaces. OpenAI hasn't shown anything. If a release date gets confirmed, it'll come from openai.com or Sam Altman's social media, not from a random Medium post.