GPT-6 Finally Arrives: What Can OpenAI's "Potato" Cook Up?

Honestly, when I saw GPT-6’s internal codename was “Spud” (potato), I almost sprayed my coffee.

Is OpenAI serious? A model that might change the trajectory of human civilization, and they call it “potato”? But then again, this might just be typical Silicon Valley geek humor—the more important something is, the more casual the name, like when Jobs called the iPhone project “Purple.”

Alright, jokes aside. GPT-6 is really here, and it arrived somewhat suddenly.

According to OpenAI’s official announcement, GPT-6’s pre-training was completed in Q1 2026. This is faster than many expected—after all, GPT-5.3 was released not long ago. But personally, I think this precisely shows OpenAI has achieved a qualitative leap in training efficiency and compute orchestration.

Regarding specific parameters, OpenAI uncharacteristically revealed some data: GPT-6 has approximately 5 trillion parameters, a 56% increase from GPT-5.3’s 3.2 trillion. But more crucially, inference speed—officially, GPT-6’s inference latency on identical hardware is 40% lower than GPT-5.3. Sounds counterintuitive, right? More parameters but faster?

I’ve looked at some technical details, and simply put: OpenAI did deep optimization on the MoE (Mixture of Experts) architecture, introducing a new “dynamic expert selection” mechanism that only activates the truly needed parameter subsets during inference. This is quite interesting because Anthropic’s Claude Opus 4.7 is using a similar approach.

But alongside GPT-6’s release, another story is brewing—Musk suing Altman.

If this happened two years ago, I’d think Musk just wanted to make headlines again. But now, looking at the evidence Musk holds and OpenAI’s recent funding moves, I think this isn’t simple.

Musk’s core allegation: OpenAI betrayed its original commitment to be “for the benefit of humanity rather than profit” and became a thoroughly for-profit company. This accusation sounds a bit like taking the moral high ground, but thinking carefully—OpenAI just completed a $122 billion private placement financing, with post-money valuation reaching $852 billion. At this scale, it’s completely not the stance a “non-profit organization” should have.

My personal feeling is that OpenAI is indeed “changing,” and changing fast. From the initial “open source pioneer” to today’s “closed source giant,” this transformation speed is dizzying. Is this right or wrong? I think it’s hard to judge with simple moral standards. Commercial AI companies need money, survival, and competition with Anthropic, Google—these are all real problems.

But the question is: how much of what OpenAI originally promised still remains?

This question, I’m afraid even Altman himself can’t answer clearly.

Back to GPT-6 itself. What I think is most worth watching in this release isn’t the parameter scale, but the potential application-layer transformation it might bring. OpenAI hinted at the launch event that GPT-6 has major breakthroughs in “multi-step reasoning” and “long-term memory”—these two capabilities are precisely the biggest bottlenecks in current AI Agent deployment.

If GPT-6 can truly make breakthroughs in these two directions, then for the entire AI industry, this might be a qualitative change from “chatbots” to “true agents.”

But honestly, my current attitude remains cautiously optimistic. After all, when GPT-5 was released, many promoted capabilities haven’t been fully delivered even now. In the AI industry, the distance between PPT and actual deployment is often farther than imagined.

One final note: I think Musk suing Altman will likely end in some kind of settlement. Both are smart people—taking it to court won’t benefit anyone. But this story’s impact might last a long time—it will make the entire AI industry rethink a question:

Who are we really building AI for?

This question might be more important than GPT-6 itself.