Did GPT-6 Get 'Released' Again Today?

A couple of days ago, I saw a friend post on social media: "GPT-6 is released," followed by "hahaha."

This person isn't the type to spread random rumors—they're in the industry. My first thought was: "How did I miss something this big?"

I casually Googled it. Found nothing. Then I had Claude Code look into it, and it rapidly spat out a bunch of results about "what time the launch event started," "what the parameter count is," "how much it improved over GPT-5"—the details were surprisingly consistent. But OpenAI's official site hadn't changed at all.

Then it clicked. April 14 was one of the rumored release dates circulating in English-language circles. OpenAI had just finished pre-training the codenamed "Spud" model at the Stargate data center on March 24, and Altman had said "releasing in a few weeks." Then everyone started guessing which day marked the end of "a few weeks." April 14 got picked. Within days, a circle of people wrote posts, added images, made up parameter counts, and cross-referenced each other, stitching together a seemingly coherent "fact." The Chinese-language versions were even more absurd—"GPT-6 releasing next week," "releasing next month," "releasing tomorrow"—it never stopped since last fall, and each round harvested a new wave of shares.

What really bothered me wasn't the fake news itself—it was how the more I thought about it, the more tangled it got.

Misinformation Actually Has an Audience

Back when Zhihu (China's Quora equivalent) was first blowing up, and again during the WeChat misinformation wave of the pandemic, there was a flood of false information. The approach back then was debunking. Guokr's "Rumor Crusher" (a Chinese science media outlet) had been doing this since 2010—503 articles in three years. They helped shut down things like the "earthquake life triangle" and "nuclear contamination spread maps." The playbook was: rumor appears → professionals dismantle it → everyone has an aha moment.

But that path seems to be getting narrower and narrower.

Because misinformation isn't a "mistake that needs correcting"—it's an emotional consumable.

Here's an analogy. When we chat, besides discussing serious matters, we also shoot the breeze. We might be talking about something substantive, and then I suddenly think of a tangential point and say something wildly exaggerated. The other person won't jump in and say "that's factually incorrect"—they'll laugh. The moment gets consumed, and nobody is harmed.

A lot of "misinformation" online plays exactly this role. Its goal isn't to be "believed" at all—its goal is to get you to click, share, or drop a comment. The harder you try to verify or debunk it, the more the algorithm sees engagement and pushes it to the next person.

Once you see this, a lot of confusing things start to make sense.

Why GPT-6 can be "released" so many times—each "release" harvests a round of traffic. The more you try to debunk it, the more you feed it.

Why those unsourced medical posts on Xiaohongshu (RED) blow up—a few images, a few lines of text, making a radical claim about some university's discovery or how some hormone actually works. The discussions below all unfold on the implicit assumption that "this is fact": "No wonder," "So that's how it is." I've read dozens of these recently; not a single one had a credible source, yet the comment sections were buzzing.

And then there was that little horse a while back. An anonymous model called HappyHorse knocked ByteDance's SEEDANCE 2.0 off the Artificial Analysis leaderboard. Overnight, happyhorse.io and happyhorse.com were domain-squatted, and a bunch of empty HuggingFace repos popped up claiming "open source" and "number one." I had Claude Code investigate its background. The first time, it actually fell for it, citing those bandwagon-jumping sources as authoritative. I ran it again with a different agent, and only then got a solid conclusion: "Sources remain unclear; recommend waiting for official confirmation."

This shook me a bit. Not because AI made a mistake—mistakes are acceptable—but because my information pipeline is fundamentally insecure.

Real Information Is Actually a Niche Market

Coming full circle, I finally see it clearly: real information is expensive.

The production side is easy to understand. Analysis, research, experiments, repeated verification—every step burns time.

Consumption is expensive too. You have to spend energy understanding convoluted phrasing, drawing conclusions that are plain and devoid of emotional payoff. Real information doesn't give you that instant dopamine hit of "oh, so that's how it is." More often it gives you "this isn't that interesting" or "I need to think about this more."

Expensive on both ends, real information is naturally a niche market. It isn't being suppressed; the supply and demand curves just look like this.

I didn't have this insight before. I always thought the problem was "how do we debunk rumors." Now I think the problem is: "In an environment where lies are cheaper and more consumable than truth, how do I preserve my own judgment?"

AI Is Making This Worse, Not Better

At first I also thought AI was the solution. Let AI help you verify, screen, and analyze.

The more I use it, the less I believe that.

AI is like an extremely capable employee. It can get work done, and do it well. But if every time you only look at the conclusion without looking at the process, you fall into a dangerous state: there is an extra curtain between you and the truth.

I've discussed this before: creation itself is a method of approximating truth. When an experiment fails, you judge where to adjust next based on what the failure looked like. When a plan doesn't work, you learn new things from the reasons it didn't work. None of this can be learned by only looking at conclusions.

And humans are naturally lazy. If AI does seven hours of an eight-hour workday and you only look at conclusions, you'll feel highly productive—but your judgment is quietly deteriorating. This is what worries me most. AI appears to amplify your output while simultaneously weakening your judgment. Fragmentation and short-form videos were already doing this; AI is now pouring more fuel on the fire.

In the context of daily work, this comes down to tool choice.

So I Pushed the Team Back to Claude Code

Over the past month or two, I've been comparing Codex and Claude Code.

At first I recommended the team switch to Codex. The reason was straightforward: it's stronger. Codex can crack hard problems that Claude Code can't. GPT-5.4 is clearly a step up in complex architecture and long-chain reasoning. On benchmarks, it scores more than ten points higher than Opus 4.6 on SWE-Bench Pro. When Opus grinds away at something for a long time without success, Codex often nails it on the first try. I was genuinely impressed at the time.

But recently I changed my mind. Not because Codex is bad—we're still using it for what it can do. It's because I realized that for most people, using Codex exclusively traps you in a state of "AI does it, I don't grow."

Claude Code's process is transparent. It actively makes plans, asks you clarifying questions, and writes out clearly what it's doing at every step, in plain human language. I'm not the foremost expert in every domain, but I can participate in the discussion, follow the reasoning, and add challenges or caveats—sometimes the angles it throws out spark ideas I wouldn't have had otherwise. After a month, I'm starting to develop judgment in several subfields I previously knew nothing about.

Codex is different. Its thought process is more machine-language-like, a jumble of weird tags that's hard to decipher. When it's done, it dumps a dense wall of summary on you. You ask it to explain, and it dumps another dense wall—still hard to parse. I actually burn more tokens with Codex than with Claude Code—it naturally takes more because it tackles harder tasks—but I grow less. I only know "it got it done" or "it didn't," not how it got there.

A month isn't long, but projected over a year, this gap becomes significant.

A Friend Asked Which One I Recommend

Claude Code.

Not because it has the strongest benchmarks (it doesn't). Not because it's the cheapest (it isn't either). It's because "making you stronger" is built into the product. Codex is like a colleague who silently works overtime and gets it done. Claude Code is like a colleague who walks you through the code. The former is faster; after a year with the latter, you are two people.

Back to that social media post at the beginning.

In the age of AI, information is getting cheaper, but people are becoming more valuable. AI getting stronger doesn't automatically solve the problem—those of us using AI have to find our own ways to preserve judgment in this environment.

So my selection criteria have changed too. I don't just look at how much AI can help me get done; I also look at whether I understand a little more after a day of collaboration. These are two different things, but I didn't used to separate them.