The day GPT 5.6 launched, I opened Codex to switch the model over. Muscle memory.
When 5.5 came out, it had quietly appeared in the model list overnight. One click and you were using it. The release and availability were practically the same thing. I’d gotten used to the rhythm: the moment the best model dropped, I could put it straight to work on the hardest thing I had.
This time I stared at the dropdown. No 5.6.
It had launched, all right. The specs were clear, three tiers, and the benchmarks were beautiful—the kind that make you want to test them against your hardest real-world task immediately. But you couldn’t select it. Reading further, I got it: this generation was going to only about 20 partners first. After that, every new customer had to be approved—by the U.S. government.
I’ve been on this frontier long enough that “release” meant “available.” This was the first time the two had been pried apart. And what pried them apart wasn’t something you could throw money at.
The Gate Didn’t Close; It Moved
A while back, Fable 5 got shut down worldwide by a single export control order. I wrote about it then, thinking the lesson was clear: geopolitics can cut off a closed-source frontier in one stroke.
5.6 showed me a subtler, nastier version.
Fable 5 was taken away from everyone at once—a clean cut. Painful, but at least fair. 5.6 is “available.” Whether you can use it just depends on whether you’re on the list. The gate didn’t close. It moved.
The old gate was money. The new one is power.
Money Is Actually the Most Reasonable Gate
Let me say a word in defense of money first, because three weeks ago I was still cursing it.
Back then, Fable 5 was pulled from subscriptions and switched to API billing. One task ran me four hundred dollars. I wrote a piece called “Intelligence Is Starting to Divide by Wealth,” arguing that access to the best intelligence had become a purely financial question. I was furious.
Looking back, money is a cold gate, but it has advantages that other gates don’t.
First, it doesn’t care who you are. Nationality, background—none of that matters. If you can pay, the door opens. Nobody sits behind it deciding you don’t belong.
Second, it pushes its own price down. When GPT-4 launched in March 2023, API prices were absurd: 60 out per million tokens. But people still paid, because it was so far ahead there was no substitute. Professionals who needed it absorbed the cost without blinking. That premium was how OpenAI recouped its investment—using the fat profits from its lead to fund the next generation.
And leads expire. Seventeen months later, GPT-4o’s blended price fell to around four dollars—down nearly ninety percent. Expensive is temporary. Eventually everyone can afford it. That’s the norm.
The beauty of the money gate is this: it filters people, but it also takes itself apart. Today’s sky-high price is the down payment on tomorrow’s mass adoption.
The Power Gate Doesn’t Play Fair
The power gate has none of those advantages.
It discriminates. Twenty partners, approved one by one—meaning a few people sit behind that door, reading names off a list, deciding who gets in.
It doesn’t loosen on its own. Money gets beaten down by competition; power tightens with every generation. This generation it’s “20 partners.” Next time, maybe fewer. After that, maybe it stops being released to the public at all, becoming something like new missile technology that ordinary people never touch.
It also completely decouples “can you use it” from “can you afford it and use it well.” You might have the money and the skill, but without the admission slip, you’re already behind before the cards are dealt. A company falls behind not because its tech is worse, but because the approval didn’t come through. That’s the most absurd part.
Ironically, this gate doesn’t even pay off for the people holding the cards. Blocking users worldwide is like smashing with your own hands the lead premium that money had secured. If users can’t access it, no one validates your lead, no one pays for it, and the competition closes in fast. Once they’ve caught up, you can’t command the kind of premium GPT-4 did. The power gate devours the very economic engine that nurtures the next generation of models.
As for how this happened, my read is: the hype outran the real-world testing. The model was hyped so hard that policymakers started treating it like a weapon, and most of them had never actually used it—they’d only heard about it. The “too dangerous to release” narrative isn’t new. In 2019, OpenAI sat on the full GPT-2 model for over half a year. That same script has been playing out ever since.
I’m not saying safety shouldn’t be managed. It should. But a blanket ban by nationality is the laziest, most parochial move in the toolbox. Besides, given how fast things diffuse now, it probably won’t stop anything anyway.
A License That Changed, Then Was Forced Back
Speaking of which, I’m reminded of a story that has nothing to do with models but makes the exact same point.
Redis. The database almost every backend engineer has used, open-source for over a decade under the BSD license. Anyone could run it, modify it, build services on it, and sell them. In 2024, Redis suddenly changed its license to one that no longer counted as open source. In plain English: before, you could use it freely; now, if you want to use it commercially, you need my permission first.
This was essentially swapping the money gate for the power gate. What had been a market transaction became, overnight, “you must first obtain permission.”
But what happened next is the interesting part. Because Redis had been open source before the change, that last open-source version was sitting right there in plain sight. Anyone could continue building on it. Within weeks, the Linux Foundation took the lead, backed by AWS, Google, and Oracle, and forked Valkey directly from that version under the same open license. How fierce was the fork? Fierce enough that Redis itself couldn’t take the pressure, and in 2025 it changed the license back.
The whole point is this: you can’t truly take back something that’s already open source. The exit is always there.
The model world has the exact same structure. A closed-source frontier model is a switch you can’t reach. You build your whole business and workflow on it, and one “compliance requirement” can shut you down—with the government now holding that hand too. An open-weight model is something you can fork, deploy, and run on your own hardware. Slower, a tier behind—fine. The point is that this supply is in your own hands, not floating out of reach.
The hedge against the power gate was never more money. It’s open source.
Final Thoughts
That open weights can still keep pace with the frontier today depends on having a catfish in the tank.
DeepSeek is that catfish. When R1 dropped in January 2025 under an MIT license, completely free, it wiped out more than $589 billion of NVIDIA’s market cap in a single day—a record drop. It’s strong enough, and commercially weak enough, to open-source a version every so often, forcing everyone toward openness. To the model world, it is what Valkey is to Redis—the thing that keeps you from comfortably exploiting users.
My only worry is whether this catfish might get caged one day.
Open source is not permanent insurance. OpenAI started with “Open” too, going from sitting on GPT-2, to GPT-3 via API only, step by step becoming the most closed shop in the business today. An open model company has incentives to turn away, given enough time. While the catfish is there, the water stays alive. If it disappears, the pool quickly becomes dangerous.
So what this piece is trying to say is simple: don’t bet what matters most on a switch you can’t reach.
The best thing about this era was that the most cutting-edge intelligence could reach ordinary people early, and relatively fairly. Money is a gate too, but it’s a key anyone can get by saving enough, and it conveniently pays the down payment for the next generation’s adoption. The power gate is different: it sends out invitations, the list is held by very few people, and it only gets shorter.
Three weeks ago I was still complaining that intelligence was starting to be about money. Now I think: intelligence you can buy with money is already the relatively fair kind. What really keeps you up at night is the intelligence you can’t buy even with money—the kind you need someone’s nod to use.
The more intelligence is treated as power, the more precious is the bit you can hold in your own hands.
References
- OpenAI, "Better Language Models and Their Implications" (Announcement on GPT-2’s staged release and withholding the full 1.5B model for over half a year under the pretext of "preventing misuse"), 2019-02, https://openai.com/index/better-language-models/
- The Decoder, "From GPT-2 to Claude Mythos: The return of AI models deemed 'too dangerous to release'" (The "too dangerous to release" narrative from GPT-2 to today), https://the-decoder.com/from-gpt-2-to-claude-mythos-the-return-of-ai-models-deemed-too-dangerous-to-release/
- OpenAI, "GPT-4" (March 2023 release, API pricing at 60 out per million tokens), https://openai.com/index/gpt-4/
- Andrew Ng on X (GPT-4 launched at roughly 4/M), 2024-08, https://x.com/AndrewYNg/status/1829190549842321758
- Wikipedia, "DeepSeek" (R1 released on 2025-01-20, MIT license, free and open), https://en.wikipedia.org/wiki/DeepSeek
- Yahoo Finance, "Nvidia stock plummets, loses record 589 billion wiped out in a single day, the largest single-day drop in history), 2025-01, https://finance.yahoo.com/news/nvidia-stock-plummets-loses-record-589-billion-as-deepseek-prompts-questions-over-ai-spending-135105824.html
- CNN Business, "China's DeepSeek shook the tech world. Its developer just revealed the cost of training the AI model" (V3 training cost roughly 294,000), 2025-09-19, https://www.cnn.com/2025/09/19/business/deepseek-ai-training-cost-china-intl
- InfoQ, "Redis Returns to Open Source under AGPL License: Is It Too Late?" (The story of the 2024 SSPL change → Linux Foundation forks Valkey → 2025 reversion to AGPL), 2025-05, https://www.infoq.com/news/2025/05/redis-agpl-license/
- HashiCorp, "HashiCorp adopts Business Source License" (August 2023: Terraform and others switch to BSL); OpenTofu, "OpenTofu Announces Fork of Terraform" (Community forks under the Linux Foundation), https://opentofu.org/blog/opentofu-announces-fork-of-terraform/