The Staggered Release Theory: What Matthew Berman Gets Right and Wrong
A popular AI creator says the US government forced OpenAI to stagger GPT-5.6 and handed Anthropic regulatory capture. The story is dramatic, the receipts are thin, and the underlying anxiety about who gets frontier access is worth taking seriously anyway.
Matthew Berman put out two videos this week that are hard to ignore. One titled “I can’t believe this happened…” and the other just “I AM SO MAD.” Three years of optimistic coverage, he says, and this is the first time he’s not feeling it. The claim underneath the emotion: the US government forced OpenAI to stagger the release of GPT-5.6, Anthropic pulled its top model “Fable” after a citizenship restriction, and the net result is regulatory capture that locks builders out of the frontier while the labs keep racing internally.
That’s a big story if true. So let’s separate what’s actually established from what’s narrative, because the gap matters more than the outrage.
What the claims actually are
Berman is stacking several distinct events into one causal chain. Anthropic ran a long stretch of what he calls “fear-based marketing,” emphasizing how dangerous their models are and limiting access to a select group of testers. Anthropic then accused Alibaba of a “distillation attack,” extracting capabilities from Claude. The government, seeing a company say “our models are dangerous” and “they’re being stolen by China” in the same breath, stepped in. Result: Anthropic’s frontier model got hit with a non-US-citizen access restriction and they turned it off rather than comply. And OpenAI, sitting on GPT-5.6 for months, got told to stagger the release.
His conclusion: Dario Amodei won. Anthropic and OpenAI keep their internal frontier models (he names “Fable” and “Mythos”) to build the next generation while everyone else waits for “morsels.”
It’s a clean story. Clean stories about messy industries are usually the part you should check first.
Where the evidence is thin
Here’s the honest problem: almost none of the load-bearing claims come with documentation in these videos. There’s no executive order cited, no agency name, no statute, no letter from OpenAI’s policy team confirming a “stagger” mandate. “The government said you have to stagger the release now” is asserted, not shown. A government directive that materially restricts a commercial product launch would normally leave a paper trail. I’d want to see it before treating it as fact.
The China piece is more grounded but Berman undercuts it himself. He repeats the Anthropic accusation that Alibaba ran a campaign to “brazenly and illicitly extract AI capabilities,” then says the argument is “a little bit flimsy because they didn’t really steal all that much data.” You can’t have it both ways. Either the distillation extraction was serious enough to justify a policy response or it wasn’t substantial. The video treats it as both the trigger for regulation and as overblown, depending on which point needs support in the moment.
And the prompt-hacking observation is correct but proves the opposite of his framing. He’s right that no model is 100% jailbreak-proof. That’s the nature of these systems. But if that’s true, then “China prompt-hacked the model” is not evidence of a security failure that demands lockdown. It’s the baseline reality of deploying any LLM. He uses it to explain why the government overreacted, which actually weakens the case that there was a serious incident worth regulating around.
The model names themselves (“Fable,” “Mythos,” “GPT-5.6 Soul preview”) don’t match any announced products I can verify. That doesn’t make them fake. Internal codenames leak. But when the whole argument rests on specific releases and specific government actions, and none of it is sourced, the appropriate posture is suspended judgment, not anger.
The real anxiety is legitimate
Strip away the unverified specifics and there’s a genuine concern here that I don’t want to dismiss. The concentration of frontier access is real and it’s been trending one direction for two years.
The pattern Berman is reacting to does exist in milder form. Labs do run staged rollouts. Early access goes to enterprise partners and red-teamers before the public gets it. Safety testing periods do delay launches. Anthropic genuinely does argue that powerful models warrant restricted release. Compute and policy both favor the incumbents. A small builder today already waits longer for the best model than a Fortune 500 company with a partnership does.
So the structural worry, that the gap between who builds the frontier and who can use it is widening, is worth taking seriously. The fear-based-marketing critique also lands partly. There is a commercial incentive to talk up danger if danger justifies you being the gatekeeper. That tension is real and underdiscussed.
What I’d push back on is the leap from “this gap exists and might be widening” to “the government just imposed de facto AI regulation that hands one CEO control of the frontier.” The first is an observable trend. The second is a specific factual claim that needs receipts these videos don’t provide.
How to think about gatekeeping without the panic
The useful frame isn’t “regulatory capture has arrived.” It’s “frontier access has always been tiered, and the question is whether the tiers are getting wider or narrower over time.”
On the narrowing side: open-weight models keep closing the gap. The best open model today does what last year’s frontier closed model did. Distillation, the very thing Anthropic is complaining about, is also what makes capability diffuse outward to everyone. If Alibaba can extract capability from Claude, so can the broader ecosystem over a longer horizon. That’s not theft so much as physics of the field.
On the widening side: the very top, the actual frontier, may genuinely pull away if compute and data advantages compound and if safety-justified restrictions become the norm. Both things can be true at once. The frontier pulls ahead while the floor rises behind it.
That’s a more boring story than “Dario won.” It’s also more likely to be the true one.
A builder reading Berman’s videos should do three things. First, don’t make roadmap decisions on unsourced release claims; wait for the executive order, the policy letter, or the on-record confirmation before you believe a government stagger mandate exists. Second, build on the capability tier you can reliably access, which means open weights and the current public frontier, because that’s the layer you control and it’s improving fast regardless of who gets the secret preview. Third, watch the actual policy mechanisms, export controls, model registration, compute thresholds, because those are where real gatekeeping would show up, not in a YouTube monologue. The catch most readers miss: outrage is a content strategy, and the angriest video about access concentration is itself a play for your attention. The concern is real. The receipts are what’s missing.