With every day that passes, another model joins the ranks at the frontier. Last week it was xAI's Grok 4.6. This week it's GLM 5.3. A month ago it was Kimi K3. And it clearly won't be long until Qwen and MiniMax and Meta's Muse Spark join them.
It turns out the frontier was temporal. It was never about how good you are - all these labs are quite good - it was about when you got there. A compounding advantage of being on the frontier works equally for everyone on the frontier.
Does that mean there won't be any leaps forward? No. But it does change the economics.
On the 27th of August, Z.ai will release the weights for GLM 5.3. MIT license. Anyone can use them, at no cost. Frontier-level weights. They might be just a touch underneath Claude Opus 5 and GPT-5.6 Sol. But it's not clear to me that anyone using these models for most things will ever notice the difference.
Which means the frontier is now really powerful and basically free.
So how do Anthropic and OpenAI compete with free? That question is going to be asked by every investor on their IPO roadshows. Given that Anthropic now expects to list at somewhere around $2 trillion, there's enough money on the table to make people extremely focused on the answer.
We've run the compete-with-free experiment before. Linux took the operating system layer, and the money moved up the stack: Red Hat sold assurance, the clouds sold operations, and the kernel itself became something everyone used and no one paid for. Free won the layer; the business moved to what sat on top of the layer. Weights are the new kernel. The labs' problem is that they're currently priced like they own the kernel.
Anthropic will promise investors that it will leap ahead, once again, as it did with Mythos. But as we saw with Mythos, that presents problems of its own. We're at a delicate point: the models are so powerful that making them more powerful isn't necessarily the safest thing for Anthropic's customers.
There is a sweet spot for "good enough" models - smart enough to do great work, not smart enough to go rogue - and the line between those sits somewhere around where the frontier is right now. Beyond this point, models will be carefully controlled, possibly even licensed. A whole new level of regulation kicks in if you want to use the best of the best. Think of the top-of-class frontier model as a luxury good - one you'll need to be carefully vetted before you're granted the privilege of paying for. Most organisations, outside of cybersecurity, will never need that capability, and won't see the point in the paperwork.
Yet that regulation may be the only durable moat the frontier labs have left. If the top tier of capability becomes a vetted, controlled substance, then the accredited lab with the solid safety record isn't being constrained by the regime - it's being deputised by it. Free weights can be brilliant. What free weights cannot be is accountable: nobody stands behind them, nobody carries liability for them, nobody vets who runs them.
The labs are already moving to harnesses. Claude Code earns its place as the first harness worthy of the name; it launched Anthropic on its current rocket trajectory and kickstarted the agentic revolution. OpenAI's Codex is also excellent. Study after study shows that a great model in a great harness is an incredibly powerful tool.
Is that enough? It feels like a narrow basis for a multi-trillion-dollar enterprise - not least because harnesses have their own problem: everything a harness does eventually gets absorbed into the models themselves (the 'Bitter Lesson'), and the models are becoming free. The migration doesn't stop at the harness. It moves through it, toward the things that can't be copied at all: trust, accountability, the standing to be vetted. The labs that thrive will be the ones that notice where the migration ends.
It could well come to pass that the Chinese models are simply banned in the United States, through some combination of security concerns and mercantilism. That's the US. The rest of the world won't care, and all of that market gravitates to the free models. That shift is well underway.
There doesn't seem to be an obvious path out. If that's right, there's about to be some carnage in the markets. Not among the hyperscalers - free weights at the frontier only increase the superexponential demand for compute - but among the model providers themselves.
A few months back I suggested to John Allsopp that Anthropic may end up acquired by one of the tech giants. Microsoft certainly needs that kind of capability. So does Amazon. Maybe Google too. The same logic holds for OpenAI.
It's not that these firms don't make great models. It's that making a great model is table stakes. That's become obvious at this very crowded frontier. The game that matters now is the one free weights can't play: being the name on the contract that wears the product liability.