The Four-Month Moat Is Not a Moat

The capability gap between frontier AI and open alternatives has narrowed to 4.4 months. That is the finding Mozilla dropped this week, and it is the most consequential number in AI right now.

4.4 months. That is shorter than a product roadmap cycle. Shorter, in many cases, than the procurement process required to deploy the model you just paid a premium for. The window between “cutting-edge” and “commoditized” has compressed to the point where the word “frontier” is doing almost no work.

Mozilla’s report — previewed by Ars Technica — compared the capabilities of open Chinese models against Silicon Valley’s paid frontier offerings. The gap has narrowed sharply, but it has not closed. Open models still trail on the most demanding tasks — Mozilla’s own CTO flags expert professional work, high-intensity retrieval, and long context as places where closed models still earn their premium. Outside those categories, open models run at roughly 30 percent of the cost of closed frontier alternatives, or about 3x cheaper. In one specific benchmark — Terminal-Bench 2.1, comparing GLM against Claude Opus — open models came in at one-fifth the cost per completed task. Those are different numbers for different claims, and conflating them is how bad procurement decisions get made.

The math is still brutal, even with the right numbers. A 3x cost premium for a 4.4-month capability lead that evaporates on the most common workloads — that is not a moat. It is a window. And windows close.

Everyone says the frontier labs have a durable lead because of compute, data, and talent density. The opposite is closer to true, and has been for eighteen months. The lead exists — but it is measured in weeks to months, not years, and it is being eroded by open-weight releases, Chinese labs running at a fraction of the cost, and distillation techniques that let anyone compress a frontier model’s behavior into something much cheaper to run. The moat is real; it just drains faster than the labs are filling it.

This reframes the Anthropic-and-OpenAI-lobbying-Washington story running in parallel today. The Register reports that both labs are angling to be designated critical AI infrastructure — the tech equivalent of too-big-to-fail. Read that through the Mozilla lens and the strategy becomes obvious: if your commercial moat is 4.4 months wide and shrinking, you need a regulatory moat to compensate. Get Washington to certify you as a national security asset, and suddenly the open Chinese models aren’t a cheaper competitor — they’re a threat vector. Procurement rules, export controls, and security clearances do the work that capability advantages no longer can.

That is a rational play. It is also exactly what incumbents do when the product stops differentiating itself. IBM did it with mainframes. Boeing does it with defense contracts. The question is whether Washington will oblige — and whether it should. The steelman for the labs is real: concentrated, auditable AI infrastructure is genuinely easier to secure than a fragmented open ecosystem. If a state actor is going to compromise AI systems, you would rather it happen at one of two companies with federal oversight than at ten thousand self-hosted deployments with none. That argument has teeth.

But the steelman breaks down here: the labs are not proposing to be regulated like utilities. They are proposing to be protected like utilities while being priced like luxury goods. The goal is captured status, not increased accountability. If Anthropic and OpenAI actually wanted the obligations that come with critical infrastructure designation, they would be publishing red-team results, opening their weights to federal auditors, and accepting price caps on government contracts. They are not doing any of those things. They are doing the part where Washington certifies them and skipping the part where Washington controls them.

The Mozilla data makes this a live question rather than a hypothetical. If open models are 4.4 months behind at roughly a third of the cost — and at one-fifth the cost on specific benchmark tasks — then the government can run capable AI on open infrastructure without depending on either lab for most workloads. The national security argument for mandatory reliance on OpenAI or Anthropic weakens considerably when the alternative is not “use a Chinese model” but “use a domestically fine-tuned open model at commodity pricing.” The labs need Washington to not notice this. Mozilla just made sure they will.

The stakes here are not abstract. If the lobbying works, the next decade of AI procurement in government, defense, healthcare, and education flows through two private companies with a 4.4-month capability lead and a regulatory designation that insulates them from competition. If it fails — or if open models continue to close the gap — the architecture of AI deployment looks much more like the internet: messy, distributed, and ultimately harder to corner.

A moat you have to beg Congress to dig for you is a political arrangement, not a competitive one — and political arrangements have a habit of reversing.