The number that reframes this entire debate is simple. In August, open-weight models ran most of the tokens flowing through one of the industry's busiest routing layers but collected barely a seventh of the money. My read is that most coverage has the emphasis backwards. The interesting development is not that open models are winning, and not that closed labs are doomed. It is that the market has quietly split into two different businesses that happen to share the word "AI," and the economics of each are diverging fast. If you want to understand where value accrues over the next year, that split is where to look.
The volume story is real, and it is dramatic
Start with what actually happened, because the pace is easy to underweight. Open-weight models grew from 7% of token volume in December to 56% in August on Vercel's AI Gateway, a platform that routes tens of trillions of tokens a month across apps built on its infrastructure. Their share rose from 7% in December 2025 to 13% by April, then to 36% in July, before crossing 50% in August. On a single day in late September, the CEO flagged a reading above 78%.
This is not confined to one gateway. Open-weight models accounted for 60% of OpenRouter's US token consumption in August, with Chinese-developed models making up the majority of that volume. The engine here matters. The surge is substantially a Chinese open-weight wave, with DeepSeek, Moonshot's Kimi line and Z.ai's GLM family supplying a large chunk of the raw tokens. These are not toy models. The best open variants now rival proprietary systems on coding and a lot of everyday reasoning, and the headline benchmark gap that stood at double digits a year ago has compressed to almost nothing on general-knowledge tests.
So on the measure of "what is actually running in production," open weights have won. If the story ended there, the closed labs would be in serious trouble.
The money has barely moved
It does not end there. Despite this token-share dominance, open-weight models captured only 14 cents of every dollar spent, while Anthropic's models took 64 cents of every dollar despite far lower token share, a figure that has never dropped below 61% since December 2025. The reason is structural. Open weights are cheap to serve because any cloud operator can host them, which triggers price competition on inference. Closed frontier models stay expensive because the lab is the only seller of that specific intelligence. Anthropic's models carry premiums of up to 4.4 times the average token price, and buyers keep paying it for the work where correctness is worth more than cost, chiefly coding agents and other high-stakes automation.
The cleanest way to see the divide is a pair of lines from the same leaderboard. One model runs 59.3% of the tokens and collects 5.1% of the spend. Another runs 1.7% of the tokens and collects 13.7% of the spend. The first is a cheap open-weight workhorse running vast volumes of routine work. The second is a premium closed model being used sparingly for tasks where the output justifies the price. Both are rational customer choices. Neither tells you who is winning on its own.
There is a measurement caveat worth stating plainly. Token share counts everything, including long contexts, cached reads and the heavy reasoning chains that agentic work produces, so it can be inflated by a handful of computationally hungry jobs. Spend reflects list prices rather than business value delivered. A single 78% day is noise. The monthly trend is signal, and the signal is a widening gap between volume and dollars.
The closed labs are cutting prices to defend the floor
The closed incumbents are not standing still, and this is where the price war turns genuinely aggressive. Anthropic's cheapest tier, Haiku 5.5, launched at $0.10 per million input tokens, which the company framed as roughly a 75% cut on the prior Haiku and up to 90% cheaper on short requests. That figure matched OpenAI's comparable small model to the penny, which tells you these cuts are competitive responses rather than acts of generosity.
OpenAI pushed further at the frontier. Its GPT-6 Sol and Luna models arrived both 50 percent cheaper than their previous iterations, with GPT-6 Sol 50 percent cheaper than Anthropic's Claude 5.5 Opus on input and output prices, while GPT-6 Luna is cheaper than even the off-peak API pricing for the DeepSeek V4.1 Flash model. Read that last clause twice. A closed US lab priced its small model below a Chinese open-weight model that was built partly to undercut on price. The whole justification for reaching for open weights was cost, and the incumbents are now attacking that justification directly.
Underneath the headline cuts is a quieter shift in how the labs want you to measure them. Rather than defend price per token, which they will lose, they are reframing the comparison as price per completed task. A pricier model that burns fewer tokens to finish a job, or gets it right the first time, can be cheaper on the work that matters even at a higher sticker rate. That is a defensible argument for high-value workloads. It does nothing for the enormous low-stakes volume where open weights are plainly good enough.
Adoption is messier than the volume charts suggest
Here is the part that complicates any simple "open weights are eating enterprise" narrative, and where I part ways with the more triumphant takes. Raw gateway volume and actual enterprise commitment are not the same thing. Menlo Ventures found enterprise open-source model usage fell from 19% of enterprise usage in 2024 to 11% in 2025, even as the models got better. The technology improved and a meaningful slice of serious buyers pulled back.
The explanation is mundane and important. Self-hosting an open-weight model is not automatically cheaper, because you swap a per-token price for GPU and operations cost, and running it well at scale is real engineering work. Data sovereignty rules keep Chinese weights out of US defence and government entirely, and out of many regulated workflows unless they run on the buyer's own infrastructure. And within the open camp the leaderboard reshuffles constantly: Google's Gemma has overtaken Meta's Llama as the most widely deployed open-weight model family, with production usage rising from 35% to 55% since July, which is its own warning against betting on any single model's permanence.
What it means
The market read is a barbell, and it is hardening. Routine, high-throughput, low-stakes work is commoditising toward the open-weight floor, where margins are thin and nobody has pricing power. High-stakes work where errors are expensive stays with premium closed models that can hold a 4x price premium because the buyer is paying for reliability, not tokens. The squeeze falls on anything stuck in the middle: a merely-adequate closed model with no distribution and no price advantage has nowhere to stand.
Follow the profit. If intelligence commoditises while usage keeps climbing, the durable margin migrates away from selling tokens and toward the layers that are scarce: the chips, the infrastructure that serves inference at scale, and the application and routing software that decides which model handles which request. The most valuable new skill for a buyer is not picking a model but building a model-swappable system that sends each task to the cheapest option that clears the quality bar. That capability, not loyalty to any one lab, is where the savings actually land.
What I would do depends on where you sit. If you are deploying, invest in routing and evaluation before you invest in any single model, because the relative value of each option is changing month to month and the ability to switch cheaply is the real asset. If you are weighing self-hosting, cost it honestly against managed inference, since the per-token saving often evaporates once GPU and ops time are counted, and only pays off at genuinely large and steady volume. If you are thinking about where equity value accrues, the token-selling business looks structurally harder than it did a year ago, and the infrastructure and tooling beneath it looks structurally better.
The tidy framing of open versus closed misses what the data is showing. Both are growing, and they are growing into different jobs. The question worth sitting with is not which side wins but how much a token of intelligence will be worth once almost anyone can serve one. On current evidence, the answer is: a lot less than the people who built it were hoping.