A Year After R1
DeepSeek-R1 was a pricing event disguised as a research event. One year later, the weights are the weapon and the price floor is the wound.
One year ago this month, DeepSeek released R1 and the market decided the story was about reasoning. Nvidia lost half a trillion dollars of market cap in a day, recovered most of it within weeks, and everyone moved on to arguing about distillation. I think the market panicked about the wrong thing, then calmed down about the wrong thing.
R1 was not primarily a research event. It was a pricing event. The reasoning was the headline; the weights were the payload.
The credible threat
Here is the mechanism, and it has nothing to do with benchmarks. A closed lab prices tokens against the next best alternative. As long as every strong model lived behind an API, the “next best alternative” was another API with the same incentives, and prices could settle politely. A downloadable model changes the game even if you never download it. The moment a good-enough open model exists, every closed price sheet is negotiating against your option to leave.
Economists call this a credible threat. You do not have to exercise the option. You just have to be able to. The threat does the work.
That is what R1 actually shipped: not a score, but an option. And through 2025, the option kept getting better. Qwen filled out every size class. GLM and Kimi pushed on agentic work. MiniMax and a half dozen others filled the niches between niches. None of them needed to beat the frontier. They needed to be close enough that a CFO would ask the question, and by mid-2025 every CFO was asking the question.
The floor and the premium
The result is a structure I would draw as a floor and a shrinking shelf above it. Open weights set the floor: this much capability costs this little, forever, because you can hold it. Closed labs live on the shelf above the floor, and the shelf’s height is exactly the premium they can justify.
What can live on that shelf? Not raw capability; the floor eats raw capability on a lag of months. What survives is everything a file of weights cannot give you: distribution, latency guarantees, liability, integration depth, safety posture, and the boring, decisive fact of already being in the procurement system. In other words, the premium is moving from intelligence to serving. The product is not the model. The product is the model showing up reliably inside someone’s workflow with someone to sue if it does not.
I have skin in this framing. My research lives in the unglamorous layer: retrieval that respects a latency budget, updates that do not erase last month, pipelines that correct themselves. A year ago that looked like plumbing. The R1 repricing is why I think the plumbing is the business.
What this means if you build
Three practical consequences, none of which require you to deploy a single open model.
Design for substitution. The option is only credible if switching is cheap. If your prompts, evals, and orchestration are welded to one vendor, you have voluntarily given the premium back. Keep the model behind an interface; keep evals that transfer; know your second source. You will feel this in your renewal negotiation even if you never switch.
Price capability, not prestige. Most requests in a production system do not need the frontier. When we instrumented query difficulty in my retrieval work, nearly half of real traffic was easy, and serving it with maximum effort was pure waste. The same distribution holds at the model level. The floor exists; route to it.
Watch tokens, not benchmarks. Market share of routed tokens is the honest metric of this shift, because it measures what people run, not what people admire. My falsifiable call for 2026: on neutral routers, open-weight models pass half of all tokens served before the year is out, and the closed labs respond not by matching price but by unbundling their own lineups into tiers. Capability where it matters, floor pricing where it does not.
The uncomfortable part
There is a version of this essay that reads as triumphalism about open source. That is not what I am saying. Frontier closed models remain better, and for a meaningful class of problems the gap is worth paying for. The uncomfortable part is for the middle: closed models that are neither frontier nor cheap. The floor rises toward them every quarter, and “slightly better than free” has never been a durable business.
R1’s anniversary is worth marking because it settled a question people were still treating as open in 2024: whether weights would be a research curiosity or an economic force. They are a force. The frontier will keep its shelf. Everyone else is learning to live with the floor.