The $1/M Frontier: What Comes After the AI Benchmark Race | Vertical
How it works<br>Integrations<br>Pricing<br>Free Tools<br>Get startedLog in
0% read
← All articlesAs I'm building Vertical, I've been thinking a lot about the future of AI. Singularity, open weights, frontier intelligence. In this article, I actually argue that the Innovator's Dilemma will soon start applying to the current models. We'll soon be caring about price and speed a lot more than raw intelligence.
When I think about AI, there are two workflows that are important for me. Productivity and application. By productivity, this is all the AI that I use for myself and to run Vertical. At the moment, my productivity is heavily subsidised by OpenAI and Anthropic, so these are the primary models that I use. By application, I mean the models that I actually use to provide Vertical as a product.
Choosing a model at the application layer is much harder than choosing at the productivity layer. Cost and latency are much more important here (unfortunately, OpenAI and Anthropic don't subsidise me, and I'm bootstrapped so no VC subsidies either). There are three factors:
Model performance : Can the model complete the task?
Cost : What's the blended cost of running this model?
Speed : Is the latency good enough for the user experience that I want to deliver?
Performance Is Already Good Enough
I was happily hacking a couple weeks ago, when Opus 5 dropped (but to my great disappointment, with no subscription reset). Later that day, a friend of mine asked me how it was. I'd had it for maybe half a day at that point, on pretty ordinary work. I told him it was good. Then I admitted I couldn't point to anything from that day in which Opus 5 felt that different from Sonnet 5.
I'm not saying that is always the case, but more and more, I feel how I use the model makes much more difference than which model I use. There are truly hard tasks in which a smarter model genuinely makes a difference, but it's becoming less and less frequent.
This is basically the Innovator's Dilemma at its finest: incumbents keep improving the thing their best customers pay for, until the product is better than what most of the market needs. At that point people stop paying for more of it, and the competition moves to price and speed. It happened to disk drives. I think it's starting to happen to intelligence.
We're starting to see it with Kimi K3 (and Chinese models more generally). Kimi is able to fight with Fable, while being significantly cheaper.
To me, things will get really interesting when we start seeing Opus 5 intelligence at the $1/M blended cost.
Why $1/M? Well, that's actually from an OpenRouter analysis of the models vs their prices.
First I pulled their public models API and computed a blended cost (assuming 3 input tokens for every output token) for the 253 models from the major labs. The median came out at $1.00 per million tokens , sitting exactly on the line.
But list prices only tell you what's on the menu, not what people order. OpenRouter also publishes rankings data with actual weekly token volume per model. I joined the last 12 weeks of volume to current prices, and the result surprised me.
I expected cheap to win. It doesn't, at least not cleanly. The two biggest models by volume are ultra cheap: DeepSeek V4 Flash and Xiaomi MiMo, both at $0.18/M. But Claude Opus 4.7, at $10/M, is still #6 by volume, with Opus 4.8 and Sonnet 4.6 close behind it. The models that actually lose are the ones in the middle. The $1-5/M tier gets the least volume of all.
Crowded under $1/M, crowded again above $5/M, nearly empty in between.
The Speed Frontier
But price is not the end of the story. The Innovator's Dilemma will not just apply to price, but also to speed. Every time one axis gets over-served, the competition moves to the next one:
Intelligence : over-served for most tasks. This is the wave that's ending.
Price : the wave we're in right now. That's what the chart above shows.
Speed : the wave nobody is seriously fighting yet.
Today, speed is expensive. Serving tokens faster to one user means batching fewer users on the same hardware. SemiAnalysis measured this tradeoff: generating roughly 2.5x faster costs about 7x more per token. Right now you can have fast or you can have cheap, but not both.
That's exactly why I think speed comes third. Once models hit the $1/M frontier, price will become good enough for a lot of applications, the same way intelligence already did for most tasks. Whatever breaks the speed-cost tradeoff (better hardware, better serving, or routing systems that only go fast where it matters) starts the next wave.
And that wave is the one I'm most excited about. Imagine if we could get millisecond latency and 10k/s tokens on AI responses. Both our application layer and productivity layer would change so drastically. No more tab switching waiting for AI to complete its work. Fast user experiences at the application layer. And more speed...