DeepSeek's v4.1 Flash now captures 78.4% of token volume while frontier models like Anthropic's still get most dollar spend. Robert Ta and Jonathan argue builders should use open-source models for most tasks, reserving expensive frontier models for high-intelligence work that needs them.
On Vercel AI Gateway, OpenWeight Models hit a record 78.4% token volume on September 18th, while roughly 64% of the dollars still went to Anthropic. DeepSeek shipped version 4.1 Flash on September 10th. If you take a look at this dollar spend, says that number 3 and 4 on the AI leaderboards here are Moonshot AI and DeepSeek. If you add in ZAI, Their combined dollar spend surpasses number 2, which is OpenAI. That's pretty crazy. And it shifts a slight inflection point that I've been hearing from the All-In podcast hosts where they keep saying like, dude, next thing is just going to be open source models fine-tuned for your use cases with domain-specific harnesses. And then shift and allocate a certain amount of compute to the frontier intelligence where you need it. But for everything else, just do this. And here's some data to back that claim. And you could go to the AI model leaderboards here on vercel.com. You can actually see the top models, labs, apps, and providers on AI in terms of usage, at least on the Vercel platform. If you trust their data and all that, you know, there's other platforms as well, but this is a good sampling, I'd say. Jonathan, from your CTO perspective, what are your thoughts on this open weight, open source versus frontier sort of battle that's going down in this inflection point in the data we're seeing?
You're right. The data point you mentioned from Olin, you know, I think we're going to be using routers and routing tokens to open source models that will just be part of our stack and set of tools that we use. That seems pretty inevitable. The lead that they kind of buried here was that the DeepSeek V4.1 launch this week seems to be very popular. It pretty much took over the chart this week. So you see less of the other tokens being used and then like more than ever before tokens being used with DeepSeek V4.1 Flash. One thing it's good for, cost is low, the caching performance is very good. And so what it allows you to do is run these long-running tasks like a goal loop or a long-running coding agent at fractions of cost of what you pay OpenAI or Anthropic.
Yeah.
And so I think you got companies, builders like us out there testing and just like saying, let's just set the agents on autonomous and have them build like infinite software now. I think that was the unlock this week, that the open source model was good enough to try some of these experience experiments. The Vercel data probably isn't representative across the board of builders, but it is a specific type of builder and kind of tools they're picking up and using. And I think that this is a like signal to latch onto. That's, uh, really interesting.
Yeah. Shout out to Vercel for making this data public. AWS Bedrock, Google App Platform. Azure, please do the same. Render, go ahead. I'd love to see all that data. It is funny what Jonathan says here. If you look at this chart, this is token volume, share of AI Gateway token volume across the top models on Vercel AI Gateway from June 22nd, 2026 to September 19th, 2026. That is the past 3 months. For the past 3 months, just to clarify what Jonathan said, which is this model just came out this week and already it's dominated 64.8% of usage for the past 3 months. And Jonathan's saying from his engineering AI builder perspective, that just means that, hey, this is maybe an inflection point that folks have figured out and trusted the competence of DeepSeek v4.1 Flash in particular on this chart to run longer horizon autonomous goal loops as opposed to just running them on Frontier Intelligence. And that could be the reason for this increase, drastic exponential increase in usage here. We could be looking back at this in the next little 3 months and be like, oh my gosh, like this was the start of when the meta became, oh yeah, just run all your long-ranging shit on, on the open weight models with your own harnesses, guys. That's just what we do now. That's how we roll. For folks out there that are tinkering and trying to get the edge on the competition, that might be something to look forward to, is more open weight model usage and tuned with your own domain-specific harnesses on long horizon tasks for your own problems.
