Open Source Is Winning the Token War. Anthropic Is Still Winning the Money War.

Here's a stat that should terrify Anthropic: DeepSeek just took the lead in token volume on Vercel's AI gateway, processing over a third of everything flowing through the platform. Here's the stat that explains why it doesn't: Anthropic still accounts for more than half of the actual money spent there.

That gap between tokens and dollars is the most interesting number in AI right now, and TechCrunch's Russell Brandom dug into it today, building on a theory Decagon CEO Jesse Zhang floated in a post titled "Everyone is wrong about open source AI in the enterprise."

Zhang's argument, drawn partly from his own company's behavior, goes like this: mature AI deployments are switching to lighter, cheaper models. His firm is doing it. Everyone is doing it. And yet total spend on expensive frontier models has barely moved. His explanation is that frontier and open source models aren't actually competing. They're two phases of the same life cycle. Companies use expensive state-of-the-art models to figure out whether a use case works at all, then hand the proven workload down to a cheap open alternative once it matures. As Zhang puts it, "The frontier labs will keep owning discovery. Open source will increasingly own production."

The numbers back him up, mostly

Zhang didn't bring much data, but as Brandom notes, it isn't hard to find. On Vercel's gateway leaderboard, DeepSeek leads token volume and Z.ai, the lab behind GLM-5.2, has climbed to fourth. Scroll down to spend, though, and Anthropic holds more than half of it, a share that has slipped only slightly over the past month, and mostly because of Anthropic's own price increases rather than customer flight.

OpenRouter tells the same story at larger scale. DeepSeek V4 Flash processes 5.3 trillion tokens a week, more than double the 2 trillion handled by Opus 4.8, the most popular frontier model. But Opus tokens cost roughly 23 times more, about $1.37 per million against 6 cents, which means Anthropic is almost certainly still capturing the bulk of the spending despite moving a fraction of the volume.

The frontier labs, in other words, have ceded the commodity shelf and kept the premium aisle. Volume is not the business. Margin is the business.

The "yet" is doing a lot of work

Brandom's headline hedges with a "yet," and the hedge is earned. A CNBC report published the same day shows what the pressure looks like up close. Vercel told CNBC that GLM-5.2 saw the fastest adoption of any model the platform has tracked this year, with daily token volume growing roughly 27x in its first full week. OpenRouter says open Chinese models can run 60% to 90% cheaper than the leading Anthropic and OpenAI options. And AI startup Lindy moved 100% of its traffic from Claude to DeepSeek in June, a switch its CEO says will save millions within months. He described watching the cost curve crash once the migration went through.

None of that contradicts Zhang's life-cycle theory. It confirms it. Lindy is exactly the kind of mature deployment that graduates from discovery pricing to production pricing. The open question is what happens to the top of the funnel. The two-tier economy only stays stable if new frontier-worthy use cases keep appearing as fast as old ones graduate downward. So far they have. AI-addressable work is expanding quickly enough that the frontier labs can hold their revenue just by dominating the early, hard, expensive phase of every deployment. If that expansion ever slows, the graduation pipeline becomes a drain instead of a flywheel.

There's a second stabilizer worth naming: some workloads may never graduate. Plenty of tasks are hard enough that no amount of distillation gets a 6-cent model over the bar, and those tasks tend to be the ones enterprises value most. That's the same dynamic we covered in our story on Claude Cowork's expansion to web and mobile: the usage data there showed agents absorbing high-volume, mid-difficulty office work, exactly the tier that commands premium pricing today and is the obvious candidate to commoditize tomorrow. Where that line settles will decide how big the premium aisle stays.

The Starbucks question, revisited

Brandom points back to his own prediction from last September, that foundation labs might end up "selling coffee beans to Starbucks," supplying commodity inputs while the application layer captured the value. Parts of that came true. Vertical AI companies did switch to lighter models, and wrapper economics have held up. But the labs kept the part of the market everyone assumed would erode first: the premium token price.

So the frontier business isn't being commoditized. It's being specialized, into an R&D tier that the rest of the market pays a steep toll to access. Whether that's a durable position or a shrinking one depends on how fast open source moves upmarket, and on how the big platform players respond. Microsoft, in particular, is making moves that bear directly on this two-tier structure, which we get into in our next story.

For now, the scoreboard reads like this: open source owns the tokens, the frontier owns the invoices, and everyone is betting the other side's advantage runs out first.