Is open source AI taking revenue away from Anthropic?

Not significantly, at least not yet. While open source models like DeepSeek dominate raw token volume on platforms like Vercel, Anthropic still accounts for more than half of total AI spend on that platform.

Why do frontier models still command most of the spending despite lower token volumes?

Frontier models cost dramatically more per token — Opus 4.8 is roughly 23x pricier than DeepSeek V4 Flash — and they continue to dominate early-stage, high-complexity deployments where teams are proving out new use cases.

What is the two-phase AI model lifecycle theory?

The idea, articulated by Decagon CEO Jesse Zhang, is that expensive frontier models are used to validate and prototype AI use cases, which then get handed off to cheaper open source alternatives once mature — meaning the two aren't direct competitors.

Where does Nvidia's Nemotron fit into this picture?

Nemotron is an emerging variable — backed by Nvidia's distribution reach and touted for its adaptability, it's expected to rapidly gain token-volume share, though its impact on overall spend dynamics remains to be seen.

Home›Large Language Models›Article

Large Language Models

Open Source AI Is Eating Token Volume — But Anthropic Still Owns the Tab

Open source models are winning on raw token volume, but Vercel and OpenRouter data suggest frontier labs like Anthropic still capture the majority of actual AI spend — pointing to a two-tiered market that may be more stable than critics assume.

Dispatch desk·July 8, 2026·4 min read

TL;DR

Open source models dominate token volume while frontier models like Anthropic's Claude still capture the majority of enterprise AI spend, suggesting a durable two-phase market rather than zero-sum competition.

What changed

Decagon CEO Jesse Zhang argued publicly that open source and frontier AI models occupy sequential lifecycle phases rather than competing head-to-head, a framing supported by Vercel and OpenRouter platform data showing Anthropic holding over half of total spend despite losing token-volume share to DeepSeek.

Why it matters

If this two-tiered model is structurally stable, it shapes how engineering teams should budget AI spend, which vendors retain pricing power, and how long frontier labs can sustain their current revenue without a capabilities moat collapse.

Editorial record

Drafted by autopilot. Reviewed when marked.

Decagon CEO Jesse Zhang dropped an interesting provocation recently, arguing that basically everyone has the open source vs. frontier AI debate completely backwards. His claim: mature AI deployments are absolutely migrating to cheaper, lighter models — he's seeing it at his own company. And yet, enterprise spending on expensive frontier models hasn't collapsed. It's barely moved.

If you find that paradoxical, you haven't been paying attention to how software actually gets deployed at scale.

The Two-Phase Model Nobody's Talking About

Here's Zhang's core insight, and it's a good one: frontier models and open source models aren't really competitors. They're sequential phases of the same product life cycle. Expensive, state-of-the-art models from labs like Anthropic get used to prove that a use case works — to prototype, to validate, to build confidence that some AI-assisted workflow is worth operationalizing. Once the use case matures and the patterns are understood, teams swap in cheaper open source alternatives and call it optimization.

Meanwhile, new use cases keep emerging, and those get kicked off on frontier models. The frontier spend never collapses because the frontier is always busy clearing ground for what comes next.

It's a reasonable theory. The data, while imperfect, doesn't contradict it.

What the Numbers Actually Show

Vercel's AI gateway dashboard offers a useful window here. DeepSeek has surged to the top of token volume rankings, processing roughly a third of all tokens flowing through Vercel's infrastructure in a recent week. Z.ai — the lab behind the GLM-5.2 model — jumped into fourth place over the same period. By raw token count, open source is clearly winning.

But scroll down to where the money actually lives, and Anthropic still accounts for more than half of total AI spend on the platform. Yes, that share has dipped slightly — partly because Anthropic's own pricing has risen, which is a fun irony — but it hasn't cratered. Not even close.

OpenRouter tells a broadly similar story across a larger, somewhat less enterprise-focused slice of the market. DeepSeek V4 Flash dominates raw usage at 5.3 trillion tokens weekly. The leading frontier model, Opus 4.8, handles just over 2 trillion. But here's the kicker: Opus 4.8 costs roughly 23 times more per token than V4 Flash — about $1.37 per million tokens versus six cents. Do the math on that multiplier, and Opus is almost certainly still capturing the bulk of total revenue even at a fraction of the volume.

And none of this yet accounts for Nvidia's Nemotron model, which appears poised to grab significant market share quickly, backed by Nvidia's formidable distribution muscle and the model's reported adaptability across deployment scenarios. That's a variable worth watching.

Why Anthropic Isn't Panicking (Yet)

There are two plausible explanations for why frontier labs are holding their ground despite losing the token-count horse race to open source.

The first is simple market expansion. The universe of AI-addressable tasks is growing faster than open source is capturing share from the top. Frontier labs keep their revenue up by continuously dominating early-stage deployments — the exploration phase — before cheaper alternatives take over in production. As Zhang frames it: "The frontier labs will keep owning discovery. Open source will increasingly own production." That's a reasonable division of labor, and it's potentially stable.

The second explanation is less flattering but equally plausible: some use cases are genuinely hard, and even "mature" AI deployments can't fully offload to a smaller model without meaningful capability regression. There's a floor to how much you can optimize away when you're doing complex reasoning, nuanced document analysis, or anything requiring sustained coherence over long contexts. The cheap model fails in ways that turn out to be expensive.

A Two-Tiered Market That Might Actually Stick

The interesting question isn't whether open source will keep growing — it will — but whether this bifurcated economy is a temporary equilibrium or a permanent structural feature of the AI industry.

Not long ago, there was a reasonable concern that frontier AI labs might end up like commodity suppliers: growing fast, capturing little value, watching margins evaporate as their outputs became interchangeable inputs to someone else's product. That fear hasn't gone away. But the data so far suggests frontier labs have more runway than the most pessimistic scenarios implied, precisely because they occupy a specific and defensible niche: the high-stakes, high-value, early-stage work that justifies premium pricing.

The catch, of course, is that "yet" is doing a lot of work in any sentence about Anthropic's resilience. Open source capability is improving fast. The gap between frontier and open source performance has been narrowing in ways that would have seemed implausible eighteen months ago. Every quarter that passes, the set of tasks that genuinely require a frontier model shrinks a little.

For now, Anthropic owns the tab. The question is how long the meal lasts.