As reported by Axios, with OpenRouter usage data via Office Chai.
DeepSeek now leads OpenRouter by token volume, open-weight models account for the majority of traffic, and the headline share collapse obscures a more important split: volume is not revenue.
OpenRouter Token Share — June 2025 → June 2026
June 2025
US-origin models (OpenAI, Anthropic, Google combined) handle ~70% of tokens on OpenRouter
Late 2025
Open-weight Chinese models (DeepSeek, Tencent, MiniMax and others) accelerate adoption among cost-sensitive developer workloads
June 2026
US-origin share falls to ~30%; open-weight models now account for the majority of total token volume; DeepSeek leads all providers by token volume
Ongoing
Premium closed-frontier models (Anthropic, OpenAI) retain disproportionate revenue per token on high-value, agentic and reasoning tasks
OpenRouter is a model-routing marketplace where developers send queries to whichever AI model they choose, making its usage logs one of the cleaner available reads on real developer preference in the wild. According to data reported by Office Chai, US-origin models — OpenAI, Anthropic, and Google combined — handled roughly 70% of the tokens flowing through the platform around June 2025. By mid-2026 that share had fallen to approximately 30%, with cheap, open-weight Chinese models taking over the top of the usage charts. Open-weight models now handle the majority of total volume on the platform, and DeepSeek leads all providers by token count. The usual caveats apply: these are usage-share figures on one platform, denominators shift across reporting periods, and the picture is moving fast.
The providers driving the open-weight surge — DeepSeek, Tencent, MiniMax and peers — have released models with competitive benchmark scores at a fraction of the inference cost of closed US labs. For developers running bulk, repetitive, or cost-sensitive workloads, the economics of switching are straightforward. OpenRouter’s routing layer makes that switching frictionless, which accelerates the trend and makes the platform a useful, if partial, signal of where developer default preferences are heading.
The key insight: Token volume and revenue do not move together when the volume is being captured by low-cost commodity models. The headline share number measures quantity consumed, not value captured — and those two things are diverging sharply at the inference layer.
The Structural Read
The token-volume collapse in US-model share is real, but reading it as a single verdict on the competitive landscape misses the more important economic split happening underneath. Open-weight Chinese models are winning bulk, commoditized inference work — the high-throughput, cost-elastic tasks where price per token is the primary selection criterion. Closed US frontier labs such as Anthropic are not primarily competing in that segment; they are pricing and positioning for reasoning-intensive, agentic, and enterprise-critical workloads where accuracy and reliability carry a premium that developers and enterprises will pay.
The result is a layered market: the commodity inference layer is going open and increasingly Chinese, while the high-value layer has not followed the same path — at least not yet. Revenue per token at the closed frontier remains multiples higher than at the open-weight tier, which means a lab can lose token-volume share while growing revenue. That is not a paradox; it is a familiar dynamic from other software markets where low-cost alternatives absorb the price-sensitive base while premium tiers hold on differentiated value.
Business Engineer — Frontier AI and the Kimi Delusion
“Volume share and revenue share are different clocks running at different speeds. The open-weight wave commoditizes the bulk layer first. Whether it eventually compresses the frontier depends entirely on whether open models close the qualitative gap on reasoning, reliability, and trust — and that gap has not closed yet.”
The Frontier AI and the Kimi Delusion framework on Business Engineer offers a useful lens here — not as settled fact but as a structured way to ask the right question: the bet that “value stays with the closed frontier” is a bet on continued differentiation, not a structural law. Open models keep improving. The trend is real and fast. The bracket has to stay open.
Three Implications
IMPLICATION 1 — INFERENCE IS COMMODITIZING FROM THE BOTTOM UP
Bulk, repetitive workloads — the largest share of raw token volume — are migrating to the cheapest capable model. This is not a temporary arbitrage; it reflects a structural economics shift at the inference layer. Routing platforms like OpenRouter accelerate it by reducing switching costs to near zero. Closed labs that depend on volume-driven revenue models face real pressure here.
IMPLICATION 2 — PREMIUM POSITIONING IS THE ONLY DURABLE MOAT AT THE MODEL LAYER
For closed US frontier labs, the strategic response is not to compete on price — it is to widen the qualitative gap on hard reasoning, agentic reliability, and enterprise trust. If that gap narrows, the volume-to-revenue dissociation closes and margin pressure follows. Anthropic’s Constitutional AI work and OpenAI’s o-series reasoning investments read differently when you frame them as differentiation bets against an open-weight commoditization wave.
IMPLICATION 3 — OPENROUTER DATA IS A LEADING INDICATOR WORTH WATCHING, WITH LIMITS
OpenRouter skews toward developer experimentation and cost-optimizing builders — not enterprise procurement or regulated-industry deployments. The platform is a signal, not a census. But as an early-adoption proxy, it has historically led broader market shifts by several quarters. The direction of travel matters even if the exact share figures should be held loosely.
Business Engineer Framework
The Map of AI — Where Does This Shift Sit in the Stack?
The Map of AI tracks 200+ companies across 9 layers of the AI stack. The OpenRouter shift is a Layer 3–4 event: inference routing and model access. Understanding which layer is commoditizing — and which layers retain pricing power — is the essential structural question for anyone building or investing in AI right now.
US-origin models losing roughly 40 percentage points of token volume on OpenRouter in twelve months is a meaningful data point, not a noise event — but its meaning depends entirely on which question you are asking. If the question is where bulk inference volume is going, the answer is open-weight and increasingly cheap. If the question is where high-margin revenue is going, the picture is more contested and the answer hinges on whether open models can close the qualitative gap on the hardest tasks. That gap is the only thing standing between a structural bifurcation and full-stack commoditization, and right now it is narrowing faster than the closed-frontier incumbents would prefer.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
