Chinese AI Models Hit 60% of OpenRouter Usage as Pricing War Threatens OpenAI, Anthropic IPOs

Chinese AI models accounted for about 1% of usage on OpenRouter in 2024. By May 2026, that figure had climbed past 60% of OpenRouter token usage, according to CNBC. The shift is breathtaking in both speed and scale.

The price gap is the obvious driver. IndexBox reports that MiniMax's M2.5 model (Shanghai) processed 4.55 trillion tokens on OpenRouter in February 2026 alone — within roughly two weeks of launch. Benchmark costs compiled by Artificial Analysis tell the story in dollars:

• Anthropic's Claude: $4,811 per benchmark run
• OpenAI's ChatGPT: $3,357
• DeepSeek: $1,071
• Kimi: $948
• Zhipu's GLM: $544

Claude costs nearly nine times more than the cheapest Chinese alternative for the same workload, CNBC noted. For builders running production workloads, those multiples add up fast.

The cost asymmetry has structural roots. U.S. labs face massive capital expenditure on the most expensive Nvidia chips, power grid constraints, and expanding data center footprints — costs that flow directly to customers. Chinese labs, constrained by U.S. chip export restrictions, were forced to optimize aggressively for efficiency rather than raw compute, CNBC reported.

The result: competitive frontier models built with less compute at lower operating cost. DeepSeek's upcoming preview already matches or nearly matches the latest from OpenAI, Anthropic, and Google on coding, agentic, and knowledge benchmarks. Moonshot, Xiaomi, and Zhipu have shipped similarly capable models in the past four months.

Even Anthropic acknowledged the gap in a May policy paper, conceding that U.S. models are only "several months ahead" of Chinese ones and warning that Beijing is "winning in global adoption on cost," according to CNBC.

Enterprises aren't waiting for the pricing gap to close — they're restructuring how they use AI entirely. The technique, called the "advisor model" by Databricks CEO Ali Ghodsi, routes most routine work through a cheap open‑source model and only calls a frontier model when the task exceeds its capabilities.

"You can curb costs really well this way," Ghodsi told.¹ The approach drastically reduces total spend while retaining quality.

CNBC reported that Figma CEO Dylan Field described three phases of AI adoption his customers go through: nobody uses it, then everyone competes on who can spend the most on tokens, then they realize costs are too high and start cutting. "Many enterprises are now entering phase three," the report stated. Figma is now selling features that cut customers' token consumption by 20–30%.

Even Google is pushing cheaper models. At I/O, CEO Sundar Pichai pitched Gemini 3.5 Flash, claiming that if the largest Google Cloud customers shifted 80% of workloads to it, they would save more than $1 billion a year. "Many companies are already blowing through their annual token budgets, and it's only May," Pichai said, per.¹

The cost pressure isn't theoretical — it's showing up in earnings. CNBC reported that Meta, Shopify, Spotify, and Pinterest all flagged rising AI and inference costs as a drag on margins during the most recent earnings season. Shopify specifically cited increased LLM costs that partially offset economies of scale.

The core premise behind OpenAI's and Anthropic's IPO valuations — that their pricing power is durable because no real alternative exists — is being tested in real time. Both are expected to file for IPOs at valuations north of $800 billion, with OpenAI's confidential filing possible as soon as this week, CNBC reported.

Prediction markets are split. Crypto Briefing noted that Polymarket odds for Anthropic hitting a $1.25 trillion valuation by year‑end surged from 48% to 76.5% in a single day, while the market for Anthropic having the second‑best AI model by May 2026 sits at 99.1%. Traders are pricing in resilience despite the commoditization threat.

There's one market where Chinese models can't compete: regulated industries. CNBC reported that Cohere, led by CEO Aidan Gomez, sells models to banks, defense agencies, and regulated industries that won't touch Chinese alternatives. Revenue grew sixfold last year. It's a narrow but high‑trust segment.

Nvidia is releasing its own free‑to‑download AI systems that companies can run on their own servers, sidestepping both Chinese models and locked‑down labs. Reflection AI recently raised at a multibillion‑dollar valuation to build American open‑source models for enterprises that want a domestic alternative, CNBC reported.

But outside of defense and banking, the case for paying a premium weakens every month. The U.S. government's AI Safety Institute noted DeepSeek downloads have risen nearly 1,000% since the R1 release in January 2025, the U.S. AI Safety Institute reported in September 2025.

OpenAI's internal view, according to a person familiar with its thinking cited by,¹ is that every new frontier model — including GPT‑5.5 — drives a surge in API and product usage, and enterprise demand is growing in a "vertical wall." Pricing pressure is reportedly not on the company's top‑ten list of concerns.

An anonymous enterprise AI CEO offered a counterpoint to:¹ the growth is real, "but it would expand even faster for frontier if this technique wasn't used." In other words, the advisor model is suppressing demand for premium models even as overall usage grows.

For builders, the message is clear: the era of paying a premium for "the best model" is ending. The winning architecture is a multi‑model routing layer that sends cheap tasks to cheap models and reserves frontier calls for genuinely hard problems. The labs that adapt their pricing to that reality will thrive. The ones that don't will watch their IPO multiples shrink.