Chinese Models Now Run 60% of OpenRouter Traffic, Threatening the IPO Premiums on Claude and GPT

The economics of frontier AI cracked open this week. According to data cited by CNBC and confirmed by Databricks CEO Ali Ghodsi, the share of Chinese open-source models on OpenRouter — the developer platform that routes API traffic across providers — has climbed from roughly 1% in 2024 to more than 60% in May 2026. DeepSeek, Moonshot AI's Kimi, and Zhipu's GLM family are doing most of the work.

The price gap is the engine. CNBC's benchmarking shows a complete pass through a standardized evaluation set costing about $4,811 on Anthropic's Claude, $1,071 on DeepSeek, and just $948 on Kimi — roughly a 5x discount for results that, on most non-frontier tasks, are within striking distance. Ghodsi described what's happening inside enterprise accounts as an "advisor model" pattern: companies default to a cheap open Chinese model for the bulk of their workload and only escalate to OpenAI or Anthropic for the small slice of queries the smaller model genuinely can't solve.

That shift lands at the worst possible moment for the two leading U.S. labs. OpenAI filed its confidential S-1 with the SEC on May 22 at a target valuation of up to $1 trillion, and Anthropic is preparing an October listing at $900 billion. Both valuations are predicated on pricing power that the OpenRouter numbers directly contradict. Even Anthropic conceded the squeeze in a policy paper this week, writing that U.S. models are only "several months ahead" of Chinese counterparts on capability and that Beijing is "winning in global adoption on cost."

The picture inside the developer ecosystem is even starker than the platform-level share number suggests. OpenRouter's traffic skews toward startups, indie builders, and cost-sensitive workloads — exactly the demand curve that incumbents need to defend to justify a trillion-dollar multiple. Tooling providers like Cursor, Continue, and Aider have all added one-click fallback to DeepSeek and Kimi over the past quarter, and Hugging Face reports DeepSeek V3.2 and Kimi K2.6 sit in the top three most-downloaded chat models for 2026 so far.

U.S. labs still have real advantages — agentic depth, long-context reliability, regulated-industry trust, and the distribution that comes with ChatGPT and Claude's consumer brands. But the IPO pitch has to explain why developers will keep paying 5x for those advantages on routine work. With both OpenAI and Anthropic now within months of public-market scrutiny, the OpenRouter chart is becoming the slide nobody on the roadshow wants to talk about.