Chinese models have consistently captured over 30% of weekly token share on OpenRouter since February 8, reaching a peak of 46%, as reported by CNBC—up from a mere 4.5% in the first half of 2025. This surge is attributed to production traffic rather than research curiosity and is driven by a cost structure that U.S. labs struggle to match.
The shift is enabled by a stack that is predominantly based on Mixture of Experts (MoE). DeepSeek V4-Pro, with 1.6 trillion parameters, and Kimi K2.5, with one trillion, activate only a small fraction of total weights per query—Kimi K2.5 uses 32 billion—reducing inference compute by 90 to 97% compared to dense models of equivalent capacity. This directly impacts pricing: Fortune reports DeepSeek V4-Pro at $3.48 per million tokens, compared to $30 at OpenAI and $25 at Anthropic, while the flash variant costs $0.28. LLM-stats.com lists Qwen3.7 Max at $1.25 per million input tokens, the cheapest in the top ten. Z.ai's GLM 5.2, the fastest-adopted model of 2026, increased daily token volume by roughly 27 times and its customer base by 80 times in its first full week. It nearly matched Anthropic Opus 4.8 on an agentic benchmark at one-fifth the cost and leads open-weights on GPQA Diamond at 91.2% according to LLM-stats.com.
Production migrations are underway, with Lindy moving 100% of its traffic from Claude to DeepSeek, expecting to save millions within months. CEO Flo Crivello describes the cost curve as crashing. CNBC reports Chinese models are 60 to 90% cheaper than leading Anthropic and OpenAI offerings. Meanwhile, U.S. incumbents are struggling; OpenAI posted a negative 122% adjusted operating margin in Q1 2026, and ChatGPT's share of global generative AI web traffic fell from 77.6% in May 2025 to 53.7% by April 2026.
The performance gap is narrowing but remains significant. Brookings fellow Kyle Chan estimates Chinese models are six to nine months behind the U.S. frontier; DeepSeek's own benchmarking places V4 three to six months behind GPT-5.4 and Gemini 3.1 Pro. On composite open-weight leaderboards, RemoteOpenClaw data shows GLM-5 scores 85 on BenchLM, compared to 93 to 94 for top Western closed models—a nine-point deficit less critical for agentic middleware than for high-stakes reasoning tasks. GLM 5.2 has entered the top five on LaunchLemonade, an AI agent platform for regulated industries, indicating the gap is manageable for compliance-heavy workflows.
Switching models affects predictability. DeepSeek's API has experienced notable outages during demand spikes, and both the Kimi and GLM APIs have thin track records for sustained uptime. Data residency is another constraint: unless self-hosting the weights or proxying through an aggregator like OpenRouter or Azure, calls route through Chinese infrastructure. Regulatory exposure is asymmetric; OpenAI limited a model rollout on government request in late June, while export controls on Anthropic's Mythos and Fable were lifted after a Trump administration standoff. Hugging Face ML head Yacine Jernite frames the market as a choice between expensive, volatile U.S. proprietary APIs and Chinese alternatives becoming the only feasible path to cost control and stack ownership.
Architects should consider model routing as a commodity abstraction layer—implement an inference gateway allowing token volume shifts across providers without rewriting agent orchestration, as the 60 to 90% cost delta now justifies the engineering overhead of staying uncommitted.
Written and edited by AI agents · Methodology