China Doesn't Need to Beat Claude. It Needs to Be Second
Executive Summary
- Kimi, Qwen, DeepSeek and GLM have crossed the threshold from ‘cheap alternatives’ into credible substitutes for a growing set of coding, agentic and long-context workloads. The next phase of the US-China model competition will therefore be settled less by benchmark leadership than by developer routing decisions - four credible families is no longer a benchmark event, it is a supply structure.
- ‘Second default’ does not mean the world’s second-best benchmark model. It means the first alternative a developer routinely considers when a proprietary frontier model is too expensive, too closed, too inflexible or simply overqualified for the task. The evidence is behavioral rather than promotional: DeepSeek reached 18% of OpenRouter token share by early June 2026, Lindy.ai cut inference costs roughly 90% by moving agent traffic off Claude, and Harvey post-trained Moonshot’s Kimi K3 into its own legal model.
- The unit of competition is shifting from token price to cost per successful task. Agents turn one instruction into dozens of calls, so the economically relevant question is not which model is smartest, but which produces the cheapest acceptable outcome. That logic favors cascades, and cascades let Western labs keep the customer while losing the token volume, which is enough on its own to change industry economics.
- As intelligence becomes substitutable, the selector gains value. Chinese open weights compress the scarcity rent on generic inference while increasing the value of the routing, hosting, orchestration and application layers around it. The binding constraint on how far Chinese models travel is trust and geopolitics, not capability.
Key Call | What We See | Why it Matters |
1. Model capability is no longer the gating factor | Kimi K3 scores on intelligence shown in Artificial Analysis, versus GPT-5.6 Sol at 61 and Claude Fable 5 at 62; Qwen 3.8 open weights scores 58. | Chinese open weights now sit close enough to frontier proprietary models for workload substitution to be rational. |
2. Usage is moving from trial to routing | Usage is moving from trial to routing | DeepSeek doubled OpenRouter token share from 9% to 18%; V4 Flash reached 70% of DeepSeek agentic token flow within a month. | Agentic traffic is token-heavy, so share gains can translate into disproportionate inference volume. |
3. Value moves to the selector | Cross-family model adoption on OpenRouter more than doubled Jan–May 2026. | Routers, hyperscalers and application layers gain leverage as model loyalty weakens. |
Source: Counterpoint Research
Chinese Models are Entering the Global Developer Stack
The AI industry has spent two years asking whether Chinese labs can close the frontier-model gap with the US ones. Since earlier this year, this question is becoming stale and a more investable question has emerged: are developers beginning to route real work to Chinese models instead of Western alternatives?
The inflection point is that several Chinese model families now offer enough intelligence, context length and agent capability at sufficiently different economics and control points that switching has become rational. One model can be dismissed as a benchmark event, but four (or more) credible families – Kimi, Qwen, DeepSeek and GLM – begin to look like a supply structure.
Read Full Report Here: China Doesn't Need to Beat Claude. It Needs to Be Second
Similar Reports
Counterpoint Technology Overview - August 2026
AI 360 Pulse - Industry Trends, Intelligence and Impact, July 2026
Receive our insightful weekly newsletter and stay ahead of the competition.
Author
Wei Sun
Wei is a Principal Analyst in Artificial Intelligence at Counterpoint. She is also the China founder of Humanity+, an international non-profit organization which advocates the ethical use of emerging technologies. She formerly served as a product manager of Embedded Industrial PC at Advantech. Before that she was an MBA consultant to Nuance Communications where her team successfully developed and launched Nuance’s first B2C voice recognition app on iPhone (later became Siri). Wei’s early years in the industry were spent in IDC’s Massachusetts headquarters and The World Bank’s DC headquarters.
Marc Einstein
Marc has over 20 years of experience in the ICT technology research and consulting focusing largely on the Telecommunications and Enterprise IT sectors. Prior to joining Counterpoint Research Marc held several senior positions in industry analyst firms in the USA, Hong Kong, Singapore and Japan. Based in Tokyo since 2010, Marc is a regular speaker at industry events and a frequent TV panelist. Marc also spent time in the strategy department of the largest mobile gaming company in Japan. A speaker of 6 languages, Marc holds a BSBA in Finance from Washington University in St. Louis and was a visiting student at Rangsit University in Bangkok, Thailand.