China‑US AI Gap Shrinks as Chinese Models Close In on Performance While Cutting Costs

According to data from together.ai, China’s Kimi K3 scores 60 on a composite benchmark covering reasoning, knowledge, mathematics, and programming, just two points behind Anthropic’s Claude Fable 5 (62). Third‑party tests show Kimi K3 achieving an Elo of 1678 on the GDPval‑AA v2 suite, surpassing GPT‑5.5, GLM‑5.2 and Claude Opus 4.8, but still trailing Fable 5’s 1760.

Despite comparable scores, Kimi K3’s inference speed is about 41 tokens per second, roughly 75 tokens slower than Fable 5, and differences remain in factual accuracy, tool‑calling stability, and complex task continuity.

The Stanford AI Index 2026 report indicates that since early 2025 the top U.S. and Chinese models have swapped the lead multiple times; by March 2026 the U.S. advantage narrowed to 2.7 percentage points and stayed in single‑digit territory for most of the year.

Cost analysis from Artificial Analysis shows Kimi K3’s average per‑task cost at $0.84 versus $2.75 for Fable 5—a 69 % reduction. Using a 7:2:1 weighting for cache‑hit, input and output tokens, Kimi K3 costs about $2.31 per million tokens, compared with $7.70 for Fable 5. For enterprises running large numbers of AI‑agent tasks, a two‑point performance gap may be less significant than a three‑fold cost gap.

Similar patterns appear with Zhipu’s GLM‑5.3, which scores 60 and costs $0.68 per task—19 % cheaper than Kimi K3 and 45 % cheaper than GPT‑5.6 Sol.

These trends suggest a shift in AI competition: rather than asking which model achieves the highest score, the focus moves to how many tasks can be completed per dollar, model openness, and deployment scale.

Enterprises can adopt a mixed‑routing strategy, assigning high‑risk, high‑complexity workloads to leading U.S. models while delegating high‑volume, routine tasks (e.g., conversational queries, code generation, search, automation) to near‑state‑of‑the‑art Chinese models to lower overall inference costs.

Open‑weight models further amplify cost advantages, as developers can download, fine‑tune, and privately deploy them, and cloud providers can optimize inference for diverse hardware.

OpenRouter data indicate growing overseas usage of Chinese models, with some markets showing higher Chinese‑model share than U.S. models, though U.S. AI assistants still lead in app‑download numbers, brand experience, and enterprise channels.

Nevertheless, the United States retains a substantial lead in compute and capital: 2025 private AI investment reached $2.859 trillion in the U.S., 23 times China’s $124 billion; the U.S. operates 5,427 data centers and controls the majority of the world’s AI‑compute resources (≈17.1 million H100‑equivalent GPUs, >60 % from NVIDIA).

While capital and chip advantages do not guarantee perpetual model leadership, they sustain the U.S. ceiling for AI capabilities. Chinese labs, constrained by compute supply, focus on hybrid‑expert architectures, model distillation, reinforcement learning, inference efficiency, and the development of domestic accelerators and advanced packaging.

Lower‑cost models may not reduce overall AI compute demand; cheaper per‑task pricing can encourage broader deployment of AI agents, increasing total token consumption and keeping demand for GPUs, ASICs, memory, and networking high.

Consequently, high‑priced API providers may face pressure, whereas hardware and infrastructure suppliers could benefit from expanding inference workloads.

In summary, the next phase of China‑U.S. AI competition will be judged not by marginal score differences but by which side can deliver sufficiently strong, affordable, and reliable AI at scale, turning cost‑effective models into the backbone of global AI infrastructure.