The short answer: yes, on raw benchmark performance, China has effectively caught up. The longer answer—the one that matters for your budget, your use case, and your vendor risk—is more complicated.

Let’s start with the number that matters most.

The Gap: From 1,300 Points to 39

Stanford’s 2026 AI Index Report tracks the performance gap between the top US and Chinese models using Elo ratings (the same system used for chess rankings). In May 2023, the gap stood at over 1,300 points. By March 2026, it had shrunk to just 39 points. That’s a lead of 2.7% for the top US model, Anthropic’s Claude Opus 4.6 (1,503 Elo), over China’s Dola-Seed 2.0-Preview.

The report’s phrasing is worth quoting directly: “The U.S.-China AI model performance gap has effectively closed.” US and Chinese models have traded places at the top of performance rankings multiple times since early 2025. In February 2025, DeepSeek-R1 briefly matched the top US model.

Bloomberg Intelligence puts the gap at a record-low 6% as of June 2026, down from 9% in May. On specific benchmarks like MMLU, HumanEval, and GSM8K, Chinese models including DeepSeek V4, Qwen 2.5 Pro, and GLM-5.1 have effectively erased the difference with US counterparts like GPT-5.5 and Claude Opus 4.

The caveat: benchmark scores are not the same as real-world utility. As Stanford notes, the benchmarks themselves have growing “reliability and gaming concerns,” with error rates up to 42% on widely used evaluations like GSM8K. A model can ace a test and still fail at tasks that matter to you.

The Chinese Models You Need to Know (and What They Actually Cost)

DeepSeek

DeepSeek V3 is the baseline. It achieved the highest mean diagnostic score (2.32) in a clinical evaluation, beating GPT-4o (2.23) and Claude 3.5 Sonnet (2.12). In coding tests, it matches Claude 3.5 on accuracy while generating solutions faster.

Pricing: DeepSeek V3 runs about $0.27 per million input tokens and $1.10 per million output tokens through the official API. On third-party hosts, it can go lower. DeepSeek V4 Flash pushes this further: $0.14 per million input tokens and $0.28 per million output tokens. DeepSeek V4 Pro costs $0.435 input and $0.87 output.

For context: a coding workload that costs $4,811 on Anthropic and $3,357 on OpenAI runs about $1,071 on DeepSeek. That is not a typo.

Qwen (Alibaba)

Qwen3.7 Max currently leads the MMLU-Pro leaderboard with a score of 0.896 across 129 evaluated models. Qwen models are among the most widely adopted open-weight families globally, backed by Alibaba Cloud as a full platform.

Pricing: Qwen3.7 Plus runs $0.32 input and $1.28 output per million tokens. Qwen2.5 72B on third-party providers works out to around $5–$10 per month for moderate usage. The pricing varies significantly by provider and deployment method.

GLM (Zhipu AI)

GLM-5.2 is a 750-billion-parameter open-weight model ranked second on the Code Arena frontend coding benchmark, operating at about one-sixth the cost of US frontier models. GLM-5 (the base version) is priced at $1.00 input and $3.20 output per million tokens.

GLM-5.1 has been clocked at 400 tokens per second—one of the fastest inference speeds globally. Speed matters. A model that’s 10% smarter but 3x slower is a worse user experience.

Kimi K3 (Moonshot AI)

Kimi K3 is the most recent headline-grabber. Released July 2026, it’s a 2.8-trillion-parameter MoE model with a million-token context window. It reached first place on Arena.ai’s Frontend Code leaderboard on July 16, scoring 1,679 points—ahead of Claude Fable 5 (1,631) and GPT-5.6 Sol (1,618). Its predecessor had ranked 18th.

The company’s own tests found K3 can match or outperform rival US models on coding and spreadsheet manipulation tasks.

Pricing: $3.00 per million input tokens and $15.00 per million output tokens, with cache-hit input at $0.30. This puts it at the same price point as Claude Sonnet 5 ($3/$15). It undercuts OpenAI’s GPT-5.6 Sol ($5/$30) and Claude Opus 4.8 ($5/$25).

One catch: K3 is large. Too large to run on personal devices. It will “probably require significant investment by institutions to run it”. The weights are expected to be released under a modified MIT license by July 27, 2026—so self-hosting is technically possible, just not cheap.

Tencent Hy3

Tencent open-sourced Hy3 under Apache 2.0 in July 2026. API pricing: 1 yuan (~$0.14) per million input tokens and 4 yuan (~$0.56) per million output tokens. That compares to $15 per million input tokens for OpenAI’s GPT-5.5. Hy3 claims performance comparable to models 2–5 times its size.

The Pricing Picture: Not All Chinese Models Are Cheap Anymore

Here’s the table that matters:

Sources: OpenAI and Anthropic pricing; DeepSeek; GLM-5; Kimi K3; Tencent Hy3.

The gap is stark—but it’s not uniform. Kimi K3 sits at the same price as Claude Sonnet 5. Some Chinese models have moved beyond the “cheap” phase to compete on capability directly. The narrative that all Chinese AI is uniformly cheaper is oversimplified.

However, the price gap at the lower end is real and significant. DeepSeek charges about $0.28 for the same output that costs $25 on Opus 4.8—a 99% discount. As of June 2026, Chinese models captured up to 46% of US enterprise token usage. San Francisco-based AI assistant startup Lindy moved some workloads from Anthropic to DeepSeek, citing millions of dollars in savings.

The Open-Weight Difference

Six of the eight top Chinese labs—DeepSeek, Alibaba (Qwen), Zhipu AI (GLM), Moonshot AI (Kimi), MiniMax, and to some extent Tencent (Hunyuan) and Baidu (Ernie)—release their core models with open weights under MIT or Apache 2.0 licenses.

The top Western systems (GPT-5, Claude, Gemini) remain exclusively closed.

This is not a trivial distinction. Open-weight models can be:

• Self-hosted on your own infrastructure
• Fine-tuned on your proprietary data
• Audited for security and compliance
• Run without sending data to third-party APIs

The trade-off: you manage the infrastructure, monitoring, and uptime. The total cost of ownership may still favor closed APIs for small teams. For enterprises with data sovereignty requirements, open-weight Chinese models are increasingly the only cost-effective option.

What the Numbers Don’t Tell You

1. Investment asymmetry

US private AI investment reached $2,859 billion in 2025—about 23 times China’s private investment. Yet the performance gap is 2.7%. This isn’t because China is “more efficient.” It’s because US capital flows into data centers and GPU clusters (5,427 data centers vs. China’s 449), while Chinese firms, under resource constraints, have prioritized algorithmic efficiency and application engineering. The US builds infrastructure; China builds models that run on less of it.

2. Research depth

US AI R&D spending was about $120 billion in 2025 vs. China’s $40 billion—a 3x gap. US research focuses on new architectures, AI for science, and interpretability. Chinese research focuses on model engineering, training efficiency, and deployment. The number of top-tier AI researchers (those publishing at NeurIPS, ICML, ICLR) in China is about 40% of the US total. Application talent in China exceeds the US. This matters for the next breakthrough, not the current one.

3. Safety and alignment

US AI companies maintain safety teams of 50–150 people. Chinese companies typically have 10–30. In AI alignment, red-teaming, and robustness evaluation, Chinese investment is visibly lower. China’s regulatory push on generative AI is starting to change this, but the capability gap in safety research remains real.

So, Has China Caught Up?

On benchmark performance: Yes. The gap is effectively closed. A 2.7% Elo difference is a single release away from flipping. Chinese models lead in some benchmarks (MMLU-Pro, frontend coding) and trail in others.

On pricing: Unequivocally yes for the low-cost tier. DeepSeek and Tencent Hy3 are 10–99% cheaper than US counterparts on listed API rates. Kimi K3 has chosen to compete on capability at near-US prices.

On ecosystem and infrastructure: No. The US maintains a massive lead in compute infrastructure, data centers, and foundational research investment. China has 12x fewer data centers. Power constraints are becoming the binding constraint for both, but the US has more of it, for now.

On open access: China leads, by a wide margin. Six of eight top Chinese labs release open-weight models. Every major US lab keeps its frontier models closed.

What This Means for You

If you’re a developer or startup: Chinese open-weight models are worth evaluating. The cost savings are not marginal—they’re an order of magnitude. Run your own benchmarks on your own tasks. Benchmark scores are a starting point, not a conclusion.

If you’re an enterprise: Selective adoption makes sense. Use cheaper Chinese models for non-sensitive workloads—coding assistance, summarization, translation, batch processing. Self-host open-weight models to keep data inside your infrastructure. Don’t swap all your OpenAI usage overnight. Do route the workloads that don’t require frontier performance to models that cost 1% as much.

If you’re a researcher: Chinese open-weight models give you access to frontier-scale systems without paying API bills. The trade-off is that you’re working with models that are a few months behind the absolute frontier, but the gap is shrinking.

If you’re just reading to understand the landscape: The “China caught up” headline is true on benchmarks. The “China surpassed the US” headline is not—yet. The real story is about price, access, and open-weight availability, not just a single Elo score.

Bottom Line

Benchmarks say China has caught up. Budgets say China has lapped the US on price. Infrastructure and research depth say the US still holds the long-term advantage.

All three statements can be true at the same time.

For most practical purposes—building applications, running inference at scale, managing costs—the Chinese models are good enough, and they’re dramatically cheaper. The smart move is to test them on your actual workloads and let the results, not the headlines, guide your decision.

Sources:

• Stanford HAI, 2026 AI Index Report
• Bloomberg Intelligence, July 2026
• DoNews, “斯坦福AI报告解析”
• Nature, “Does China’s latest AI model finally equal US rivals?”
• TechRepublic, “Chinese AI Models Challenge OpenAI and Anthropic on Cost”
• Notebookcheck, “Kimi K3 tops Frontend Code Arena”
• Edgen.tech, “腾讯开源Hy3 AI模型”
• Sohu, “斯坦福AI报告2026”