Chinese AI models in 2026: how close are they?

Chinese AI models in 2026: have they caught up?

For everyday chat, nearly. In LMArena's rankings (8 October 2026), where people vote on anonymous answers, Moonshot's Kimi K3 sits 16th, close behind the top US models. On hard reasoning and coding tests they're clearly behind: the best Chinese model scores 46 on the Artificial Analysis Intelligence Index, against 58 for Anthropic's Claude Opus 5.5.

That matches my own view: cheap models sit slightly off the frontier, are still very powerful, and are perfect for everyday tasks. The cheapest Chinese models also cost a small fraction of the US flagships per token, though OpenAI and Anthropic now sell budget models at similar prices.

A correction, dated 11 October 2026: the August version of this page said the top Chinese model was one point off the best US model. That was wrong even on the August data, where the gap was two to three points, and Artificial Analysis has since changed how it calculates its index, so August and October scores can't be compared.

The main Chinese AI models in October 2026

These are the newest notable models from each major Chinese lab, with release dates from Artificial Analysis and the labs' Hugging Face pages.

Seven of the models in this table shipped between 16 July and 21 September 2026, a new one every week and a half on average. Most are released as open weights, which means anyone can download and run them. Kimi K3's license asks companies that sell model access and earn over $20 million a year to sign a separate agreement with Moonshot. ByteDance's models power its Doubao chatbot app, and DeepSeek, Alibaba's Qwen and Moonshot's Kimi all run their own consumer chat apps too.

Plainly's own history follows the same labs. Its past models are DeepSeek V4 Flash, DeepSeek V4.1 Flash, GLM 5.3 Flash and Xiaomi's v2.6 Pro, all from Chinese labs and all released in 2026.

How Chinese AI models score against ChatGPT, Claude and Gemini

Two measures, checked the same week. The Artificial Analysis Intelligence Index (11 October 2026) combines hard reasoning, knowledge, coding and agent tests, and the scores below are each model's highest effort setting. LMArena's text leaderboard (8 October 2026) ranks models by people's votes between two anonymous answers, out of 414 models. Prices are per million tokens, input then output, from each lab's pricing page on 11 October 2026.

*Xiaomi's own endpoint as listed on OpenRouter; its pricing page didn't load. LMArena marks Gemini 4 Argon's ranking as preliminary. Gemini 4 Argon is in a restricted preview and isn't on Google's public API price list.

Where Chinese models are close: everyday chat

LMArena is the nearest public measure of everyday chat, because the votes come from people asking their own questions. On it, the gap is small. Kimi K3 sits 37 points below first place. On LMArena's rating scale, that means voters would pick the top model in about 55% of head-to-head matchups. DeepSeek V4.1 Flash and GLM-5.3-Flash, the cheap Chinese workhorses, sit about 50 points off first place and roughly 30 points above GPT-6 Luna, the model on ChatGPT's free plan.

On the harder Artificial Analysis index, DeepSeek V4.1 Flash and GPT-6 Luna score 39 and 38. On both measures, the cheap Chinese models match or beat the model on ChatGPT's free plan, and open Chinese models such as MiMo-V2.6-Pro and GLM-5.3 score higher still.

Where they're behind: the hardest work

At the top, the US leads clearly. Six US models score between 52 and 58 on the Artificial Analysis index, from Claude Opus 5.5 (58) and Claude Sonnet 5.5 (56) down to GPT-6.1 Sol (52). The best Chinese model, MiMo-V2.6-Pro, scores 46. Most of the top US models are only weeks old: Claude Opus 5.5 shipped on 22 September, GPT-6.1 Sol on 29 September and Gemini 4 Argon on 30 September.

The pattern has held for a while. In September 2025, the US government's AI standards center tested DeepSeek's models against US ones and found the best US model solved over 20% more software engineering and cyber tasks (NIST's Center for AI Standards and Innovation, 30 September 2025). I expect the US labs to keep the lead at the very top.

What Chinese AI models cost

Per token, the cheapest Chinese models cost a small fraction of US flagships: DeepSeek V4.1 Flash's off-peak input price is a thirteenth of GPT-6.1 Sol's and a sixty-seventh of Claude Fable 5.1's. The gap with US budget models has closed, though. GPT-6 Luna and Claude Haiku 5.5 list slightly below DeepSeek's off-peak price. Xiaomi's MiMo-V2.6-Flash has the lowest output price of the main Chinese models, about half of DeepSeek V4.1 Flash's. The Chinese flagships, Kimi K3 and Qwen3.8 Max, are priced close to GPT-6.1 Sol. The reasons behind these prices, and why a cheap token can still mean an expensive task, are in why Chinese AI models are cheaper.

My own bill shows the everyday end. As a very heavy AI user, I spent $18.90 in total on two Chinese chat models, DeepSeek V4 Flash and GLM 5.3 Flash, from 16 August to 15 September 2026, paying per use through OpenRouter.

Can you use Chinese AI models outside China?

Yes. Open weights mean the model can run anywhere, so US and other non-Chinese clouds host the same models. On 11 October 2026, OpenRouter listed 29 hosts for DeepSeek V4.1 Flash, including CoreWeave, Together AI and Fireworks, and Amazon Bedrock among the hosts for Kimi K3. The weights also stay available if a lab changes course. Why Chinese labs publish them at all is covered in why China is open-sourcing its AI models.

The labs' own apps are a different matter, because your chats go to the lab. Are Chinese AI models safe to use covers the apps, the hosts and the published safety tests, and DeepSeek vs ChatGPT compares the two apps directly.

What this means for you

For everyday tasks, the cheap Chinese models are close enough, and on LMArena they rank above the model on ChatGPT's free plan. For the hardest coding or analysis, the US flagships still score clearly higher.

You don't need to track this table to get the benefit. Plainly runs one capable open-source model and switches when a better one comes out. You never pick. It costs $5 a month for unlimited chat. The case for letting the app choose is in why you don't need to pick a model.