Chinese AI Models vs US Frontier: Who Leads in October 2026

Chinese AI models vs US frontier labs, as of 4 October 2026: the best Chinese open-weight models trail the top US closed models by 12 points on Artificial Analysis' independent index (46 against 58), sit 18 points below first place on LMArena's text leaderboard, and sit within about one point of the leaders on some coding benchmarks that the labs report themselves. They cost a fraction per token and several ship open weights. This article separates verified facts from our reading of them.

How did Chinese AI models catch up? A dated timeline

Dates below come from the labs' own news pages where they exist (DeepSeek, Z.ai) and otherwise from the creation date of the Hugging Face weights repository, which can precede public availability.

Date	Release	What it was
26 Dec 2024	DeepSeek-V3	Open-weights mixture-of-experts model
20 Jan 2025	DeepSeek-R1	Reasoning model, MIT licence
11 Jul 2025	Kimi K2	Moonshot, about 1T parameters
22–28 Jul 2025	Qwen3-Coder, GLM-4.5	Alibaba's 480B coder (Apache-2.0) and Z.ai's GLM-4.5 series
22 Oct 2025	MiniMax-M2	Agent and coding model
1 Dec 2025	DeepSeek-V3.2	MIT licence
12 Feb 2026	GLM-5	Z.ai flagship generation 5
24 Apr 2026	DeepSeek-V4 preview	V4-Pro 1.6T total and 49B active parameters, V4-Flash 284B and 13B, 1M context by default
2 Jun 2026	MiniMax-M3	About 428B parameters, 23B active, 1M context
13 Jun 2026	Kimi K3	2.8T parameters; repository date, Kimi's own pricing page is dated 28 Jul
5–8 Aug 2026	Qwen3.8	27B dense (Apache-2.0) and 2.4T-A95B
13 Aug 2026	DeepSeek-V4-Pro GA	General release of V4-Pro
18 Aug 2026	GLM-5.3	Weights followed on 25 Aug
10 Sep 2026	DeepSeek-V4.1-Flash	Current Flash model

Our read. The cadence matters more than any single model. Between the V4 preview in April and V4.1-Flash in September, DeepSeek alone shipped four dated releases, and four other labs shipped flagship updates in the same window.

Where do Chinese AI models lead or trail on benchmarks?

Three sources, three answers. Read each one for what it measures.

Artificial Analysis Intelligence Index v4.3.2 (leaderboard, read 4 October 2026, page undated) averages ten evaluations that include Terminal-Bench 4.0, GDPval-AA and Humanity's Last Exam. It is independent of the labs.

Model	Lab	Index	Weights
Claude Opus 5.5 (max)	Anthropic	58	Closed
GPT-6 Astra (max)	OpenAI	53	Closed
MiMo-V2.6-Pro	Xiaomi	46	Open
GLM-5.3 (max)	Z.ai	45	Open
Kimi K3 (max)	Moonshot	44	Open
DeepSeek V4.1 Flash (max)	DeepSeek	39	Open
MiniMax-M3	MiniMax	29	Open

The six highest open-weights entries in AA's open-source table are all from Chinese labs. Qwen3.8 Max (0902) also scores 45, but LMArena lists it as proprietary.

LMArena text leaderboard (page stamped 2 September 2026, 7,999,020 votes) ranks by human preference. First place, claude-fable-5, scores 1507. Kimi K3 is 12th at 1489 (±5), GLM-5.3 is 20th at 1482, and DeepSeek-V4-Pro-0813 is 52nd at 1460. Several US models rank below Kimi K3 there. On LMArena's WebDev leaderboard (same date) Kimi K3 is 4th at 1674 against 1765 for first place.

Lab-reported coding benchmarks come from model cards. This table is a single one, from Z.ai's GLM-5.3 card.

Benchmark	GLM-5.3	Kimi K3	DeepSeek V4-Pro-0813	Fable 5	GPT-5.6 Sol	Opus 4.8
Terminal Bench 2.1	88.2	88.3	87.9	88.0	88.8	85.0
DeepSWE v1.1	66.9	67.5	62.7	69.7	72.7	58.0
Terminal Bench 3.0	28.3	17.4	not reported	33.7	34.6	21.1

DeepSeek's own V4.1-Flash model card lists 90.6 on Terminal-Bench 2.1 against 89.1 for Opus-5.0 and 88.8 for GPT-5.6 Sol, and 74.2 on DeepSWE against 74.0 for Opus-5.0. Those are DeepSeek's numbers, run in its own harness, and the same card shows a Terminal-Bench 3.0 score of 30.0 against 43.3 for Opus-5.0.

Why the sources disagree

Self-reported tables pick their own harness and their own comparison models. The same Opus 4.8 scores 85.0 on Terminal Bench 2.1 in Z.ai's table and 84.6 in Kimi's. DeepSeek's V4.1-Flash card shows 65.5 to 74.2 on DeepSWE depending on the agent scaffold. Labs also compare against older US versions (Opus 4.8, Fable 5), while AA scores the newer Opus 5.5 and Fable 5.1.

Our read. On agentic terminal and repository tasks the gap looks small. On AA's broader index it is 12 points. Arena voters land in between. Treat each as one measurement, and test candidates on your own tasks.

Does open weights or price per token explain the gap?

Open weights are real but uneven. DeepSeek V4 and V4.1-Flash are MIT, Qwen3.8-27B is Apache-2.0. GLM-5.3 uses an MIT-style licence with a security-review clause for very large model-as-a-service operators. Kimi K3's licence requires a separate agreement from model-as-a-service businesses above US$20M in 12-month revenue and name display for products above 100M monthly users or US$20M monthly revenue. MiniMax-M3's community licence requires a "Built with MiniMax M3" notice for commercial use. Several Chinese frontier models, such as Qwen3.8 Max, are API-only.

List prices per 1M tokens, read 4 October 2026 from Anthropic, DeepSeek, Z.ai, Moonshot and MiniMax:

Model	Input	Output	Example cost
Claude Opus 5.5	$4	$20	$15.00
Kimi K3	$3	$15	$11.25
GLM-5.3	$1.40	$4.40	$4.86
DeepSeek V4-Pro (off-peak to peak)	$0.66–1.32	$1.98–3.96	$2.28–4.55
DeepSeek V4.1 Flash (off-peak to peak)	$0.15–0.30	$0.60–1.20	$0.54–1.08
MiniMax-M3	$0.30	$1.20	$1.08

The example assumes 3M uncached input tokens plus 150K output tokens, for instance $4 × 3 + $20 × 0.15 = $15.00 for Opus 5.5. Caching lowers every row. Per-token price is also not per-task cost: AA's cost-per-task column shows GPT-6.1 Sol (max) at $0.72 with index 52, against $2.01 for GLM-5.3 (max) at 45, because models spend different amounts of reasoning tokens.

Are the training-efficiency claims credible?

These are claims as published, not findings of ours. The DeepSeek-V3 paper reports 2.788M H800 GPU hours and, assuming a $2 per GPU-hour rental price, $5.576M, and says this excludes prior research and ablation runs. DeepSeek's R1 paper appeared in Nature on 17 September 2025. For V4-Pro, the model card says it needs 27% of V3.2's single-token inference compute and 10% of its KV cache at 1M context. MiniMax's M3 card claims 9× faster prefill and 15× faster decode than M2 at 1M context. We have not reproduced any of these.

What did chip export controls change?

Reported facts, kept neutral. A CNAS analysis (16 January 2026) describes a 13 January BIS rule that licenses Nvidia H200 exports to China case by case, capped at 50% of US sales of that chip, plus a 25% import tariff set by proclamation on 14 January. CNBC reported on 19 August 2026 that Chinese firms reportedly reach advanced Nvidia compute through Southeast Asian data centres, that a White House official accused Moonshot of using GB300 chips in Thailand, and that a policy researcher said the current regime covers physical chips, not remote access. Z.ai's GLM-5.3 card also lists Huawei Ascend deployment support.

Our read. Public data cannot isolate the effect of controls on release speed. We would not claim they helped or hurt.

What should developers and companies choose?

Four checks matter more than a leaderboard rank.

• Data residency. DeepSeek's privacy policy says it collects, processes and stores personal data in the People's Republic of China. Z.ai's data processing terms say customer data is generally processed in Singapore and that content is not stored. Self-hosting open weights removes the question.
• Licence. Read the text on the model page, as the thresholds above differ by model.
• Terms. Anthropic's Claude Code gateway docs say it does not support routing Claude Code to non-Claude models.
• Fit. Run 50 to 400 of your own tasks. Our post on cutting LLM costs shows how we score routing choices.

Need	Reasonable starting point
Cheapest agent tokens	DeepSeek V4.1 Flash, MiniMax-M3
Strongest open weights on independent index	MiMo-V2.6-Pro, GLM-5.3, Kimi K3
Self-host on one GPU	Qwen3.8-27B (Ollama lists 18GB)
Strictest data control	Self-hosted open weights, or a US or EU host you have a contract with

For coding picks see the best open-weight coding models, and for subscriptions see the Chinese coding plans compared. To route Claude Code to other models, read Claude Code with any model.

How Danube Labs can help

Danube Labs builds an evaluation set from your real tasks, runs several models against it and recommends the cheapest one that passes, with a written note on licences and data location. Details are on the AI systems page, or book a free intro call.