Is DeepSeek still the best Chinese AI model?

It is the most famous, and its V4 generation is current — V4-Flash shipped 31 July 2026 and V4-Pro on 13 August 2026. But Kimi K3 now scores highest among open models on the Artificial Analysis index, and GLM and Qwen ship competitive models at lower prices.

Which Chinese AI model is cheapest?

The efficient tier — DeepSeek V4-Flash, GLM-5.3-Flash, Qwen 3.8 Flash — all serve answers for a few cents per million tokens, an order of magnitude below frontier pricing. Exact figures move monthly; the order of magnitude is the stable fact.

Are Chinese AI models as good as ChatGPT?

On everyday tasks, yes — you would struggle to tell the difference. The top US frontier still scores higher on the hardest benchmarks, but most questions are not the hardest benchmarks.

Why are Chinese AI models so cheap?

Less funding, export bans on the best chips and brutal domestic competition forced them to optimise for efficiency.

Are Chinese AI models safe to use?

The models are open weights — anyone can run them anywhere, and zero-data-retention providers serve them without keeping your prompts.

Have Chinese AI models caught up? The 2026 picture, dated

Yes — at the low and middle end they own the best cost-to-output in the market, and a new one lands almost weekly. The US frontier still wins the very top. Most people do not use the very top. The table with dates and prices.

Have Chinese AI models caught up?

At the low and middle end, they own the best cost-to-output in the market, and a new one lands almost weekly. The US frontier still wins the very top. Most people do not use the very top. The table has the dates.

The score table

Model	Released	AA Index (v4.1.1)	API price in/out, $/MTok	Licence
DeepSeek V4-Flash	31 Jul 2026	52	A few cents in, ~20¢ out — order of magnitude below frontier	MIT
GLM-5.3-Flash (Z.ai)	26 Aug 2026	57	$0.075 / $0.25 (50% launch promo to 9 Sep 2026; list $0.15 / $0.50)	MIT
Kimi K3 (Moonshot)	16 Jul 2026	60 — top open model	$3.00 / $15.00	Restricted
Qwen 3.8 Flash (Alibaba)	~23 Aug 2026	Not yet indexed	$0.15 / $0.47	—

Benchmark scores from the Artificial Analysis Intelligence Index v4.1.1, captured 29 August 2026. Prices from OpenRouter's model API, 2 September 2026; they move, so read them for shape.

The frontier reference

On the same index as of 29 August 2026: Sol (max) scores 61, Fable 5 scores 62, and GPT-5.6 Luna — the model on ChatGPT's free tier — scores 52. So the top open Chinese model sits one point off the best US frontier, a capable Chinese model ties the free-tier ChatGPT model at a fraction of the serving cost, and the whole efficient band runs from 52 to 60. That band is where everyday questions live.

The cadence

One model proving parity would be an event. This is a drumbeat. DeepSeek has shipped open weights continuously since January 2025. Moonshot, MiniMax, Xiaomi, Alibaba, Baidu and Z.ai now cover the whole spectrum from lightweight to heavy, and a cost-efficient release lands every few weeks. Why they can do this — less money, fewer chips, more craft — is the story in why Chinese AI models cost less, and why they publish the weights at all is in why China open-sources while the US stays closed.

Where the US still wins

The frontier earns its price at the top end. The hardest coding, the most demanding analysis, anything where the first answer must be the right one — that is US frontier territory, priced accordingly. GPT-OSS vs DeepSeek shows what happens when the US plays the open-weights game itself: the gap to the Chinese open models is large. The frontier is excellent. It is also not where Tuesday's email lives.

The Pareto point

The useful frame is cost-to-output. Chinese models own the frontier of best output per dollar at low and medium prices; US models own the absolute top. If your question sits below the top — and everyday questions do — the Chinese models are the efficient option. Where that leaves the buyer is in free vs $5 vs $20, and where the prices go next is in the future of cheap AI.

The honest limit

Benchmarks move weekly and this table is dated — re-check it before quoting it. Kimi K3's licence restricts commercial use, which matters if you are building rather than asking. And "caught up" is a claim about the band where people actually work; on the hardest problems the US frontier still leads, and I expect it to keep leading there.