China’s new AI:Is DeepSeek really 6× more efficient than Claude?

Efficiency headlines work because they hit a real worry: AI can get expensive, and nobody likes paying double for the same result. “6× more efficient than Claude” promises exactly that — six times more for your money. Before acting on it, a sober look pays off: what was measured, what was claimed, and what it concretely means for a business that wants to make money or cut costs with AI.

What “6× more efficient” actually means

“Efficiency” isn’t a single value but at least three different things — and that’s exactly where catchy numbers come from:

• Cost per token. What does one processed unit of text cost? Here, Chinese open-weight models are indeed often dramatically cheaper.
• Compute per response. How much GPU time/energy does a request take? This is the technical heart of the efficiency story.
• Quality per euro. The decisive — and uncomfortable — one: a model that’s half as good but a sixth of the price isn’t automatically “more efficient” for your use case.

The honest verdict: a specific “6×-versus-Claude” figure is not independently established. What can be shown cleanly are the mechanisms these models use to cut compute. And those are the real story.

The two levers behind the efficiency

Lever 1: mixture-of-experts (MoE). Classic models activate all their parameters for every request. MoE models split the network into many “experts” and activate only a few per request. DeepSeek, for instance, activates only about 37 of 671 billion parameters per request. The model “knows” as much as a huge one but computes like a small one — cutting compute many-fold versus an equally large classic (“dense”) model.

Lever 2: sparse attention. The second cost driver in language models is the “attention” mechanism, which relates every chunk of text to every other — its cost grows more than proportionally with text length. DeepSeek’s approach (“DeepSeek Sparse Attention”) no longer computes every relationship, only the relevant ones. That cuts cost especially on long contexts — exactly where enterprise AI works with large documents, files or codebases.

Both levers together — plus aggressive pricing — make up the efficiency story. Neither is “magic”, and neither is exclusively Chinese; Western providers use MoE too. But Chinese labs push hard here and usually release their models as open weights.

DeepSeek, Kimi, Qwen, GLM: who “Chinese AI” actually is in 2026

Behind the blanket headline isn’t a single model but a whole field:

• DeepSeek — the efficiency front-runner. V3.2 introduced sparse attention; April 2026 brought the V4 series with a large V4 Pro (around 1.6 trillion parameters) and a lean, very cheap V4 Flash (around 284 billion). Both open-weight.
• Kimi (Moonshot), Qwen (Alibaba), GLM (Zhipu) — all strong, openly available MoE models with their own strengths depending on task and language.

The common pattern: high performance, low cost, open weights. That combination is exactly what makes them interesting for the mid-market — and raises the decisive question.

Efficiency is not the same as “best AI”

The uncomfortable truth behind every efficiency headline: cheaper doesn’t mean better. On standard workloads — classification, extraction, standard text, simple lookups — efficient models often deliver practically equivalent results at a fraction of the cost. On the hardest tasks — multi-hour autonomous agent runs, complex reasoning, demanding code migrations — Western flagship models like Claude still lead. The gap is closing, but in 2026 it’s still there.

In practice: the question isn’t “Chinese or Western” but “which workload needs which model”. The counterweight — the expensive flagship — is covered in our take on Claude Fable 5 for the mid-market, and the broader picture in our overview of AI in the mid-market.

The catch for the mid-market: data residency

This is where it gets serious for owner-led companies. If you use a Chinese provider’s official API, your data leaves the EU and is processed on servers outside European jurisdiction. For patient data, client correspondence or internal strategy documents, that is an exclusion criterion in most cases — regardless of how cheap the model is.

Add the regulatory frame: the EU AI Act has been in force since August 2024 and becomes fully applicable for most obligations on 2 August 2026. Data governance, transparency and documentation apply regardless of which country the model comes from.

The opportunity: open-weight means self-hosting

This is where the story turns. Because DeepSeek & co. publish their weights openly, you aren’t tied to the China API. The same models can run at an EU host or — depending on size and hardware — locally in-house. Then you get the same efficiency, the same low cost, but your data stays under your control.

What hardware realistically suffices, when running locally pays off and where the limits are, we break down in our piece on AI in the mid-market. The fundamental “run your own model or buy a commercial API” question is one to work through deliberately before committing.

What the mid-market should do now

Three principles:

• Read efficiency headlines as a signal, not an order. “6× more efficient” isn’t a reason to rebuild your whole AI setup — but a good prompt to review your costs and model choices.
• Decide by workload, not by origin. Cheap efficient models for volume and routine, flagship models for the hardest tasks. A model portfolio often cuts costs by more than half without hurting quality where it matters.
• Settle data residency first. Before a Chinese model goes into production: run it via EU hosting or local self-hosting, not via the China API with sensitive data.

The through-line we see in every one of these model debates: the win isn’t following the latest “model X beats model Y” video, but a sober architecture — which workload belongs on which model, and where the data is allowed to be processed.

Frequently asked questions about Chinese AI models