In January 2025 DeepSeek-R1 shook the markets because it came within a few weeks of OpenAI's best reasoning model. Almost two years later the picture is more nuanced. On Epoch AI's Capabilities Index, China's strongest model, Moonshot's Kimi K3.1 Pro, is about six months behind the US frontier, and the gap has widened rather than closed. At the same time, Chinese labs now build almost all of the best open-weight models and handle 69% of the tokens on OpenRouter. For a company in Croatia, Germany or Austria the real questions are practical: are DeepSeek, Qwen or Kimi good enough, are they really cheaper, and can you use them under GDPR? This guide answers all three with data.
The frontier race in one chart
Epoch AI's Capabilities Index (ECI) combines dozens of benchmarks into one score, anchored so that Claude 3.5 Sonnet equals 130 and GPT-5 equals 150. The chart below shows, for each country, the best score of any model released up to that date. The US frontier is now Claude Fable 5.2 at 173.4 (9 November 2026), ahead of Google's unreleased Gemini 4 Carbon (169.8) and OpenAI's GPT-6 Astra (169.0). China's best is Kimi K3.1 Pro at 161.4. Europe's best, Mistral Large 4 from France, reaches 154.6.
Source: Epoch AI, Capabilities & benchmarking data (ECI scores, downloaded 14 November 2026). Value = highest ECI of any model from that country released up to the end of each quarter. Includes unreleased and limited-access US models listed by Epoch.
Two things stand out. First, the shape: China caught up dramatically in late 2024 and early 2025 (Qwen 2.5, then DeepSeek-R1), but since then the US frontier has moved faster, driven by the Claude 5 family, GPT-6 and Gemini 4. Second, Europe: Mistral spent most of 2025 at the level of Mistral Medium 3, and only Mistral Large 4 in October 2026 brought it back into the conversation.
How big is the gap in months?
A more intuitive way to read the data is to ask: how long ago did the US first reach the score of China's best model? When DeepSeek-R1 launched, the answer was about one month (OpenAI's o1). Today Kimi K3.1 Pro matches GPT-5.5 Pro from April 2026, a gap of six months. Mistral Large 4 matches GPT-5.2 Pro from December 2025, a gap of almost eleven months.
Source: own calculation from Epoch AI ECI scores. Months between the release of the non-US model and the release of the first US model with an equal or higher ECI.
Six months sounds small, but in 2026 it is the difference between a model that can run multi-hour agentic tasks reliably and one that cannot. Our article on AI agent time horizons shows how fast that capability doubles. For chat, summarisation, extraction and translation, however, a six-month-old frontier level is more than enough.
Where China leads: open weights
The picture flips when you look only at models whose weights you can download and run yourself. Of the eight strongest open-weight models in Epoch's data, six are Chinese: two Kimi versions, three DeepSeek versions and GLM-5.3 from Z.ai. The best US open model, Meta's Muse Spark Open (155.4), and Mistral Large 4 (154.6) come in sixth and seventh. OpenAI's gpt-oss models are far behind. Of the 32 Chinese models Epoch scored in 2026, 23 have open weights; of the 52 US models, only 7 do.
Source: Epoch AI ECI scores, November 2026, models classed as open weights. Licences differ: DeepSeek V4.x and GLM-5.3-Flash are unrestricted; Kimi K3.1 Pro, GLM-5.3, Muse Spark Open and Mistral Large 4 have restricted-use licences; Kimi K3 is non-commercial.
Note the licence column. "Open weights" does not always mean "free for commercial use". Kimi K3 was released under a non-commercial licence, and several others carry use restrictions. Read the licence before you build a product on a model. We cover the trade-offs in open-weight vs closed models.
Usage: developers already voted
Developers who pay per token have moved fast. OpenRouter, the largest model marketplace, reported in September 2026 that open-weight models accounted for 72% of weekly tokens and Chinese labs for 69%; the two shares almost overlap. Labs outside the US and China, Mistral included, fell from almost 12% of tokens at the start of 2025 to 2%. OpenRouter also says over 40% of US startups on its platform use DeepSeek in production, up from about 13% in January 2026.
Source: OpenRouter, "The token state of play" newsletter (22 September 2026). Token share is not spend share: in spend, DeepSeek, Moonshot and Z.ai together are in the high single digits.
Keep this in perspective. Token share is dominated by high-volume, low-price workloads such as coding agents and bulk processing. In spend, OpenRouter says DeepSeek, Moonshot and Z.ai together reach only the high single digits. Enterprise budgets still go mostly to Anthropic and OpenAI, as our analysis of enterprise LLM market share shows.
Cheap tokens are not the same as cheap tasks
Chinese models are famous for low prices. DeepSeek's API lists V4.1 Flash at $0.30 per million input tokens and $1.20 per million output tokens at peak time, half that off-peak. Claude Opus 5.5 costs $4 and $20.
| Model | Input / 1M tokens | Output / 1M tokens |
|---|---|---|
| DeepSeek V4.1 Flash (peak / off-peak) | $0.30 / $0.15 | $1.20 / $0.60 |
| DeepSeek V4 Pro (peak / off-peak) | $1.32 / $0.66 | $3.96 / $1.98 |
| Claude Sonnet 5.5 | $2 | $10 |
| Claude Opus 5.5 | $4 | $20 |
| Claude Fable 5.2 | $8 | $40 |
Sources: api-docs.deepseek.com pricing (peak = 01:00–04:00 and 06:00–10:00 UTC on weekdays) and platform.claude.com pricing, November 2026. Standard rates without caching or batch discounts.
But a model that needs more tokens, more retries or more reasoning to finish a job is not cheaper in practice. Artificial Analysis measures the cost of running its whole benchmark suite per task. On that measure, OpenAI's GPT-6.1 Sol (high) scores 50 for $0.59 per task, while DeepSeek V4.1 Flash at maximum reasoning scores 33 for $0.60. Kimi K3.1 Pro, the strongest Chinese model, scores 46 at $3.68, more than Claude Opus 5.5 (high) at $3.46 for a score of 52.
Source: Artificial Analysis model leaderboard (Intelligence Index and cost per task, retrieved 14 November 2026). Selected models and reasoning levels. The Intelligence Index uses a different benchmark mix than Epoch's ECI, so absolute values are not comparable.
The lesson for buyers: compare models on cost per successful task in your own workload, not on list price per token. Our guides on model routing and inference economics show how to measure it.
The European question: GDPR, security and trust
For EU companies the hardest question is not performance but data. There are two very different ways to use a Chinese model:
- The vendor's app or API. Prompts and files go to servers in China. Italy's data protection authority (Garante) blocked the DeepSeek app on 30 January 2025, and in June 2025 Berlin's data protection commissioner reported it to Apple and Google as illegal content under the Digital Services Act because of unlawful transfers to China. Greece's authority ordered DeepSeek to appoint an EU representative. For personal data, client data or trade secrets this route is very hard to justify.
- Open weights on your own or EU-hosted infrastructure. If you run DeepSeek, Qwen or GLM weights on servers in the EU, no data leaves your control. The remaining issues are the licence, supply-chain security of the model files, and model behaviour: Chinese models are trained to follow Chinese content rules on political topics, which matters for public-facing chat but rarely for document processing or code.
Access also runs the other way. Since September 2025 Anthropic has refused service to companies more than 50% owned by entities in China and other unsupported regions, regardless of where they operate, and OpenAI does not serve China either. For a Croatian or German subsidiary of a Chinese group, that can rule out Claude entirely. Manus, the agent company founded by a Chinese team and now based in Singapore, shows how much corporate structure matters in this market.
Our rule of thumb for EU companies: use frontier closed models (Claude, GPT, or Mistral for EU-hosted needs) via EU-region endpoints for anything involving personal or confidential data and complex agentic work. Consider Chinese open-weight models only self-hosted or through an EU provider, for high-volume, low-risk tasks where cost per task is proven lower in your own tests. Never paste client data into a consumer app hosted in China.
What this means for companies in Croatia and DACH
- The frontier is American, and the gap is growing. For demanding agentic and reasoning work, Claude Fable 5.2, Claude Opus 5.5 and GPT-6 are 6–12 months ahead of everything else.
- Open weights are Chinese, with Mistral as the European exception. Mistral Large 4 is the only non-Chinese model in the open-weight top seven besides Meta's, and the natural choice when data sovereignty is a hard requirement.
- Price per token is a trap. Measure cost per finished task. The cheapest token often isn't the cheapest result.
- Compliance decides first. Under GDPR and the EU AI Act, where data is processed matters more than benchmark points.
The bottom line
China has not overtaken the US in AI, and on Epoch's data it is falling slightly further behind at the frontier. But it has won the open-weight race and much of the developer token volume. For European businesses the sensible strategy is a portfolio: frontier models from Anthropic and OpenAI for complex and sensitive work, Mistral where EU hosting is required, and carefully self-hosted open-weight models, Chinese or not, where tests show a real cost advantage.
Which models belong in your AI stack?
AI Workshop benchmarks Claude, GPT, Mistral and open-weight models on your own tasks, calculates cost per task, and checks GDPR and EU AI Act fit before you commit. Our Claude Certified Architects build the routing and hosting setup that keeps sensitive data in the EU.
Talk to an AI consultant