DeepSeek
Identity card
The company
- Name
- DeepSeek ↗
- Founded
- 2023
- Head office
- Hangzhou, China
- Chief executive
- Liang Wenfeng
The product
- Flagship model
- V4-Pro
- Subscription
- free
- API
- 1.32 $ / 3.96 $ input / output, per million tokens
The data
- Law
- China
- Storage
- China
Usage
- Web traffic, Sept. 2026
- 3.4 %
Key points
In January 2025, DeepSeek released a model on a par with the best, free and open access. Data goes to China, the app is restricted in Italy, and its safeguards give way under attack.
- Developers
- Open-weight models
- Free
- Personal or professional data
- Politically sensitive subjects
Source: Similarweb, share of visits to generative AI sites, worldwide; records published by Similarweb and repeated by the press.
Yes for a developer who wants a high-performing, inexpensive model or one to install on their own servers. Look elsewhere for any personal or professional data (Vibe, Claude).
Strengths
- 1st on OpenRouter, among developersDeepSeek is 1st on OpenRouter, the platform that routes developers' queries, with 29.8% of tokens processed in the week of 26/10/2026.Source: OpenRouter, company ranking, 26 October 2026 ↗
- Open and efficient modelsThe V4 models are published under the MIT licence and can be downloaded; the Flash version activates only 13 billion of its 284 billion parameters for each word.Source: Wikipedia, 7 November 2026 ↗ ; Huawei Central, 24 April 2026 ↗
- Top of the downloads in January 2025After the release of R1, on 20/1/2025, the app climbed to the top of the downloads; on 27/1/2025, Nvidia's share price lost nearly 17% in one session.Source: Wikipedia, 7 November 2026 ↗ ; The Kathmandu Post, citing Reuters, 28 January 2025 ↗
Weak points
- Data stored in ChinaThe privacy policy says that data is collected, processed and stored in China, and kept as long as the account exists.Source: DeepSeek, privacy policy, 10 February 2026 ↗
- Safeguards that are easy to bypassCisco researchers succeeded with 100% of their attacks against R1, against 26 to 96% for the other models tested (o1-preview, Claude 3.5 Sonnet, Gemini 1.5 Pro, GPT-4o, Llama 3.1).Source: Cisco, 31 January 2025 ↗
- Restricted in ItalyThe Italian data protection authority restricted the processing of Italians' data from 30/1/2025.Source: Garante per la protezione dei dati personali, 30 January 2025 ↗
The company DeepSeek
The DeepSeek entry →Company history
DeepSeek was born in July 2023 in Hangzhou, in eastern China, as the AI laboratory of High-Flyer, an investment fund founded by Liang Wenfeng. The company is funded by this fund.
The story continues
In January 2025, its R1 model and app went around the world: a Chinese model, released as open access, rivals the best American ones. DeepSeek has continued to publish its models as open access since.
Its V4 models, presented in April 2026, run partly on Chinese Huawei chips.
Sources: Wikipedia, 7 November 2026 ↗ ; Huawei Central, 24 April 2026 ↗
Leadership
- Founder and Chief Executive Officer
- Liang Wenfeng ↗
DeepSeek plans
- DeepSeekFree0 $
DeepSeek does not publish a price in euros: the amount paid in France depends on the exchange rate and VAT. Pricing pages, up to date as at 9 November 2026.
Content of the 1 plans, prices in France and the United States
VibeMistral Large1.50 $0.50 $
PerplexitySonar2.50 $0.25 $- DeepSeekV4-Pro3.96 $1.32 $2.6 times the output price of Vibe
Meta AIMuse Spark 1.14.25 $1.25 $
QwenQwen3.8-Max6 $2 $
GrokGrok 4.76 $2 $
GeminiGemini 4 Argon10 $2 $
ClaudeOpus 5.520 $4 $
ChatGPTGPT-6 Astra50 $10 $
Copilot: no API of its own. Flagship model of each company, standard rate, up to date as of 9 November 2026.
DeepSeek API price for 2 models
| Model | Input, $ per million tokens | Output, $ per million tokens |
|---|---|---|
| V4-Pro | 1.32 | 3.96 |
| V4.1-Flash | 0.30 | 1.20 |
Prices up to date as of 9 November 2026, read on the companies' pages, or from 2 concordant sources; the date of each record is in the page's price file Methodology.
DeepSeek models
- Company
- DeepSeek
- Models offered
- DeepSeek V4.1 Flash 1017by default, free (no known paid plan for the app); 21st of 71 for general intelligence (official page ↗)
Links read on 10/11/2026 on the company's pages, press as a fallback. Rank: Epoch AI ECI index (CC BY 4.0). Performance of AI models
Data management
How DeepSeek handles conversations
- Training models on conversationsYes: your inputs are used to train the models, according to the privacy policy.Source: DeepSeek, privacy policy, 10 February 2026 ↗
- Opting out of trainingWrite to [email protected] to exercise the right to object; the policy does not describe any setting in the app.Source: DeepSeek, privacy policy, 10 February 2026 ↗
- Retention periodAs long as the account exists.Source: DeepSeek, privacy policy, 10 February 2026 ↗
- Applicable lawChinese law: DeepSeek is a company based in Hangzhou, and data is stored in China, where the 2017 intelligence law requires cooperation with State services.Source: China Law Translate, national intelligence law, 2017 ↗
Score on this criterion: 8 out of 100, 10th of 10. Data stored in China; blocked in Italy (1/2025); transfers to a ByteDance subsidiary observed in South Korea (4/2025). The ethics ranking
Environment
Environment score
45 5th out of 10- Simple query 40 %58
- Advanced queries 20 %58
- Decarbonised electricity 20 %17
- Environmental approach 20 %33
Black line: average of the 10. The environmental ranking
Our view
DeepSeek has very efficient models, which activate only a small share of their parameters (13 billion out of 284 for Flash), and shorter answers since October 2026. Reasoning, on the other hand, is switched on by default in the API, and the servers are in China, on a grid at 525 gCO2e/kWh, 15% above the world average. The company publishes no energy figure, no report and no target.
Per-query figure published by DeepSeek
Nothing publishedDeepSeek publishes no per-query figure, either for energy or for emissions. 3 of 10 companies publish one: Google, Mistral AI and OpenAI.
Company figures cannot be compared with each other: scopes and methods differ. The "Simple query" criterion of the environmental ranking combines them with external estimates, for lack of a common basis: the methodology sets out the limits.
- Gemini0.24 Wh
Copilotnothing published- Vibe1.14 gCO2e
- Claudenothing published
- ChatGPT0.34 Wh
- Qwennothing published
- Meta AInothing published
- Perplexitynothing published
- DeepSeeknothing published
- Groknothing published
On the Éclairages blog: What is the electricity consumption of AI? ↗ ; Which AIs pollute the most? ↗
Scores and criteria
The ranking →The Celsius score is the average of 3 scores: performance, ethics and environment. Each combines several criteria. Each row places DeepSeek among the 10; clicking a criterion opens its definition and the justification for DeepSeek's score.
5 public rankings 100 %438th of 10
Definition. Synthesis of 5 public performance rankings, where models are ranked by tests or by votes from internet users: each ranking gives 100 to the best of the 10 and 0 to the last.
Justification. LMArena: 35th (rank among 419 models); Artificial Analysis: 33 (index v5.0.1); Epoch AI: 158.8 (ECI index); compar:IA: 43rd (rank among 121 models); ARC-AGI-2: 78.1 (% passed).
Content produced 30 %33joint 7th of 10
Definition. Content the assistant produces on request (sexual, misleading images, minors) and results of independent tests.
Justification. Strict written rules, but 100% of attacks succeeded (Cisco, 1/2025) and 94% of malicious queries were executed (NIST, 9/2025).
Use of data 30 %810th of 10
Definition. Use of conversations for training, retention period, storage location, sanctions by the authorities.
Justification. Data stored in China; blocked in Italy (1/2025); transfers to a ByteDance subsidiary observed in South Korea (4/2025).
Safety 20 %810th of 10
Definition. Independent safety evaluations (Future of Life Institute, SaferAI, Phare) and resistance to attacks.
Justification. Grade F from the Future of Life Institute (summer 2026), no published safety framework; its open models at least allow outside testing.
Governance 20 %17joint 9th of 10
Definition. Control of the company and deliberate choices: military uses, shareholding, commitments kept or abandoned.
Justification. Chinese law; avoids 85% of questions sensitive for Beijing (Promptfoo, 1/2025); open models under the MIT licence.
Simple query 40 %58joint 3rd out of 10
Definition. Consumption of a typical text query, on the default model served: figure published by the company, third-party estimate, failing that the size of the model.
Justification. Very sparse models (13 billion active parameters for Flash) and shortened responses in October 2026; reasoning by default in the API.
Advanced queries 20 %58joint 3rd out of 10
Definition. Weight of the resource-hungry uses the product promotes: video and image generation, agents, reasoning enabled by default.
Justification. No image or video; Expert mode that reasons at length.
Decarbonised electricity 20 %17joint 8th of 10
Definition. Share of low-carbon electricity in the data centres and efforts to increase it: local grid, on-site gas, renewable or nuclear electricity contracts.
Justification. Servers in China, grid at 525 g CO₂e/kWh, 15% higher than the world average (458).
Environmental approach 20 %33joint 6th out of 10
Definition. What the company publishes and commits to: energy and emissions figures, environmental report, targets, model optimisation.
Justification. Real efficiency and rock-bottom prices, but no energy figure, no report and no target.
Compare
or DeepSeek or QwenThe 2 Chinese families with open models, close in Celsius score: DeepSeek more used by developers.Celsius score: 35 against 39 or DeepSeek or ChatGPTDeepSeek is free and open-weight, under Chinese law; ChatGPT is more widely used and better rated on ethics.Celsius score: 35 against 62 or DeepSeek or Vibe2 alternatives to the American giants: one in China, the other in Europe.Celsius score: 35 against 40Other comparisons
DeepSeek or Claude−39DeepSeek or Gemini−36DeepSeek or Copilot−23DeepSeek or Meta AI−13DeepSeek or Perplexity−1DeepSeek or Grok+9Celsius score gap.
Ranked 9th of 10 in the assistant ranking, scores as at 11 November 2026. Facts and figures in the entry collected from 7 to 9/11/2026, each with its source; foundation and headquarters: Wikipedia, 9 November 2026 ↗.
Continue in the ranking
The full ranking →CompanyDeepSeek4models measured, of which 1 with an entryIts assistant, its models, its governance and the electricity of its data centres.The entry →ModelDeepSeekDeepSeek V4.1 Flash 101721stout of 71 on general intelligenceThe best ranked for general intelligence among the models DeepSeek offers.The entry →Comparison2 AI assistantsDeepSeek or Qwen35 against 39Celsius score out of 100The 2 Chinese families with open models, close in Celsius score: DeepSeek more used by developers.The comparison →The hub newsletter
This month's ranking, news on the impact of AI and one figure from the AI Impact Database. One newsletter a month at most, free of charge. Past issues
