Skip to content

DeepSeek Statistics: Models, API Pricing, Users and Accuracy

DeepSeek ships a new model every few months and prices its API far below rivals. It dominates China, barely registers in the US, and trails the US frontier in tests.

By Jimmy PalFounder, GTM 8020

Published 13 min read View as Markdown

DeepSeek's newest model, DeepSeek-V4-Pro, has 1.6 trillion parameters and a 1 million token context window, and its Flash API model costs $0.15 per million uncached input tokens at off-peak rates. Usage is uneven, though. Microsoft puts DeepSeek's market share at 89% in China, while only 1% of US adults used it in a given week in March 2026. Below are 34 numbered statistics on DeepSeek's models, pricing, adoption, accuracy and search engine. Each one links to a DeepSeek release, paper or pricing page, or to an original third-party study. All sources were read on 8 November 2026.

Start a free trial, no card. See which brands AI assistants name when your buyers ask for a recommendation.

What are the key takeaways?

  • DeepSeek ships fast. It released V3, R1, V3.1, V3.2, V4 and V4.1-Flash between December 2024 and October 2026, and its API now serves two models: deepseek-flash and deepseek-v4-pro.
  • Its models are big but cheap to run. V4-Pro has 1.6 trillion total parameters with 49 billion active, and the Flash model costs $0.30 per million input tokens at peak and half that off-peak.
  • Adoption is concentrated. Microsoft puts DeepSeek's market share at 89% in China, 56% in Belarus and 43% in Russia, against 1% weekly use among US adults (Epoch AI and Ipsos).
  • Independent tests are less flattering than DeepSeek's own. NIST's CAISI found V4 lags the US frontier by about 8 months, and NewsGuard's January 2025 audit gave DeepSeek an 83% fail rate on news prompts.
  • DeepSeek runs its own search engine. It says the engine handles hundreds of millions of search requests a day and filters "SEO spam, GEO manipulation, and low-quality AIGC".

What models has DeepSeek released?

DeepSeek has shipped nine model releases since December 2024, from V3 and R1 through V4 and V4.1-Flash, roughly one every two to three months. Its API now runs two models: deepseek-flash, served by V4.1-Flash-1017, and deepseek-v4-pro, served by V4-Pro-0813. Older model names have been retired.

Date Release What changed
26 Dec 2024 DeepSeek-V3 671B MoE parameters, 37B active, 14.8T training tokens
20 Jan 2025 DeepSeek-R1 Reasoning model released under the MIT License
28 May 2025 DeepSeek-R1-0528 AIME 2025 score rose from 70.0 to 87.5
21 Aug 2025 DeepSeek-V3.1 One model with thinking and non-thinking modes
1 Dec 2025 DeepSeek-V3.2 and V3.2-Speciale First DeepSeek model to integrate thinking into tool use
24 Apr 2026 DeepSeek-V4 Preview V4-Pro and V4-Flash with 1M token context
13 Aug 2026 DeepSeek-V4-Pro GA Peak and off-peak API pricing introduced
10 Sep 2026 DeepSeek-V4.1-Flash 552B parameters, native image understanding
17 Oct 2026 DeepSeek-V4.1-Flash-1017 Benchmark update and public Search and Fetch APIs

Sources: DeepSeek API change log and the release notes linked below.

  1. DeepSeek-V3 has 671 billion parameters, 37 billion active per token, and was trained on 14.8 trillion tokens. It generated 60 tokens a second, three times faster than V2. Source: DeepSeek, Introducing DeepSeek-V3, 26 December 2024.
  2. DeepSeek-V3's full training took 2.788 million H800 GPU hours. Pre-training on 14.8 trillion tokens used 2.664 million of them. Sources: DeepSeek-V3 technical report, arXiv and DeepSeek-V3 on GitHub.
  3. DeepSeek-R1 launched with 6 distilled small models, and DeepSeek said the 32B and 70B versions were on par with OpenAI-o1-mini. Source: DeepSeek, R1 release, 20 January 2025.
  4. The R1 paper was published in Nature on 17 September 2025. Source: Nature, DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning.
  5. R1-0528 lifted R1's AIME 2025 score from 70.0 to 87.5 and its GPQA score from 71.5 to 81.0. Source: DeepSeek API change log, 28 May 2025.
  6. DeepSeek-V3.2 was trained on synthesized agent data covering more than 1,800 environments and 85,000 complex instructions. DeepSeek says V3.2-Speciale reached gold-level results at IMO, CMO, the ICPC World Finals and IOI 2025. Source: DeepSeek, V3.2 release, 1 December 2025.

How big is DeepSeek-V4?

DeepSeek-V4-Pro has 1.6 trillion total parameters with 49 billion active per token, and V4-Flash has 284 billion with 13 billion active. Both were pre-trained on more than 32 trillion tokens and support 1 million tokens of context, which DeepSeek made the default across all of its official services.

  1. V4-Pro: 1.6 trillion total parameters, 49 billion active. V4-Flash: 284 billion total, 13 billion active. Source: DeepSeek, V4 Preview release, 24 April 2026.
  2. Both V4 models were pre-trained on more than 32 trillion tokens. Source: DeepSeek-V4-Pro model card, Hugging Face.
  3. At 1 million tokens of context, V4-Pro needs 27% of the single-token inference FLOPs and 10% of the KV cache that DeepSeek-V3.2 needed. Source: DeepSeek-V4-Pro model card, Hugging Face.
  4. 1 million tokens is now the standard context length across all official DeepSeek services. Source: DeepSeek, V4 Preview release.
  5. V4.1-Flash is a 552 billion parameter model that activates 8 billion parameters for input and 16 billion for output. Its KV cache needs a quarter of the HBM and an eighth of the SSD storage of the previous generation. Source: DeepSeek, V4.1-Flash release, 10 September 2026.
  6. V4.1-Flash-1017 scores 42.6 on Humanity's Last Exam without tools and 77.0 on DeepSWE v1.1, up from 36.8 and 74.2 for the September V4.1-Flash release. Source: DeepSeek API change log, 17 October 2026.
  7. The legacy model names deepseek-chat and deepseek-reasoner were retired after 24 July 2026. Source: DeepSeek, V4 Preview release.

How much does the DeepSeek API cost?

DeepSeek's Flash model costs $0.30 per million input tokens and $1.20 per million output tokens at peak, and V4-Pro costs $1.32 and $3.96. Off-peak rates are half of peak, a scheme DeepSeek introduced in August 2026. Its new Search and Fetch APIs cost $0.75 per 1,000 successful requests each.

Price per 1M tokens deepseek-flash peak deepseek-flash off-peak deepseek-v4-pro peak deepseek-v4-pro off-peak
Input, cache hit $0.006 $0.003 $0.044 $0.022
Input, cache miss $0.30 $0.15 $1.32 $0.66
Output $1.20 $0.60 $3.96 $1.98

Source: DeepSeek, Models and Pricing, read 8 November 2026.

  1. Off-peak rates are half of peak rates. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday. Weekends and Chinese public holidays are off-peak in full. Source: DeepSeek, Models and Pricing.
  2. Both API models support a 1 million token context and up to 384,000 output tokens. Concurrency limits are 2,500 for deepseek-flash and 500 for deepseek-v4-pro. Source: DeepSeek, Models and Pricing.
  3. The Search and Fetch APIs cost $0.75 per 1,000 successful requests each. Source: DeepSeek, Models and Pricing.
  4. At launch in January 2025, R1 cost $0.55 per million input tokens on a cache miss and $2.19 per million output tokens. Source: DeepSeek, R1 release.
  5. From 8 February 2025, V3 cost $0.27 per million input tokens on a cache miss and $1.10 per million output tokens. Today's off-peak Flash price for cache-miss input, $0.15, is 44% lower. Source: DeepSeek, Introducing DeepSeek-V3.

Check whether ChatGPT, Perplexity, Claude and Gemini name your brand. Start a free trial, no card.

How many people use DeepSeek?

The clearest adoption figures come from third-party studies. Microsoft's telemetry-based research gives DeepSeek an 89% market share in China, 56% in Belarus and 43% in Russia. In the US, an Epoch AI and Ipsos survey found 1% of adults used DeepSeek in the past week, against 31% for ChatGPT.

  1. DeepSeek's market share is 89% in China, 56% in Belarus, 49% in Cuba and 43% in Russia. Source: Microsoft AI Economy Institute, Global AI Adoption in 2025, January 2026.
  2. DeepSeek usage in Africa is 2 to 4 times higher than in other regions. Microsoft links this to the free chatbot, the MIT-licensed model and promotion by partners such as Huawei. Source: Microsoft AI Economy Institute, January 2026.
  3. 1% of US adults used DeepSeek in the past week in a survey of 2,021 adults fielded 3 to 6 March 2026. Source: Epoch AI and Ipsos, National AI Usage Survey topline.
AI service US adults who used it in the past week
ChatGPT 31%
Google Gemini 21%
Microsoft Copilot 11%
Meta AI 8%
Grok 5%
Claude 3%
Perplexity 2%
DeepSeek 1%

Source: Epoch AI and Ipsos, KnowledgePanel survey of 2,021 US adults, 3 to 6 March 2026. 49% had not used any AI service in the past week.

  1. DeepSeek jumped from outside the top 20 to #3 in Cloudflare's generative AI ranking between 28 January and 3 February 2025, then settled between #6 and #10 for the rest of the year. Source: Cloudflare Radar 2025 Year in Review, internet services.
  2. DeepSeek finished #10 in Cloudflare's 2025 year-end generative AI ranking, and performed better on weekends than weekdays, as did Grok and Perplexity. Source: Cloudflare Radar 2025 Year in Review.
  3. Downloads of DeepSeek models on model-sharing platforms rose nearly 1,000% between January 2025 and NIST's September 2025 evaluation. Source: NIST, CAISI evaluation of DeepSeek AI models, 30 September 2025.
  4. DeepSeek-V4.1-Flash was downloaded 1,842,868 times in the last month, and V4-Pro 335,063 times, according to Hugging Face counters read on 8 November 2026. Sources: DeepSeek-V4.1-Flash and DeepSeek-V4-Pro on Hugging Face.

For how DeepSeek compares with the wider market, see our AI search adoption statistics.

How accurate and safe is DeepSeek?

Independent testers rate DeepSeek below the US leaders. NIST's CAISI found DeepSeek V4 trails the frontier by about 8 months, even though it is the most capable Chinese model CAISI has tested. NewsGuard's January 2025 audit gave DeepSeek an 83% fail rate on news prompts, and it ranked tied for 10th of 11 chatbots.

  1. DeepSeek V4's capabilities lag the US frontier by about 8 months, and it is the most capable model from China that CAISI has evaluated. Source: NIST, CAISI evaluation of DeepSeek V4 Pro, 1 May 2026.
  2. DeepSeek V4 was more cost efficient than GPT-5.4 mini on 5 of 7 benchmarks, ranging from 53% less expensive to 41% more expensive. Source: NIST, CAISI evaluation of DeepSeek V4 Pro.
CAISI benchmark GPT-5.5 Opus 4.6 GPT-5.4 mini DeepSeek V4 Pro
SWE-Bench Verified 81% 79% 73% 74%
GPQA-Diamond 96% 91% 87% 90%
ARC-AGI-2 semi-private 79% 63% – 46%
PortBench 78% 60% 41% 44%

Source: NIST CAISI, 1 May 2026.

  1. DeepSeek's most secure model in 2025, R1-0528, complied with 94% of overtly malicious requests when a common jailbreak was used, against 8% for US reference models. Source: NIST, CAISI evaluation, 30 September 2025.
  2. Agents built on R1-0528 were 12 times more likely than US frontier models to follow hijacking instructions. Source: NIST, CAISI evaluation, 30 September 2025.
  3. DeepSeek models echoed 4 times as many inaccurate and misleading CCP narratives as US reference models. Source: NIST, CAISI evaluation, 30 September 2025.
  4. DeepSeek failed 83% of NewsGuard's news and information prompts in January 2025, finishing tied for 10th of 11 chatbots. Source: NewsGuard, 29 January 2025.

Do people trust DeepSeek?

Not on balance. In the Reuters Institute's 2026 survey of nine countries, DeepSeek sat in the "net distrust" group with Meta AI, Snapchat My AI and Grok, where slightly more people distrust it than trust it. The institute also found that use and awareness of DeepSeek "seem to have stalled" since early 2025.

  1. DeepSeek is "net distrusted", with a low share of trusters and a slightly larger share of distrusters, while ChatGPT, Gemini, Copilot and Claude are trusted on balance. Source: Reuters Institute, Generative AI and News Report 2026, 29 October 2026.
  2. Use and awareness of DeepSeek "seem to have stalled" after its early 2025 surge, while average weekly ChatGPT use rose from 22% to 34% across the six countries with earlier data. Source: Reuters Institute, 29 October 2026.

For the full picture on trust across engines, our LLM citation statistics show where each assistant gets its answers.

How does DeepSeek search the web?

DeepSeek's web search runs on an in-house search engine that DeepSeek built itself. DeepSeek says the engine indexes hundreds of billions of pages, handles hundreds of millions of requests a day and filters "SEO spam, GEO manipulation, and low-quality AIGC". Since October 2026 developers can query it directly through a paid Search API.

  1. DeepSeek's in-house engine "reliably handles hundreds of millions of search requests from around the world every day" and "indexes hundreds of billions of web pages". Source: DeepSeek API change log, 17 October 2026.

The same note says the engine filters "SEO spam, GEO manipulation, and low-quality AIGC through strict detection mechanisms". The Search API returns 1 to 50 results per query and supports operators such as site: and -site:. Source: DeepSeek, Search and Fetch guide.

For brands, that line matters. DeepSeek is telling marketers it will discount pages built only to game AI answers. The pages most likely to be cited are the ones that state facts plainly and are referenced by other sites. Our AI crawler statistics cover how AI bots read and fetch pages.

What do these numbers mean for brands?

DeepSeek matters most where it dominates: China, Russia, Belarus, parts of Africa, and developer tools that run its cheap API. For US buyers it is a minor consumer channel at 1% weekly use. Brands selling into its strong markets should test it; most others should focus on ChatGPT, Gemini and Google first.

  • Know where your buyers are. An 89% share in China against 1% weekly use in the US means DeepSeek's importance depends almost entirely on geography.
  • Expect DeepSeek inside other products. At $0.15 per million uncached input tokens off-peak, its models are cheap to build into other tools, and DeepSeek's docs list integrations with coding agents such as Claude Code, Codex and OpenCode.
  • Do not try to game it. DeepSeek says its search filters GEO manipulation. Clear, sourced pages and third-party coverage are the safer route.
  • Check the engines your buyers use most. Here is how to check if ChatGPT mentions your brand.

Where does Mentionova fit, and where does it not?

Mentionova tracks brand mentions and citations in ChatGPT, Perplexity, Claude, Gemini, Google AI Overviews and Reddit. DeepSeek is not one of those engines today. Brands whose buyers rely on DeepSeek should test it by hand; brands whose buyers use the larger US and European assistants can track those in Mentionova.

What the live pricing page shows on 8 November 2026:

  • Starter costs $99 a month billed annually, or $125 month to month, with 25 tracked prompts, 100 agentic tasks a month and 3 answer engines.
  • Scale costs $320 a month billed annually, or $400 month to month, with 100 tracked prompts, 500 agentic tasks and 6 answer engines.
  • Enterprise costs $799 a month billed annually, or $999 month to month, with 250 tracked prompts, 2,500 agentic tasks and 6 answer engines.
  • Every paid plan has a free trial, no card, and needs no pixel, code or tag.

Where it does not fit: if most of your buyers are in China or Russia, DeepSeek and local assistants will matter more than any engine Mentionova covers.

Where does each number come from?

Every statistic above links to its source. DeepSeek figures come from its API documentation, release notes, change log and pricing page at api-docs.deepseek.com, its Hugging Face model cards, the V3 paper on arXiv and the R1 paper in Nature. Third-party figures come from NIST, Microsoft, Epoch AI and Ipsos, Cloudflare, NewsGuard and the Reuters Institute. All were read on 8 November 2026.

What do people ask about DeepSeek?

What is the latest DeepSeek model?

As of 8 November 2026, the DeepSeek API serves DeepSeek-V4.1-Flash-1017 under the name deepseek-flash, released on 17 October 2026, and DeepSeek-V4-Pro-0813 under deepseek-v4-pro. V4.1-Flash is a 552 billion parameter model with native image understanding.

How many parameters does DeepSeek have?

It depends on the model. V3 has 671 billion parameters with 37 billion active. V4-Pro has 1.6 trillion with 49 billion active, V4-Flash has 284 billion with 13 billion active, and V4.1-Flash has 552 billion with 8 billion active for input and 16 billion for output.

How much does DeepSeek cost?

The chat app is free. The API charges per million tokens: deepseek-flash costs $0.30 input and $1.20 output at peak, and deepseek-v4-pro costs $1.32 and $3.96. Off-peak rates are half of those.

Is DeepSeek open source?

DeepSeek publishes open weights and technical reports for its models. R1 was released under the MIT License, and the V4-Pro weights on Hugging Face carry an MIT license tag.

How many people use DeepSeek?

The best figures come from third parties. Microsoft puts DeepSeek's market share at 89% in China and 43% in Russia. In the US, 1% of adults used it in a given week in March 2026, according to Epoch AI and Ipsos.

Can brands influence what DeepSeek recommends?

Its web answers draw on DeepSeek's own search engine, which DeepSeek says filters SEO spam and GEO manipulation. That rules out shortcuts. The reliable route is clear, factual pages on your own site plus coverage on independent sites that cover your category.

Why does AI recommend my competitors and not me?

Usually because the pages and publications AI engines cite mention your competitors and not you. Check which sources each engine uses for your buying questions, then close the gaps. Our Gemini statistics show how much each engine relies on brand-owned pages.

Find out which AI engines name your brand today. Start a free trial, no card.

100+ agents on call · 24×7

Put the agentsto work.

100+ growth agents run your competitor research, watch six answer engines, write content briefs and build the reports — around the clock. You keep final approval.