DeepSeek’s share of tokens routed through OpenRouter doubled from 9% in January 2026 to 18% by June. Over the same stretch, deepseek.com traffic went the other way, falling to 346.4 million visits in July. This post covers verified DeepSeek statistics for 2026: user counts, web traffic, token share, API pricing, US government benchmark results and funding.
DeepSeek Statistics
How Many People Use DeepSeek?
QuestMobile put DeepSeek at 130 million monthly active users in China as of May 2026. That places it third among China’s AI-native apps, behind ByteDance’s Doubao at 382 million and Alibaba’s Qwen at 167 million.
The category itself reached 499 million monthly active users, up 85.4% year on year. QuestMobile counts AI-native apps separately from device-maker AI features and plugins, and the figures cover China only.
| App | Monthly active users | Period |
|---|---|---|
| Doubao (ByteDance) | 382 million | May 2026 |
| Qwen (Alibaba) | 167 million | May 2026 |
| DeepSeek | 130 million | May 2026 |
| All China AI-native apps | 499 million | May 2026 |
Source: QuestMobile, 2026 First-Half AI Application Market Development Insight Report, data as of May 2026
These are China-only app users, not global accounts. User counts for rival assistants are tracked separately in this breakdown of Grok user and revenue data.
DeepSeek Website Traffic Statistics
Similarweb recorded 346.4 million visits to deepseek.com in July 2026, down 7.25% on the month. Visitors viewed 3.37 pages per visit and stayed an average of 5 minutes 33 seconds.
The site ranked 131st globally in July 2026 and 6th in Similarweb’s Programming and Developer Software category. Its global rank was 113 three months earlier.
| Metric | Figure | Period |
|---|---|---|
| Total visits | 346.4 million | July 2026 |
| Change vs previous month | Down 7.25% | July 2026 |
| Pages per visit | 3.37 | July 2026 |
| Average visit duration | 5 min 33 sec | July 2026 |
| Bounce rate | 38.39% | July 2026 |
| Global rank | 131 | July 2026 |
Source: Similarweb estimated data, deepseek.com, July 2026
Where deepseek.com Traffic Comes From
China sent 49.27% of desktop visits to deepseek.com in July 2026. Russia followed at 9.15% and the United States at 6.08%.
Direct traffic drove 77.92% of desktop visits, with organic search second. The audience was 63.25% male and 36.75% female, and 25 to 34 year olds formed the largest age group.
| Country | Share of desktop visits |
|---|---|
| China | 49.27% |
| Russia | 9.15% |
| United States | 6.08% |
| Hong Kong | 3.24% |
| Brazil | 2.93% |
| Others | 29.32% |
Source: Similarweb estimated data, deepseek.com desktop traffic, July 2026
Similarweb lists claude.ai, chatgpt.com and gemini.google.com among the closest sites by audience affinity, with visit-level detail for the last of those in these Gemini usage and revenue figures.
DeepSeek Token Share Statistics on OpenRouter
OpenRouter logged more than 450 trillion tokens between 1 January and 14 June 2026. DeepSeek began the year just under 10% of weekly token flow, fell to 5% across February and March, then reached 18% by June.
Share here means token volume, input and output combined, not share of spend. OpenRouter notes the two diverge sharply for low-cost models.
V4 launched on 24 April 2026. By the end of May, V4-Flash accounted for 70% of DeepSeek’s agentic token flow, and DeepSeek had been the top model on OpenRouter since mid-May.
Agentic requests burn roughly 15 times more tokens than ordinary human chat, per OpenRouter. Agentic token volume overtook human token volume around 1 February 2026, which is the shift covered in these enterprise AI agent adoption figures.
Hobbyist users routed nearly a third of their tokens to DeepSeek models by early June. Chinese models as a group passed American ones in OpenRouter token share in early June 2026, after US models supplied about three quarters of tokens through 2025.
Source: OpenRouter request logs, 1 January to 14 June 2026
DeepSeek API Pricing Statistics for V4
DeepSeek lists deepseek-v4-pro at $0.435 per million cache-miss input tokens and $0.87 per million output tokens. deepseek-v4-flash is listed at $0.14 and $0.28.
Cache-hit input is billed on a separate line: $0.003625 per million tokens on V4-Pro and $0.0028 on V4-Flash. Both models carry a 1M context window and a 384K maximum output.
| Specification | deepseek-v4-pro | deepseek-v4-flash |
|---|---|---|
| Model version | V4-Pro-0813 | V4-Flash-0731 |
| Total / active parameters | 1.6T / 49B | 284B / 13B |
| Context length | 1M | 1M |
| Max output | 384K | 384K |
| Input, cache hit | $0.003625 / 1M | $0.0028 / 1M |
| Input, cache miss | $0.435 / 1M | $0.14 / 1M |
| Output | $0.87 / 1M | $0.28 / 1M |
| Concurrency limit | 500 | 2,500 |
Source: DeepSeek API Docs, Models & Pricing and V4 Preview release notes, checked 13 August 2026; parameter counts from DeepSeek’s 24 April 2026 release post
The same pricing page states that DeepSeek plans to raise overall API pricing, with a significant increase expected. No replacement rates and no effective date have been published.
Third-party endpoints price V4 differently. OpenRouter reported in June 2026 that V4-Flash on its cheapest endpoint cost $0.09 input and $0.18 output per million tokens, against $5 and $30 for GPT-5.5.
DeepSeek Benchmark Statistics vs US Frontier Models
The Center for AI Standards and Innovation, part of NIST, evaluated DeepSeek V4 Pro in April 2026 across nine benchmarks in five domains. Two were held out and non-public: ARC-AGI-2’s semi-private dataset and CAISI’s own PortBench.
CAISI concluded that V4’s capabilities lag the frontier by about 8 months, and that it performs similarly to GPT-5 rather than to the models DeepSeek’s own report benchmarks against.
| Benchmark | GPT-5.5 | Opus 4.6 | GPT-5.4 mini | DeepSeek V4 Pro |
|---|---|---|---|---|
| CTF-Archive-Diamond | 71% | 46% | 32% | 32% |
| SWE-Bench Verified | 81% | 79% | 73% | 74% |
| PortBench | 78% | 60% | 41% | 44% |
| FrontierScience | 79% | 72% | 74% | 74% |
| GPQA-Diamond | 96% | 91% | 87% | 90% |
| ARC-AGI-2 semi-private | 79% | 63% | – | 46% |
| OTIS-AIME-2025 | 100% | 92% | 90% | 97% |
| PUMaC 2024 | 96% | 95% | 93% | 96% |
| SMT 2025 | 99% | 94% | 92% | 96% |
Source: NIST CAISI, Evaluation of DeepSeek V4 Pro, accuracy on each benchmark, testing conducted April 2026
The gap is not uniform. On mathematics V4 Pro scored 97% on OTIS-AIME-2025 against 100% for GPT-5.5, but on the two held-out evaluations it fell to 44% on PortBench and 46% on ARC-AGI-2 semi-private.
CAISI’s aggregate measure separates the models more cleanly than any single test. V4 Pro scored 800 plus or minus 28 on its IRT-estimated Elo, against 1260 for GPT-5.5, 999 for Opus 4.6 and 749 for GPT-5.4 mini in April 2026 testing.
Scores for the Anthropic model in this table sit alongside wider adoption data in these Claude usage and enterprise figures.
Cost told a different story. Against GPT-5.4 mini, the only US reference model CAISI judged comparable on both price and capability, V4 cost less on 5 of 7 benchmarks, ranging from 53% less expensive to 41% more expensive.
That comparison used developer-reported prices of $1.74 input and $3.48 output per million tokens for V4 Pro. DeepSeek’s page now lists $0.435 and $0.87. Cross-model scoring methods are compared further in this review of leading model benchmark results.
DeepSeek Funding and Revenue Statistics
Caixin Global reported on 17 July 2026 that DeepSeek raised about 50 billion yuan, roughly $7.4 billion, in its first external funding round. Post-money valuation came in above 350 billion yuan.
Founder Liang Wenfeng committed 20 billion yuan of his own capital. Tencent invested 10 billion yuan and battery maker CATL 5 billion yuan, with NetEase and JD.com also participating.
| Investor | Commitment | Reported |
|---|---|---|
| Liang Wenfeng (founder) | 20 billion yuan | July 2026 |
| Tencent | 10 billion yuan | July 2026 |
| CATL | 5 billion yuan | July 2026 |
| Round total | About 50 billion yuan | July 2026 |
Source: Caixin Global, 17 July 2026
The Information reported in July 2026 that DeepSeek’s annualized revenue is approaching $500 million, with gross margins of 70% to 80% on access to V4. DeepSeek has published no revenue figures of its own.
The company has said chip supply, not capital, is its constraint. Vendor positions across accelerators and memory are tracked in these AI hardware market share statistics.
What the 2026 DeepSeek Statistics Show
Consumer web traffic is contracting. Developer token share doubled inside six months.
DeepSeek’s position rests on price against capable-enough output, which is what both the OpenRouter logs and the CAISI cost comparison measure. On CAISI’s pre-committed benchmark suite, the capability gap to the US frontier held at about 8 months as of April 2026.
FAQs
How many people use DeepSeek?
DeepSeek had 130 million monthly active users in China as of May 2026, per QuestMobile. That ranked third among China’s AI-native apps, behind Doubao at 382 million and Qwen at 167 million.
How much does the DeepSeek API cost?
DeepSeek lists V4-Pro at $0.435 per million cache-miss input tokens and $0.87 per million output tokens. V4-Flash is $0.14 and $0.28. Rates checked on DeepSeek’s pricing page on 13 August 2026.
Is DeepSeek V4 as good as GPT-5.5, as some Reddit threads claim?
Not on NIST CAISI’s suite. V4 Pro scored 800 on CAISI’s aggregate Elo measure against 1260 for GPT-5.5 in April 2026 testing, though it reached 97% on OTIS-AIME-2025 against GPT-5.5’s 100%.
Why do Reddit developers say DeepSeek is so much cheaper?
Price gaps run wide at the endpoint level. OpenRouter reported in June 2026 that V4-Flash on its cheapest endpoint cost $0.09 input and $0.18 output per million tokens, against $5 and $30 for GPT-5.5.
Is DeepSeek growing or shrinking in 2026?
Both, on different measures. Visits to deepseek.com fell 7.25% to 346.4 million in July 2026, while DeepSeek’s share of OpenRouter token volume rose from 9% in January to 18% in June.
Sources
https://www.nist.gov/news-events/news/2026/05/caisi-evaluation-deepseek-v4-pro
https://openrouter.ai/blog/insights/deepseek-v4-adoption/