W.059DeepSeek says users of its Harness agent consume about 92T tokens a day, almost nine-tenths of Google’s May figure for all its products. On one modeled mix, that bills about $0.5–1B a year at its cheaper model’s list prices
DeepSeek said Harness, the agent it released on August 13, had 1.54M daily users on November 4 who consumed about 92T tokens a day, about 60M each. That is at most about 88% of the daily rate implied by Google’s May figure for all its products, which was a floor. On one modeled mix of cached input and output, it would bill about $1.4M–$2.8M a day at the list prices of DeepSeek’s cheaper Flash model, or $0.5B–$1.0B a year.
What DeepSeek said
DeepSeek published Harness’s user figures for the first time on November 5. Counting only people who use Harness with DeepSeek’s official API, it reported 1.54M daily active users and about 92T tokens consumed a day as of November 4, and said users had grown by more than a million in a month. A chart in the post showed 271K on September 30 (a seven-day average) and 429K on October 7 (method not stated); Chinese outlets describe the November figure as a single day, while Yicai Global describes all three as weekly. DeepSeek said it plans an official plugin marketplace and is preparing a support program for plugin authors, including token incentives. Bloomberg reported the figures on November 8. No outside party has audited them.
The token count is large for one product. About 92T a day is roughly 60 million tokens per daily user, the volume of an agent that rereads large amounts of code and context on each step. DeepSeek did not split it into input, output or cached tokens.
Harness reached 1.54M daily users less than three months after launch
Daily active users of DeepSeek Harness with DeepSeek’s official API, as DeepSeek reported them on November 5, 2026. September 30 is a seven-day average; October 7’s method is not stated; November 4 is reported as a single day, so the steps may mix measures.
Against other disclosed volumes
Few companies publish token counts, and none defines them the same way. ByteDance said in June that its Doubao models average 180T tokens a day. Google said in May it processes over 3.2 quadrillion tokens a month across its products, over 105T a day, and in October that its first-party model APIs handle nearly 32 billion a minute, about 46T a day. DeepSeek’s figure for one agent is 2.0 times Google’s API rate, though Google’s count is of tokens served through its APIs and DeepSeek’s may include heavily cached context.
At list prices, caching decides the bill. DeepSeek charges $0.003 per million tokens for cached input on Flash off-peak, against $0.15 for fresh input and $0.60 for output. If all 92T were cache hits, the day would bill about $0.28M; if all were output, $55M. On a modeled agent session of 95.5% cached input, 3.2% fresh input and 1.3% output, published by Capital and Compute (modeled for V4-Pro) and labelled modeled rather than metered, the blended Flash price is about $0.0153 per million tokens, which puts the 92T at $1.4M a day off-peak and $2.8M at peak. The same mix on V4-Pro would be $6.2M–$12.4M a day. DeepSeek has not said which model Harness users run, or what it earns from them.
About 92T tokens a day from one agent, against over 105T for all of Google’s products in May
Tokens a day, trillions. Each is the company’s own figure or a reported one, converted to a day: a month ÷ 30.44, a minute × 1,440, a year ÷ 365. Definitions differ (company-wide, API-only, one product), and the dates run from March to November 2026. OpenRouter’s figure is a run rate claimed by an investor and reposted by its chief executive; OpenAI’s was reported by the Wall Street Journal.
On a modeled mix, 92T tokens a day bills $1.4M–$2.8M on Flash, $6.2M–$12.4M on V4-Pro
$ million a day, at DeepSeek’s list prices on November 9, 2026, applying one modeled agent session to all 92T tokens: 95.5% cached input, 3.2% fresh input, 1.3% output. Peak hours (seven a weekday, UTC) bill double; actual billing would fall between the two. List prices only: DeepSeek has not disclosed the real mix or any discounts.
The figures behind the charts
| $ per million tokens, off-peak | Cache hit | Cache miss | Output | Blended |
|---|---|---|---|---|
| deepseek-flash | 0.003 | 0.15 | 0.6 | 0.0153 |
| deepseek-v4-pro | 0.022 | 0.66 | 1.98 | 0.0673 |
| Modeled share of tokens | 95.5% | 3.2% | 1.3% | 100% |
What this changes on 06.ai
- Token-prices pack: DeepSeek’s cached-input price, $0.003 per million off-peak, is one-fiftieth of its fresh-input price. For agent workloads the cache-hit price, not the headline price, sets the cost; worth a line on how the pack weights them.
- Compute-demand pack: a single agent at about 92T tokens a day is a new data point for per-user consumption (about 60M tokens per daily user). Worth tracking if DeepSeek publishes the figure again.
Sources and method
| Item | Figure | Source |
|---|---|---|
| DeepSeek | As of November 4, counting users of the official API: 1.54M daily active users and about 92T tokens consumed a day; about 35,000 third-party plugins; about 60% of active users made or installed one; plugin marketplace planned; v0.1 released August 13 | DeepSeek Harness team, Nov 5, via IT Home |
| DeepSeek | 271K on Sep 30, a seven-day average; 429K on Oct 7 | The Paper, Nov 6 |
| DeepSeek | Nov 4 user and token figures are for a single day; growth of more than a million in a month | Cailian Press, via Sina, Nov 6 |
| DeepSeek | Figures described as weekly | Yicai Global, Nov 6 |
| Prices | deepseek-flash: $0.003, $0.15, $0.6 per million tokens (cache hit, cache miss, output), off-peak; deepseek-v4-pro: $0.022, $0.66, $1.98; peak hours bill double | DeepSeek API docs, read Nov 9 |
| Prices | Flash peak and off-peak rates | Apidog |
| Mix | A modeled agent session of 3M cached input, 100K fresh input and 40K output tokens, labelled “modeled, not metered” | Capital and Compute |
| Over 3.2 quadrillion tokens a month across products | Google I/O, May 19, 2026 | |
| Nearly 32 billion tokens a minute through first-party model APIs | Alphabet Q3 2026 remarks, Oct 28 | |
| ByteDance | Doubao models average 180 trillion tokens a day | IT Home, Jun 23, 2026 |
| OpenRouter | 10 quadrillion annual token run rate, posted by an investor and reposted by its CEO; not published by OpenRouter | explainx.ai, Nov 2026 |
| OpenAI | 15 billion API tokens a minute by end-March, per the Wall Street Journal | AI Daily Post |
We read DeepSeek’s post through Chinese reprints; Bloomberg’s page was not available to us. Per user: 92T ÷ 1,540,000 = 59.7M. Growth: 1,540,000 ÷ 429,000 = 3.59×, a single day against a figure whose method is not stated. Per day: Google 3.2 quadrillion ÷ 30.44 = 105.1T; 32B × 1,440 = 46.1T; OpenRouter 10 quadrillion ÷ 365 = 27.4T; OpenAI 15B × 1,440 = 21.6T; DeepSeek ÷ Google all products = 88%. Mix: 3,000,000 + 100,000 + 40,000 = 3,140,000 tokens; shares 95.54%, 3.18%, 1.27%. Blended price: Σ share × price = $0.01529 (Flash) and $0.06726 (V4-Pro) per million. A day: 92,000,000 million-token units × blended price, doubled at peak: Flash, off-peak $1.41M; Flash, peak $2.81M; V4-Pro, off-peak $6.19M; V4-Pro, peak $12.38M. A year: × 365.
Wire reports are written within hours of the news, on a six-hour schedule. Each lists its sources and shows the arithmetic behind its figures. Not investment advice.