// Chapter 03 · Usage

Follow the Tokens

Which open models carry the traffic, how much the use of models changes, and how the picture changes from one platform to another.

Data: OpenRouter rankings-daily (CC BY 4.0), 1 January 2025 – 14 October 2026.

31%

of the open-weight tokens on OpenRouter from October 2025 to September 2026 went to DeepSeek’s models; Tencent (17%), Xiaomi (12%) and Z.ai (12%) followed.

OpenRouter rankings-daily (CC BY 4.0).

Illustration: layered landscape with a pale sun
One window, measured daily

From October 2025 to September 2026, DeepSeek’s models took about a third (31%) of the tokens processed by open-weight models in OpenRouter’s daily ranking. The versions themselves change much faster: none of January 2026’s ten most used versions, open or closed, was still in September’s top ten. The ranking shows intense competition for traffic.

01 Most used

DeepSeek takes a third of the open-weight tokens of the whole year

We add up the tokens processed by open-weight models in OpenRouter’s daily top 50 from October 2025 to September 2026. Of that total, DeepSeek’s models took 31.3%, and DeepSeek V4 Flash alone 18.1%.

Open-weight token consumption in OpenRouter’s ranking, October 2025 to September 2026

DeepSeek Other developers

  1. 1

    DeepSeek V4 Flash DeepSeek

    18.1% of open tokens

  2. 2

    MiMo-V2.5 Xiaomi

    9.8% of open tokens

  3. 3

    Hy3 Tencent

    7.5% of open tokens

  4. 4

    GLM 5.3 Flash Z.ai

    6.3% of open tokens

  5. 5

    Hy4 preview Tencent

    5.8% of open tokens

  6. 6

    DeepSeek V4.1 Flash DeepSeek

    5.0% of open tokens

  7. 7

    DeepSeek V4 Pro DeepSeek

    4.6% of open tokens

  8. 8

    Nemotron 3 Ultra NVIDIA

    4.6% of open tokens

  9. 9

    MiniMax M3 MiniMax

    4.5% of open tokens

  10. 10

    GLM 5.2 Z.ai

    4.0% of open tokens

Denominator: tokens processed by open-weight models in OpenRouter’s daily top 50, summed from October 2025 to September 2026 (60% of the top 50’s tokens). Tokens are traffic on one platform, not users or market share. Source: OpenRouter (openrouter.ai/rankings), as of 15 October 2026 (CC BY 4.0).

Labs stay in the lead longer than their models. DeepSeek was the most used open lab in nineteen of the twenty-one complete months since January 2025, but no single model led for more than seven months. A new version overtook the previous one from the same lab in a median of four days.

Qwen, which takes more than half of the downloads in chapter 01, took only 1.0% of these tokens.

None of January 2026’s ten most used versions was still in September’s top ten

Open-weight (verified) Not verified as open (mostly closed)

JanFebMarAprMayJunJulAugSepclaude-4.5-sonnetclaude-4.5-sonnet: rank · Jan 1, Feb 5gemini-3-flash-previewgemini-3-flash-preview: rank · Jan 2, Feb 3, Mar 5, Apr 4, May 5minimax-m2.5minimax-m2.5: rank · Feb 1, Mar 2, Apr 8kimi-k2.5kimi-k2.5: rank · Feb 2, Mar 8mimo-v2-promimo-v2-pro: rank · Mar 1, Apr 7claude-4.6-sonnetclaude-4.6-sonnet: rank · Mar 6, Apr 2, May 4, Jun 8qwen3.6-plusqwen3.6-plus: rank · Apr 1deepseek-v4-flashdeepseek-v4-flash: rank · May 2, Jun 1, Jul 3, Aug 6, Sep 9hy3-previewhy3-preview: rank · May 1, Jun 4mimo-v2.5mimo-v2.5: rank · Jun 3, Jul 1, Aug 3, Sep 7minimax-m3minimax-m3: rank · Jun 2, Jul 4, Aug 9hy3hy3: rank · Jul 2, Aug 2, Sep 10deepseek-v4-flashdeepseek-v4-flash: rank · Aug 1, Sep 5glm-5.3-flashglm-5.3-flash: rank · Sep 1hy4-previewhy4-preview: rank · Sep 2
JanFebMarAprMayJunJulAugSepclaude-4.5-sonnetclaude-4.5-sonnet: rank · Jan 1, Feb 5gemini-3-flash-previewgemini-3-flash-preview: rank · Jan 2, Feb 3, Mar 5, Apr 4, May 5minimax-m2.5minimax-m2.5: rank · Feb 1, Mar 2, Apr 8kimi-k2.5kimi-k2.5: rank · Feb 2, Mar 8mimo-v2-promimo-v2-pro: rank · Mar 1, Apr 7claude-4.6-sonnetclaude-4.6-sonnet: rank · Mar 6, Apr 2, May 4, Jun 8qwen3.6-plusqwen3.6-plus: rank · Apr 1deepseek-v4-flashdeepseek-v4-flash: rank · May 2, Jun 1, Jul 3, Aug 6, Sep 9hy3-previewhy3-preview: rank · May 1, Jun 4mimo-v2.5mimo-v2.5: rank · Jun 3, Jul 1, Aug 3, Sep 7minimax-m3minimax-m3: rank · Jun 2, Jul 4, Aug 9hy3hy3: rank · Jul 2, Aug 2, Sep 10deepseek-v4-flashdeepseek-v4-flash: rank · Aug 1, Sep 5glm-5.3-flashglm-5.3-flash: rank · Sep 1hy4-previewhy4-preview: rank · Sep 2
Month by month: the 15 versions that were first or second in some month (of the 40 that were in the top ten), with a bar over the months they were in the top ten; hover to see the rank. A version is a dated OpenRouter identifier; several September versions succeed January models of the same lab, but the data do not declare that relation. Shares of the tokens of each day’s 50 most used versions; complete months from January 2026 to September 2026. Source: OpenRouter (openrouter.ai/rankings, CC BY 4.0).

02 New arrivals

In September 2026, 3 in 4 tokens in the top 50 went to versions that had entered it in the previous 90 days

Figure 3.1

Tokens going to versions that arrived in the top 50 in the previous 90 days, by month (October 2026: 14 days)

Share of observed tokens going to versions that first reached the top 50 within the previous 90 days, by month, April 2025 to October 2026. September 2026: 76.3%.0%25%50%75%100%Apr ’25May ’25Jun ’25Jul ’25Aug ’25Sep ’25Oct ’25Nov ’25Dec ’25Jan ’26Feb ’26Mar ’26Apr ’26May ’26Jun ’26Jul ’26Aug ’26Sep ’26Oct ’26*Claude 3.7 SonnetGrok Code Fast 1MiniMax M2.5Hy3 previewMiMo-V2.5DeepSeek V4.1 Flash76%
Share of observed tokens going to versions that first reached the top 50 within the previous 90 days, by month, April 2025 to October 2026. September 2026: 76.3%.0%25%50%75%100%Apr ’25Jul ’25Oct ’25Jan ’26Apr ’26Jul ’26Sep ’26Oct ’26*Claude 3.7 SonnetGrok Code Fast 1MiniMax M2.5DeepSeek V4.1 Flash76%

Source OpenRouter (openrouter.ai/rankings), as of 15 October 2026 (CC BY 4.0)

* October 2026: measured over 14 days only.

How we measured it
Monthly bars, April 2025 to October 2026. Each bar is the share of the month’s observed tokens that went to versions in their first 90 days in the top 50, with the “other” row in the base. Hover a bar for the same share with 60- and 180-day windows: alternative definitions, not a confidence interval. Above some bars, the month’s largest new arrival, in green when it is open-weight; hover any bar for its own. Token counts use each provider’s tokenizer.

Across complete months the share runs 40%–76%; with a 60-day window 28%–64%, with 180 days 70%–91%. The three windows are alternative definitions; the 90-day one is the headline.

Source
OpenRouter rankings-daily (CC BY 4.0). Source: OpenRouter (openrouter.ai/rankings), as of 2026-10-15T16:43:23Z
Base
All tokens observed each month, including the “other” row outside the top 50
Evidence
E3our tabulation

In September 2026 the five most used versions, Z.ai’s GLM-5.3 Flash, Tencent’s Hy4 Preview, DeepSeek V4.1 Flash, OpenAI’s GPT-5.6 Luna and a DeepSeek V4 Flash snapshot dated 31 July, had all reached the top 50 since 11 July; between them they took 45% of all the month’s tokens, the ‘other’ row included.

Part of the turnover is one lab replacing a dated version with the next.

03 Staying power

Six months after entering the top 50, only 19% of the versions appear in it

This chart counts model versions. Of the 196 that first entered the top 50 between February 2025 and 22 April 2026, how many appeared in it 7, 28, 91 and 175 days later? After three months, only 38% did. Often a new version of the same model, or another model, has taken its place.

Figure 3.3

Share of versions in the top 50 on a given day after their first appearance

Source OpenRouter (openrouter.ai/rankings), as of 15 October 2026 (CC BY 4.0)

How we measured it
Presence on that day only, not continuous survival: a version may have left and returned in between.

Versions already in the top 50 in January 2025 are left out: their real first day may come before the series. Versions that arrived later, with shorter follow-up, are reported separately and give higher values.

Source
OpenRouter rankings-daily (CC BY 4.0). Source: OpenRouter (openrouter.ai/rankings), as of 2026-10-15T16:43:23Z
Base
196 versions whose first day in the top 50 fell between February 2025 and 22 April 2026, early enough to be followed for 175 days
Evidence
E3our tabulation

Entering the top 50 is easier than staying in it: most versions give way within a few months, often to the next version from the same lab, so a ranking taken today says little about the one in half a year.

Each figure is presence on that day, not continuity: a version can leave and come back, and one that leaves may still be in use, its tokens counted in the “other” row.

04 By subject

Which open models are used most in each subject

OpenRouter sorts a sample of its requests by subject. In that sample a closed model is the most used in every subject; the open-weight model with the largest share is Kimi K3, with 9.6% of the tokens on health & wellness.

The most used open-weight model in each subject, and the most used closed model, share of the subject’s tokens

Most used open-weight model Most used closed model

  1. Health & wellness

    Kimi K3 9.6%

    best closed: Claude Sonnet 4.6 14.1%

  2. Software & technology

    Kimi K3 6.1%

    best closed: Claude Opus 5.5 20.0%

  3. Finance

    Kimi K3 5.1%

    best closed: Claude Opus 5.5 11.5%

  4. Legal

    DeepSeek V4.1 Flash 3.8%

    best closed: Claude Opus 5.5 10.1%

  5. Marketing & advertising

    DeepSeek V4.1 Flash 1017 2.7%

    best closed: Claude Opus 5.5 13.7%

Show all 17 subjects
  1. Health & wellness

    Kimi K3 9.6%

    best closed: Claude Sonnet 4.6 14.1%

  2. Relationships

    GLM 5.2 7.4%

    best closed: Gemini 4 Argon 26.4%

  3. Travel & hospitality

    DeepSeek V4.1 Flash 1017 7.0%

    best closed: Claude Sonnet 5.5 10.8%

  4. Software & technology

    Kimi K3 6.1%

    best closed: Claude Opus 5.5 20.0%

  5. Business operations

    DeepSeek V4.1 Flash 1017 5.4%

    best closed: Claude Opus 5.5 11.1%

  6. Science & engineering

    DeepSeek V4.1 Flash 1017 5.2%

    best closed: GPT-6.1 Sol 12.6%

  7. Finance

    Kimi K3 5.1%

    best closed: Claude Opus 5.5 11.5%

  8. Government & politics

    Kimi K3 4.9%

    best closed: Claude Opus 5.5 13.1%

  9. Sales & commerce

    DeepSeek V4.1 Flash 1017 4.4%

    best closed: Claude Sonnet 5.5 12.2%

  10. Customer service

    GLM 5.3 4.4%

    best closed: Claude Sonnet 5.5 7.9%

  11. General

    DeepSeek V4.1 Flash 1017 4.2%

    best closed: GPT-6 Astra 7.7%

  12. Gaming

    DeepSeek V4.1 Flash 1017 4.0%

    best closed: Claude Opus 5.5 10.0%

  13. Education

    DeepSeek V4.1 Flash 1017 4.0%

    best closed: GPT-6.1 Sol 11.6%

  14. Legal

    DeepSeek V4.1 Flash 3.8%

    best closed: Claude Opus 5.5 10.1%

  15. Lifestyle

    DeepSeek V4.1 Flash 1017 3.7%

    best closed: Claude Opus 5.5 12.0%

  16. Arts & entertainment

    DeepSeek V4.1 Flash 1017 3.4%

    best closed: Claude Opus 5.5 9.2%

  17. Marketing & advertising

    DeepSeek V4.1 Flash 1017 2.7%

    best closed: Claude Opus 5.5 13.7%

Weeks of 21 September to 2 November 2026, as displayed on 9 November. Source: OpenRouter rankings, as displayed.

05 Another platform

How models move on Cloudflare

Everything above comes from one intermediary, OpenRouter, where developers call models to generate text. To check whether what it shows holds elsewhere, we look at another platform that publishes how its open models are used: Cloudflare’s Workers AI, through Cloudflare Radar. It runs the models on Cloudflare’s own network, and the answer there is different.

On Cloudflare’s Workers AI most of the work is not text generation but embeddings, which turn texts into numbers for search. One embedding model, bge-m3, took 83.7% of the inference requests Workers AI served in October 2026.

Share of Workers AI inference requests: embeddings and text generation, by month

0%25%50%75%100%Jun ’25Jan ’26Oct ’2686.7% Embeddings8.5% Text generation
0%25%50%75%100%Jun ’25Jan ’26Oct ’2686.7% Embeddings8.5% Text generation
Monthly shares of all Workers AI inference requests by task, from Cloudflare Radar’s model data; requests, not tokens; the other tasks (speech, images, translation...) make up the rest. Source: Cloudflare Radar (used with permission).
Every month, every task
Share of each month’s Workers AI inference requests, by task
MonthEmbeddingsText generationSpeech recognitionText-to-imageTranslationOther tasks
June 202538.1%31.2%15.2%7.1%6.3%2.1%
July 202570.0%17.1%8.3%2.6%1.3%0.7%
August 202569.9%16.6%7.3%4.7%0.7%0.9%
September 202534.2%41.2%12.6%6.6%4.3%1.1%
October 202541.9%28.5%17.0%5.6%5.8%1.2%
November 202538.7%28.7%22.9%1.5%6.6%1.5%
December 202543.1%26.0%25.4%2.1%2.1%1.4%
January 202632.7%27.9%32.1%1.9%3.5%1.8%
February 202635.3%27.7%31.4%2.8%1.0%1.8%
March 202639.0%33.9%20.3%1.7%1.6%3.5%
April 202647.0%26.9%17.3%4.6%0.9%3.2%
May 202633.9%10.1%52.0%1.7%0.6%1.7%
June 202675.9%12.5%7.9%0.9%1.1%1.6%
July 202682.7%9.9%4.1%0.6%1.8%0.9%
August 202684.3%8.4%3.4%0.5%2.3%1.0%
September 202684.3%9.2%4.1%0.5%0.8%1.1%
October 202686.7%8.5%2.8%0.4%0.7%0.9%

Workers AI is used above all to search: since June 2026, 76% to 87% of the requests each month turned text into embeddings, and text generation, the work OpenRouter measures, only 8% to 13%. An open model on this platform is more often a component of a search system than a chat assistant.

Radar counts requests, not tokens: embedding one sentence and writing a long answer count the same, so its shares are never added to OpenRouter’s. The data show the change in mix, not its cause: Radar publishes no accounts, so one large customer and many small ones look the same. Use has no single ranking: it depends on where it is observed and on what activity is counted.

Within text generation no model keeps the lead: five models from four labs took turns, with six changes in 17 months. And the models that lead OpenRouter are rarely the ones that lead here: of the 85 top-five places on OpenRouter since June 2025, only 6 went to models Workers AI also serves. That counts what each platform makes available, not what is popular on it.

06 A serving provider

Seen from a provider, the same versions lead and the leader changes every month

OpenRouter sends each request to one of many providers. Novita AI is one of them: an inference provider that serves open-weight models. It shared aggregated usage for this report, which lets us look at the same models from the side of a provider.

In the usage data Novita AI shared, the ten most used model IDs took 72.3% of 58.1 trillion tokens between June and November 2026, and the top three 34.6%.

The ten most used model IDs in the usage data Novita AI shared, June–November 2026, share of all tokens

Also among OpenRouter’s three most used versions in September 2026 Others

  1. 1

    DeepSeek V4.1 Flash DeepSeek

    12.2% of tokens

  2. 2

    GLM-5.2 Z.ai

    11.6% of tokens

  3. 3

    DeepSeek V4 Flash DeepSeek

    10.8% of tokens

  4. 4

    DeepSeek V4 Pro DeepSeek

    9.6% of tokens

  5. 5

    GLM-5.3 Flash Z.ai

    9.4% of tokens

  6. 6

    DeepSeek V4 Flash 0731 DeepSeek

    6.4% of tokens

  7. 7

    MiniMax M3 MiniMax

    4.4% of tokens

  8. 8

    Ling 3.0 Flash inclusionAI

    2.9% of tokens

  9. 9

    GLM-5.3 Z.ai

    2.5% of tokens

  10. 10

    MiMo V2.6 Pro Xiaomi

    2.5% of tokens

The most used model ID of each month

  1. JuneDeepSeek V4 Flash
  2. JulyGLM-5.2
  3. AugustDeepSeek V4 Flash 0731
  4. SeptemberDeepSeek V4.1 Flash and GLM-5.3 Flash · together 41.9% of the month
  5. OctoberDeepSeek V4.1 Flash and GLM-5.3 Flash · together 42.4% of the month
  6. November, partialDeepSeek V4.1 Flash 1017
Tokens are input plus output; dated versions count as separate IDs; 117 model IDs in all. One provider’s data: not market shares, users or customers. Source: Novita AI, aggregated usage and realized prices, June–November 2026 (preliminary contribution, November 2026; November partial; tencent/hy3 excluded).

The leader changed from month to month: DeepSeek V4 Flash in June, GLM-5.2 in July and DeepSeek V4 Flash 0731 in August, while DeepSeek V4.1 Flash and GLM-5.3 Flash together took 41.9% of September’s tokens and 42.4% of October’s. Both are also among OpenRouter’s three most used versions in September.

Chapter 03 · Closing

What the ranking tells us

On OpenRouter, the versions that gather the most tokens are replaced quickly. In September 2026, 76% of the tokens in the top 50 went to versions that had entered it in the previous 90 days, and six months after entering, only 19% of the versions appeared in it.

Labs last longer than their models. DeepSeek was the most used open lab in 19 of the 21 complete months since January 2025, across several releases, and the five most used versions again gathered about half of the traffic: 50% in the third quarter of 2026.

For a team that builds on these models, a ranking is a snapshot of one month. It pays to review the alternatives every few months and to be able to change model without rewriting the application.

Versions in the top 50
On OpenRouter, the versions that gather the most tokens are replaced quickly. In September 2026, 76% of the tokens in the top 50 went to versions that had entered it in the previous 90 days, and six months after entering, only 19% of the versions appeared in it.
Labs’ staying power
Labs last longer than their models. DeepSeek was the most used open lab in 19 of the 21 complete months since January 2025, across several releases.
Building on these models
A ranking is a snapshot of one month. It pays to review the alternatives every few months and to be able to change model without rewriting the application.