03 Evidence & Methods

Follow the Tokens

Where each chart’s numbers come from, in one card per chart. The full calculations are in the files named on each card.

The charts

  1. The most used open-weight models over the year

    What it shows
    The ten open-weight models with the most tokens in OpenRouter’s daily top 50, October 2025 to September 2026, and each developer’s share.
    Source
    OpenRouter daily rankings (CC BY 4.0), extract of 15 October 2026; our open-weight registry.
    How we calculated it
    Tokens of each model (its dated versions together) summed over twelve months, as a share of the open-weight tokens in the top 50.
    Main limit
    Open-weight only from the day the weights were public; tokens outside the top 50 cannot be attributed; not users or market share.
    Files
    analysis/tokens/by_model/models_monthly.csv · analysis/tokens/by_model/open_leader_tenure.csv · pipelines/web/summary_2027.py
  2. New versions in the top 50

    What it shows
    The share of each month’s tokens that went to versions that entered the top 50 in the previous 90 days.
    Source
    OpenRouter daily rankings (CC BY 4.0).
    How we calculated it
    Daily top-50 tokens by version, each version dated by its first day in the top 50; the “other” row stays in the denominator. 60- and 180-day windows and a count by product line as checks.
    Main limit
    The first day in the top 50 is not a release date; the data do not say why a version gains traffic.
    Files
    analysis/tokens/turnover_monthly.csv · data/processed/openrouter/daily_rankings.csv
  3. Six months later

    What it shows
    How many versions are in the top 50 on day 7, 28, 91 and 175 after their first appearance.
    Source
    OpenRouter daily rankings (CC BY 4.0).
    How we calculated it
    Versions followed from their first day in the top 50, among those that arrived early enough to be observed for 175 days.
    Main limit
    Presence on a given day, not continuity; a version that leaves may still be used, inside the “other” row.
    Files
    analysis/identity/persistence_audit.csv
  4. By subject

    What it shows
    For each of 17 subjects, the most used open-weight model and the most used closed model, with their shares.
    Source
    OpenRouter rankings page, ‘Domain’ section, as displayed on 9 November 2026.
    How we calculated it
    Shares as displayed for the models in each subject’s list.
    Main limit
    OpenRouter classifies a sample of requests and does not publish how it is drawn.
    Files
    analysis/tokens/by_domain/domains_2026_11_09.csv · analysis/tokens/by_domain/domains_open_share_2026_11_09.csv
  5. Another platform: Workers AI

    What it shows
    The share of Workers AI runs by task (embeddings, text generation) and by model.
    Source
    Cloudflare Radar, Workers AI (with Cloudflare’s permission).
    How we calculated it
    Monthly shares of inference events by task and by model, as published by Radar.
    Main limit
    Inference events, not tokens; one platform; no accounts or volumes.
    Files
    data/processed/cloudflare-s11/monthly_models.csv

More data

Figures from the research that the chapter’s charts do not show.

  • Over the twelve months from October 2025 to September 2026, open-weight models took 59.9% of the tokens in OpenRouter’s daily top 50.
  • Hugging Face downloads by when the model was first seen, read on 1 October 2026: from 1 August 2026, 8.4% (101 models); between 1 November 2025 and 1 July 2026, 35.9% (371); on 1 October 2025 or earlier, 55.7% (1,204).

Sources and notes

Show the 21 notes

Evidence levels: E1 = figure published by the source; E2 = recomputed by us and matching a figure the source publishes; E3 = computed by us from the source data, with no published counterpart (reproducible pipeline and tests). All OpenRouter figures: Source: OpenRouter (openrouter.ai/rankings), as of 2026-10-15T16:43:23Z. Token counts are as reported by each provider and use different tokenizers. Shares are of all observed tokens, including the "other" row, unless stated otherwise.

  1. 1
    Computed from data/processed/openrouter/daily_rankings.csv with pipelines/tokens/reassess.py (load). Monthly leader by version and its share of all observed tokens: Claude 3.5 Sonnet 46.7% (2025-01) and 33.1% (2025-02); then Claude 3.7 Sonnet (2025-03, 2025-04), GPT-4o mini (2025-05), Claude Sonnet 4 (2025-06 to 2025-08) and Grok Code Fast 1 (2025-09 to 2025-12). Leaders in 2026, from Anthropic, MiniMax, Xiaomi, Qwen, Tencent, DeepSeek and Z.ai, held 7.4–13.7% in complete months (above 12.5% only in 2026-07, xiaomi/mimo-v2.5-20260422, 13.7%); a different version led in each month of 2026. Partial month 2026-10 (14 days): deepseek/deepseek-v4.1-flash-20260910, 20.7%. The five most-used versions of 2026-09, with their first day in the observed top 50 and their share of the month's tokens: z-ai/glm-5.3-flash-20260826 (2026-08-26, 9.9%), tencent/hy4-preview-20260827 (2026-08-28, 9.7%), deepseek/deepseek-v4.1-flash-20260910 (2026-09-10, 8.9%), openai/gpt-5.6-luna-20260709 (2026-07-11, 8.9%), deepseek/deepseek-v4-flash-20260731 (2026-07-31, 7.6%); 45.0% together. Distinct developers with a version in the top 50: 23 (2025-Q1), 27 (2026-Q3); leading developer share 39.3% (Anthropic, 2025-Q1), 20.8% (DeepSeek, 2026-Q3). E3. Ten most-used versions of 2026-09 (subsection The Ten Most-Used Versions in September 2026): pipelines/tokens/top_versions.py, analysis/tokens/top_versions_2026_09.csv (tokens, share_of_top50, share_of_all_observed, first_day_top50, age_days_on_sep30, effective class); together 0.67 of top-50 tokens and 0.62 of all observed tokens. analysis/tokens/top10_monthly_2026.csv: no version of the 2026-01 top ten in the 2026-09 top ten; 40 distinct versions in the monthly top tens of 2026-01 to 2026-09. E3.
  2. 2
    OpenRouter, rankings dataset (rankings-daily), CC BY 4.0, retrieved 2026-10-15; dataset description notes that token counts use each model's own tokenizer. Raw extracts and hashes in data/manifests/openrouter_2026-10-15.json. E1 (description and licence).
  3. 3
    analysis/tokens/by_model/models_monthly.csv, months 2025-10 to 2026-09, tokens of versions with an open-weight class on the day of use summed by product line (pipelines/web/summary_2027.py, d03); developers' shares and tenure in analysis/tokens/by_model/open_leader_tenure.csv (DeepSeek: 19 months at the top among open-weight developers; longest run of a single product line: 7 months).
  4. 4
    analysis/identity/v09r_monthly.csv, scenario pit_strict, 2026-07 to 2026-09 pooled: OW-P 41.6%, OW-R 0%, OW-N 21.2%, CW 0%, unknown 31.0%, other 6.2%. Pass 26: Kimi K3 counts as open from 2026-07-27, the day its weight files were first published (Hugging Face commit history); the first archived copy of its repository, on 2026-07-23, was an "Upcoming release" page without weights. With the earlier date the floor was 62.9% (independent review C-09). analysis/tokens/free_sensitivity_matched.csv: removing :free rows from numerator and denominator ("other" kept) gives 62.8% → 59.8% under pit_strict and 63.2% → 60.0% under current classes, with the same method. Free share of OW-N tokens 32.7% and of OW-P tokens 4.4%: current classes, analysis/tokens/free_sensitivity.csv. E3. DEC-076, DEC-080.
  5. 5
    analysis/tokens/openness_bounds.csv: upper bound 1.0 and confirmed closed share 0.0 in every month; the nine recorded CW decisions (DEC-093, DEC-094) are suspended pending revalidation (DEC-099); no later lower bound exceeds any earlier upper bound. E3. DEC-019, DEC-024.
  6. 6
    analysis/tokens/turnover_monthly.csv (level version) and analysis/tokens/turnover_checks.csv (column headline). Age = days since a version's first day in the observed top 50 from 2025-01-01; recent if 0–90 days inclusive; months 2025-01 to 2025-03 censored. Range 39.8% (2026-01) to 76.3% (2026-09) for complete months; 80.7% for 2026-10 (14 days observed). Below 50% in 2 of the 18 complete months (2025-04 to 2026-09): 2025-12 (41.8%) and 2026-01 (39.8%); the other 16 are between 51.9% (2025-07) and 76.3%. 60-day window 28.3% (2025-12) to 64.0% (2026-09); 180-day window 70.0% (2026-01) to 90.7% (2026-09), from 2025-07; complete months only, like the 90-day range (2026-10, partial: 73.3% and 91.5%). E3. DEC-074, DEC-077. Figure 3.6 (section Two Speeds): pipelines/ecosystem/two_speeds.py, analysis/ecosystem/two_speeds.csv (= visualizations/ecosystem/fig3_6_data.csv). Panel a: pipelines.tokens.reassess data and age rule (days from the version's first day in the observed top 50, from 2025-01-01, on each day of use); denominator: all observed tokens of September 2026 including the "other" row (daily_totals.csv, 30 observed days); the 0–90-day band, 0.76, reproduces note 5's 76.3%. Panel b: chapter 01's population and cap (capped 30-day downloads of the 1,676 tracked models with a non-zero reading on 2026-10-01; pipelines/ecosystem/audit.py); first reading = first monthly snapshot from 2024-08 with downloads > 0; bands: first reading 2026-08 or later 0.08 (101 models), 2025-11 to 2026-07 0.36 (371), 2025-10 or earlier 0.56 (1,204). Qwen2.5: analysis/ecosystem/qwen_breakdown_monthly.csv, 2026-10, 0.24. E3.
  7. 7
    Versions are joined only by exact identifier. Name-based matching is not used. The public data contain almost no alias evidence linking dated snapshots, so a "family" level built from OpenRouter's aliases gives the same result as the version level. DeepSeek V4 Flash snapshots dated 2026-04-23 (leader in 2026-06) and 2026-07-31 (leader in 2026-08) count as two versions. DEC-021, DEC-074. Name-based sensitivity (Pass 26, independent review B-05): a trailing date stamp of eight digits (YYYYMMDD) or four (MMDD or YYMM) is removed from each version identifier, which joins 12 product lines (analysis/tokens/product_line_merges.csv); the 90-day share is then 68.0% in September 2026 (76.3% by version) and above half in 13 of 18 complete months (16 by version), with the lowest month still January 2026 (39.8%) (analysis/tokens/turnover_product_line_sensitivity.csv, pipelines/tokens/reassess.py). It is a check, not a series: DEC-021 keeps the published measure on exact identifiers. Leaders at three levels (Pass 28): research/editorial/interpretation-pass28/scripts/turnover_levels.py, outputs data/turnover_levels_monthly.csv and data/product_line_successions.csv, from data/processed/openrouter/daily_rankings.csv and daily_totals.csv through pipelines.tokens.reassess.load(). Version: the exact identifier without its variant suffix (the chapter's unit); product line: the version with a trailing date stamp removed (reassess.DATE_SUFFIX, the 12 merges of analysis/tokens/product_line_merges.csv, no other joining, DEC-021); developer: the namespace before the slash, a publisher rather than a model family. Leaders by tokens in each complete month; shares of all observed tokens, the "other" row in the denominator. 2026: 9 leader changes in 9 months by version (9 leaders), 8 by product line (7 leaders), 2 by developer (3 leaders: google 2026-01 to 2026-03, anthropic 2026-04 and 2026-05, deepseek 2026-06 to 2026-09). DeepSeek V4 Flash, share of all observed tokens: snapshot 20260423 0.10 (2026-06), 0.11 (2026-07), 0.06 (2026-08); snapshot 20260731 0.12 (2026-08). E3.
  8. 8
    analysis/identity/persistence_audit.csv (built by pipelines/identity/robustness.py), level version, complete follow-up (n = 196): 62.2% at 7 days, 55.1% at 28 days, 38.3% at 91 days, 18.9% at 175 days. Cohort: versions whose first top-50 day falls between 2025-02-01 and 2026-04-22 (175 days before the last observed day, 2026-10-14); versions first seen in January 2025 are left-censored and excluded. Cohorts with incomplete follow-up are reported separately and give higher values (for example 76.6% of 94 versions at 7 days). A day outside the top 50 counts as absent. E3. DEC-018.
  9. 9
    analysis/tokens/concentration_monthly.csv and concentration_u_test.csv. Top 5 by variant, mean of daily share, all tokens: 62.1% (2025-Q1), trough 35.2% (2026-04), 49.8% (2026-Q3); decline 26.8 points, recovery 14.5 points; monthly values 2025-12 to 2026-03 between 35.6% and 40.0% (concentration_monthly.csv, level variant, k 5, all_tokens, mean_of_daily_share). Across 24 combinations: decline 26–38 points, recovery 14–18 points by variant and version; by developer 86.3% → 59.7% (trough 2026-06) → 66.9%, recovery 5–8 points. E3. DEC-074.
  10. 10
    Cloudflare Radar, AI Insights and Data Explorer (radar.cloudflare.com/ai-insights; radar.cloudflare.com/explorer?dataSet=ai.inference), Workers AI inference distribution by model and by task, retrieved 2026-11-04 between about 12:07 and 12:09 UTC through the Data Explorer page's own loader, which relays the documented endpoints /radar/ai/inference/summary/{MODEL,TASK} and /radar/ai/inference/timeseries_groups/{MODEL,TASK} (limitPerGroup=100, no token). 81 responses: summary and daily series by model and by task for each calendar month from 2025-05 to 2026-10, weekly series in two windows, and empty task summaries for 2024-12 to 2025-04. Raw bodies in data/raw/cloudflare-radar/s11-2026-11-04/explorer_bundle.json (1,450,942 bytes, SHA-256 1a61358c…1441f), every body hashed and checked on load; manifest data/manifests/cloudflare_radar_s11_2026-11-04.json. Every response reports units = inferences. Unit change: Cloudflare changelog, "Updated Workers AI popularity metric in Cloudflare Radar", 18 June 2026 ("This change will affect all new measurements as well as historical data"). Radar data are licensed CC BY-NC 4.0; used here with permission. Method: research/cloudflare/01-source-and-methodology.md (Research Sprint 11). E1 (shares as published).
  11. 11
    analysis/cloudflare-s11/task_shares_model_based_monthly.csv, complete months 2025-06 to 2026-10, all inferences of the month as denominator. Text generation 8.4% (2026-08) to 41.2% (2025-09); 8.4% to 12.5% from 2026-06. Embeddings first in 15 of 17 months (exceptions: text generation in 2025-09, 41.2% against 34.2%; speech recognition in 2026-05, 52.0% against 33.9%); median 39.0% over 2025-06 to 2026-04 (32.7% to 70.0%); 75.9%, 82.7%, 84.3%, 84.3%, 86.7% from 2026-06 to 2026-10. Weekly embedding shares (Radar's task series, data/processed/cloudflare-s11/weekly_tasks.csv): 34.7% (week of 2026-05-25), 53.9%, 71.5%, 82.9% (week of 2026-06-15). Daily sensitivity (task_daily_sensitivity.csv): median of daily embedding shares within 3.4 points of the monthly share in 15 of 17 months (2025-08: 3.7; 2025-09: 9.2). E3.
  12. 12
    data/processed/cloudflare-s11/monthly_models.csv and analysis/cloudflare-s11/model_summary_monthly.csv: @cf/baai/bge-m3 17.6% (2025-06) and 83.7% (2026-10) of all inference requests (Radar counts requests, not tokens), overall leader in 14 of 17 months (bge-large-en-v1.5 in 2025-07 and 2025-08, whisper-large-v3-turbo in 2026-05); 96.6% of embedding inferences in 2026-10 (embeddings_top3_monthly.csv). Text generation (text_generation_summary_monthly.csv, within_task_rankings_monthly.csv): denominator = inferences of every id catalogued as Text Generation, retired ids included (31 to 60 ids a month). Leaders: llama-4-scout-17b-16e-instruct 53.4%, 56.7%, 35.8% (2025-06 to 2025-08) and 8.9% (2026-10); gpt-oss-20b 31.4%, 35.1%; gpt-oss-120b 12.9% to 26.8% (2025-11 to 2026-04) and 16.8% (2026-09); gemma-4-26b-a4b-it 26.7%, 37.2%, 17.2%; glm-4.7-flash 20.2% (2026-08) and 8.2% (second, 2026-10). Five most-used models: 76.2% and 78.2% (2025-06, 2025-07), 35.3% (2026-10). With Radar's task series as denominator (the method of the superseded first Cloudflare analysis, 7e72020): leader shares 4.0 to 18.3 points higher before 2026-06 and 0.4 to 1.0 after; ranking order unchanged. Model shares as published: E1; within-task shares: E3.
  13. 13
    pipelines/cloudflare_s11/identity_matches.csv (rule and evidence per pair) and analysis/cloudflare-s11/openrouter_top5_on_workers_ai.csv. OpenRouter top five by version and tokens, from this chapter's data (2026-10 has 14 days). Same-release slots: moonshotai/kimi-k2.5-0127 (2026-02, not in Workers AI's data that month), z-ai/glm-5.2-20260616 (2026-07; Workers AI rank 8, 3.5%), deepseek/deepseek-v4-flash-20260731 (2026-08, OpenRouter's leader with 12.3% of tokens; rank 11, 2.8%; and 2026-09, rank 11, 3.6%), z-ai/glm-5.3-flash-20260826 (2026-09, OpenRouter's leader with 9.9%; rank 6, 6.4%; and 2026-10, rank 8, 5.0%). Workers AI leaders on OpenRouter (openrouter_matched_llms_monthly.csv): gpt-oss-120b best rank 7 (2025-12, 3.3% of tokens). Matched models per month: 3 to 8; rank correlations (openrouter_rank_correlation.csv) not interpreted. E3.
  14. 14
    analysis/tokens/coverage_monthly.csv; 650 observed days; 2025-06-15 and 2025-07-15 absent in the source and not interpolated. Share of each month's tokens in the "other" row (column other_share_of_month): 3.5% (2025-03) to 9.8% (2026-05). Revision probe of 2026-10-19 against the 2026-10-15 extraction: 11,526 matching rows, 0 changed (analysis/tokens/revision_check.csv). E3. DEC-026, DEC-073.
  15. 15
    data/processed/openrouter/daily_totals.csv. Median daily tokens, first 30 days 7.25×10¹⁰, last 30 days 2.26×10¹³ (ratio about 312). Daily totals as published: E1; ratio: E3.
  16. 16
    analysis/tokens/variant_split.csv; analysis/tokens/concentration_u_test.csv (top 5, mean of daily share, all tokens, 2025-Q1: variant 62.1%, version 67.7%). :beta was 29.9% of top-50 tokens in 2025-01 and 0% from 2025-08. E3. DEC-077.
  17. 17
    analysis/tokens/turnover_checks.csv and turnover_checks_meta.csv. Re-entries: 248 of 378 versions have more than one spell in the top 50. Strict window (<90 days): at most about 1 point lower. Top-50-only denominator: 3–6 points higher. Excluding versions listed in OpenRouter's model catalog before 2025 (72 versions; catalog_first_seen in registry/model_identity.csv, from archived captures of OpenRouter's /api/v1/models, DEC-075 rev. 1): 2025-04 changes from 71.95% to 71.13% and 2025-07 from 51.90% to 51.89%; the other months are unchanged. Between 1% and 29% of each month's recent-arrival tokens belong to versions whose first top-50 day came through a variant identifier. E3. DEC-077.
  18. 18
    analysis/tokens/turnover_monthly.csv, columns entrants and exits, level version, 2025-02 to 2026-09. First month left-censored and last month right-censored for this count. E3. DEC-018.
  19. 19
    Classes and statuses in registry/model_openness.csv; definitions in research/openness-taxonomy.md. OW-P: weights in a public, ungated repository and licence file matching a permissive licence (apache-2.0, mit, bsd-3-clause). OW-R: the same, with a verified restrictive licence (none in the registry). OW-N: weights public and ungated; licence non-commercial, outside the policy list, or declared without a matching licence file (the licence has not necessarily been read). CW: explicit primary source required; nine recorded decisions are suspended under DEC-099, so none is confirmed. UNK: none of the above, conflicting evidence, or gated access with unread terms. All 142 open classes are auto_verified (automated checks); no class is human_approved. Point-in-time validity (pit_strict): open only from the first archived public capture of the repository (141 versions), or from the date we retrieved the evidence where no archived capture exists (1 version). Repository metadata and licence files hashed. DEC-019, DEC-022, DEC-023, DEC-024.
  20. 20
    research/tokens/human-review-package.md and analysis/tokens/review_candidates.csv. 2025-Q1 shares: Claude 3.5 Sonnet 20.9%, Claude 3.7 Sonnet 16.6%, Claude 3.5 Sonnet (2024-06-20) 0.5%; Llama versions about 6% combined; all 0% in 2026-Q3. No decision has been recorded for any of these versions. registry/human_decisions.csv holds nine historical closed-weight decisions on other versions (DEC-093, DEC-094); all nine are suspended under DEC-099 (registry/decision_suspensions.csv), so none counts as a verified closed-weight classification. DEC-078.
  21. 21
    analysis/cloudflare-s11/task_series_catalog_coverage_test.csv and task_reconciliation.csv. Largest monthly gap between the model series summed by catalog task and Radar's task series: 7.95 points; with the models no longer in the current Workers AI catalog removed and the rest renormalised: at most 0.30 points in all 18 months (0.03 or less in 11). Retired ids: 2.3% to 11.4% of monthly inferences before 2026-06, 0.3% to 0.6% after; most are text-generation models. Radar's task series, text generation: 8.1% to 34.9% (2025-06 to 2026-10; task_summary_monthly.csv). Model-to-task mapping: data/processed/cloudflare-s11/model_catalog.csv (153 ids; current catalog first, serving prefixes stripped, keyword rules for retired ids). Our inference, not documented by Cloudflare. E3.