TokenMaxxer
Get started
Live rankings

AI Model Rankings

Share

Live rankings from real developer usage. Models are ranked by the tokens developers actually run through Claude Code, Codex, Cursor and every other tool that syncs to TokenMaxxer. View all models

Usage data through Oct 6, 2026

Tokens tracked
695B
38.85B this week
Spend at list price
$509K
$21.69K this week
Models in use
98
from 11 providers
Tools
9
CLIs, editors, agents

Top Models

Weekly token usage of models across TokenMaxxer, last six months.

0
25B
50B
75B
100B
125B
  1. Apr 8 – 14, 2026. Total 3.4B. Auto 354M, Others 3B.
  2. Apr 15 – 21, 2026. Total 3.2B. Auto 338M, Others 2.9B.
  3. Apr 22 – 28, 2026. Total 4.3B. GPT-5.5 983M, Auto 679M, Others 2.6B.
  4. Apr 29 – May 5, 2026. Total 6.8B. GPT-5.5 1.9B, Auto 1.1B, Others 3.8B.
  5. May 6 – 12, 2026. Total 4.2B. GPT-5.5 1.4B, Auto 567M, Others 2.3B.
  6. May 13 – 19, 2026. Total 4.2B. GPT-5.5 787M, Auto 170M, Others 3.2B.
  7. May 20 – 26, 2026. Total 4.8B. GPT-5.5 1.5B, Auto 523M, Others 2.7B.
  8. May 27 – Jun 2, 2026. Total 8.7B. GPT-5.5 3.7B, Opus 4.8 3B, Auto 366M, Others 1.6B.
  9. Jun 3 – 9, 2026. Total 7.7B. Opus 4.8 5B, GPT-5.5 1.2B, Auto 482M, Fable 5 252M, Others 687M.
  10. Jun 10 – 16, 2026. Total 12.2B. GPT-5.5 4.7B, Opus 4.8 3.3B, Fable 5 2.2B, Auto 333M, Others 1.6B.
  11. Jun 17 – 23, 2026. Total 9.2B. Opus 4.8 5.7B, GPT-5.5 2.3B, Auto 140M, Others 1.1B.
  12. Jun 24 – 30, 2026. Total 11.8B. Opus 4.8 9B, GPT-5.5 1.7B, Others 1B.
  13. Jul 1 – 7, 2026. Total 17.7B. Opus 4.8 9.5B, Fable 5 4.7B, GPT-5.5 1.1B, Auto 842M, Others 1.6B.
  14. Jul 8 – 14, 2026. Total 24.8B. Opus 4.8 9.2B, GPT-5.6 Sol 6.5B, Fable 5 4.4B, GPT-5.5 1.6B, Auto 905M, Others 2.2B.
  15. Jul 15 – 21, 2026. Total 58.7B. GPT-5.6 Sol 32.2B, Opus 4.8 10.5B, GPT-5.5 6.2B, Fable 5 5.2B, Auto 454M, Others 4.2B.
  16. Jul 22 – 28, 2026. Total 43.2B. GPT-5.6 Sol 19.3B, Opus 4.8 7.3B, Opus 5 6.6B, Fable 5 3.5B, GPT-5.5 2.2B, Auto 303M, Others 4B.
  17. Jul 29 – Aug 4, 2026. Total 36.7B. Opus 5 15.4B, GPT-5.6 Sol 13.1B, Fable 5 2.4B, Opus 4.8 1.3B, Auto 860M, GPT-5.5 7.2M, Others 3.6B.
  18. Aug 5 – 11, 2026. Total 34B. GPT-5.6 Sol 14.4B, Opus 5 11.6B, Fable 5 2.2B, Opus 4.8 1.3B, Auto 997M, GPT-5.5 23.6M, Others 3.5B.
  19. Aug 12 – 18, 2026. Total 36.8B. Opus 5 17.2B, GPT-5.6 Sol 12.2B, Fable 5 2.8B, Auto 1.1B, Opus 4.8 91.4M, Others 3.3B.
  20. Aug 19 – 25, 2026. Total 32.7B. Opus 5 14.6B, GPT-5.6 Sol 13.6B, Fable 5 1.7B, Auto 1.3B, Opus 4.8 172M, GPT-5.5 20.9M, Others 1.3B.
  21. Aug 26 – Sep 1, 2026. Total 74.9B. GPT-5.6 Sol 48.5B, Opus 5 13.1B, Fable 5 1.7B, Auto 248M, Opus 4.8 122M, GPT-5.5 42.7M, Others 11.2B.
  22. Sep 2 – 8, 2026. Total 115B. GPT-5.6 Sol 74.4B, Opus 5 18B, GPT-6 Astra 14.5B, Fable 5 867M, Auto 838M, Opus 4.8 153M, GPT-5.5 15.8M, Others 6.3B.
  23. Sep 9 – 15, 2026. Total 24.2B. Opus 5 12.5B, GPT-6 Astra 7.6B, Auto 1.3B, GPT-5.6 Sol 834M, GPT-5.5 8M, Opus 4.8 5.6M, Others 1.9B.
  24. Sep 16 – 22, 2026. Total 27.7B. Opus 5 15.8B, GPT-6 Astra 4.4B, Opus 5.5 1.3B, Auto 1.3B, GPT-5.6 Sol 781M, Opus 4.8 121M, Fable 5 40.3M, Others 3.9B.
  25. Sep 23 – 29, 2026. Total 30.9B. Opus 5.5 20.5B, GPT-6 Astra 3.3B, Opus 5 3.1B, GPT-5.6 Sol 538M, Auto 459M, Opus 4.8 123M, Others 3B.
  26. Sep 30 – Oct 6, 2026. Total 38.9B. Opus 5.5 16.6B, GPT-6 Astra 10.6B, Opus 5 600M, Opus 4.8 334M, Auto 232M, GPT-5.6 Sol 111M, Fable 5 20.9M, Others 10.4B.
Apr ’26Jun 17Aug 26
Share of the last 26 weeks

Each column is a trailing seven-day window; the last one ends today, so it matches every “this week” figure on the page. The chart opens on six months because most tools only started syncing in mid-2026, so earlier weeks mostly show the tools that keep long histories.

LLM Leaderboard

Models ranked by tokens processed across every tool on TokenMaxxer.

  1. 1.Opus 5.5by Anthropic
    16.6B tokens
    down 19%
  2. 2.GPT-6 Astraby OpenAI
    10.6B tokens
    up 225%
  3. 3.GPT 6 1 SOLby OpenAI
    7B tokens
    up >999%
  4. 4.Fable 5.1by Anthropic
    958M tokens
    up 554%
  5. 5.GPT-6 Solby OpenAI
    924M tokens
    down 48%
  1. 6.Opus 5by Anthropic
    600M tokens
    down 81%
  2. 7.Sonnet 5by Anthropic
    495M tokens
    up >999%
  3. 8.Claude Sonnet 5 5by Anthropic
    392M tokens
    up 87%
  4. 9.Codex Auto Reviewby OpenAI
    348M tokens
    up 1.2%
  5. 10.Opus 4.8by Anthropic
    334M tokens
    up 171%

Trailing 7 days (Sep 30 – Oct 6) against the 7 days before them.

Top Models by Tool

Each tool's leading models, sized by share of tokens on TokenMaxxer. Pick a tool to rank what runs inside it.

Codex45% of all tokens all time · 311B tokens
  1. 1GPT-5.6 Solby OpenAI
    68%
    210B
  2. 2GPT-6 Astraby OpenAI
    13%
    40.3B
  3. 3GPT-5.5by OpenAI
    8.0%
    24.7B
  4. 4GPT-5.3 Codexby OpenAI
    4.3%
    13.4B
  5. 5GPT 6 1 SOLby OpenAI
    2.3%
    7.1B
  1. 6Codex Auto Reviewby OpenAI
    1.5%
    4.6B
  2. 7GPT-6 Solby OpenAI
    0.8%
    2.6B
  3. 8GPT-5.6 Lunaby OpenAI
    0.7%
    2.3B
  4. 9GPT-5.6 Terraby OpenAI
    0.7%
    2.3B
  5. 10GPT-5.4by OpenAI
    0.7%
    2.3B

Cost per Session

Crunching individual sessions: what one coding-agent session typically costs, by session length, across finished sessions started in the last 60 days.

Cursor and Grok Build are left out: they report each request or prompt with its agent steps rolled into one row, so their sessions can't be grouped by how many model calls they ran.

Session Length

Crunching individual sessions: how long a coding-agent session runs, counted in model calls, across finished sessions started in the last 60 days.

A turn is one model call (a reply or a single agent step), not one message you typed, so one prompt can run dozens. Length is the time from a session's first turn to its last, idle gaps included. A session with a turn in the last 6hours is left out until it settles, so none is counted half-finished. Cursor and Grok Build are left out: they report each request or prompt with its agent steps rolled into one row, so their turn counts aren't comparable.

Market Share

Compare token share by model author on TokenMaxxer, week by week.

0%
25%
50%
75%
100%
  1. Apr 8 – 14, 2026. Total 3.4B. Anthropic 69%, OpenAI 16%, Cursor 14%, Google 0.6%.
  2. Apr 15 – 21, 2026. Total 3.2B. Anthropic 55%, OpenAI 32%, Cursor 13%, Google <0.1%.
  3. Apr 22 – 28, 2026. Total 4.3B. Anthropic 51%, OpenAI 30%, Cursor 19%, Google 0.2%.
  4. Apr 29 – May 5, 2026. Total 6.8B. Anthropic 46%, OpenAI 34%, Cursor 18%, Google 1.0%, Moonshot / Kimi 0.2%, Zhipu / GLM 0.2%.
  5. May 6 – 12, 2026. Total 4.2B. Anthropic 45%, OpenAI 37%, Cursor 15%, Google 2.5%.
  6. May 13 – 19, 2026. Total 4.2B. Anthropic 72%, OpenAI 19%, Cursor 6.1%, Google 2.2%, Alibaba / Qwen 0.1%.
  7. May 20 – 26, 2026. Total 4.8B. Anthropic 51%, OpenAI 33%, Cursor 14%, Google 2.3%, Alibaba / Qwen <0.1%.
  8. May 27 – Jun 2, 2026. Total 8.7B. Anthropic 49%, OpenAI 43%, Cursor 7.9%, Google 0.7%, Others <0.1%.
  9. Jun 3 – 9, 2026. Total 7.7B. Anthropic 72%, OpenAI 16%, Cursor 12%, Google 0.3%, Alibaba / Qwen 0.2%, Others <0.1%.
  10. Jun 10 – 16, 2026. Total 12.2B. Anthropic 54%, OpenAI 39%, Cursor 6.0%, Moonshot / Kimi 0.6%, Zhipu / GLM 0.3%, Google 0.3%, Others <0.1%.
  11. Jun 17 – 23, 2026. Total 9.2B. Anthropic 68%, OpenAI 26%, Cursor 4.0%, Google 2.3%, Zhipu / GLM <0.1%, Moonshot / Kimi <0.1%.
  12. Jun 24 – 30, 2026. Total 11.8B. Anthropic 84%, OpenAI 15%, Cursor 1.1%, Google 0.3%.
  13. Jul 1 – 7, 2026. Total 17.7B. Anthropic 86%, OpenAI 6.4%, Cursor 6.3%, Google 1.1%.
  14. Jul 8 – 14, 2026. Total 24.8B. Anthropic 60%, OpenAI 35%, Cursor 4.8%, Zhipu / GLM 0.3%, Google <0.1%, xAI <0.1%.
  15. Jul 15 – 21, 2026. Total 58.7B. OpenAI 67%, Anthropic 29%, Zhipu / GLM 1.5%, Cursor 1.4%, xAI 0.2%, Moonshot / Kimi <0.1%, Google <0.1%, Others <0.1%.
  16. Jul 22 – 28, 2026. Total 43.2B. OpenAI 56%, Anthropic 43%, Cursor 0.9%, xAI 0.4%, Google <0.1%.
  17. Jul 29 – Aug 4, 2026. Total 36.7B. Anthropic 58%, OpenAI 38%, Cursor 2.5%, xAI 0.4%, Google 0.2%, Moonshot / Kimi <0.1%, Zhipu / GLM <0.1%.
  18. Aug 5 – 11, 2026. Total 34B. Anthropic 52%, OpenAI 45%, Cursor 3.1%, xAI <0.1%, Google <0.1%, Zhipu / GLM <0.1%, Others <0.1%.
  19. Aug 12 – 18, 2026. Total 36.8B. Anthropic 61%, OpenAI 36%, Cursor 3.1%, xAI <0.1%, Google <0.1%, Others <0.1%.
  20. Aug 19 – 25, 2026. Total 32.7B. Anthropic 52%, OpenAI 43%, Cursor 4.1%, xAI 0.2%, Google <0.1%, Zhipu / GLM <0.1%, Others 0.9%.
  21. Aug 26 – Sep 1, 2026. Total 74.9B. OpenAI 79%, Anthropic 20%, Cursor 0.4%, xAI 0.1%, Google <0.1%, Others <0.1%.
  22. Sep 2 – 8, 2026. Total 115B. OpenAI 79%, Anthropic 20%, Cursor 0.8%, xAI 0.2%, Zhipu / GLM <0.1%, Others <0.1%.
  23. Sep 9 – 15, 2026. Total 24.2B. Anthropic 56%, OpenAI 37%, Cursor 6.1%, Zhipu / GLM 0.3%, xAI 0.2%, Moonshot / Kimi <0.1%, Others <0.1%.
  24. Sep 16 – 22, 2026. Total 27.7B. Anthropic 73%, OpenAI 20%, Cursor 4.9%, xAI 2.1%, Zhipu / GLM 0.3%, Google <0.1%.
  25. Sep 23 – 29, 2026. Total 30.9B. Anthropic 78%, OpenAI 19%, Cursor 1.5%, xAI 0.6%, Zhipu / GLM 0.1%, Google <0.1%.
  26. Sep 30 – Oct 6, 2026. Total 38.9B. Anthropic 50%, OpenAI 49%, Cursor 0.6%, xAI 0.3%, Google <0.1%.
Apr ’26Jun 17Aug 26
Share of the last 26 weeks
This week · Sep 30 – Oct 6, 2026
  1. 1.Anthropic19 developers this week
    50%
    down −27.9 points
  2. 2.OpenAI12 developers this week
    49%
    up +29.3 points
  3. 3.Cursor1 developer this week
    0.6%
    down −0.9 points
  1. 4.xAI3 developers this week
    0.3%
    down −0.3 points
  2. 5.Google1 developer this week
    <0.1%
    0
  3. 6.Zhipu / GLM0 developers this week
    0.0%
    down −0.1 points

Shares are for the trailing 7 days (the chart's last column), with the change in share points against the 7 days before. Labs outside the chart's top 8 fold into Others. Unattributed models (“Unknown”) always fold into Others.

Price vs. Usage

What a million tokens costs at list price against how widely each model is used, over the trailing 30 days.

15 models · 2 not plotted (no list price)

Developers
0
10
20
30
Effective price per 1M tokens · log scale

Tap a dot to read that model's full profile.

Ranked by developers · last 30 days
  1. 1.Opus 5.5by Anthropic
    22devs
    $0.34 / 1M
  2. 2.Opus 5by Anthropic
    19devs
    $0.70 / 1M
  3. 3.GPT-5.6 Solby OpenAI
    16devs
    $0.53 / 1M
  4. 4.GPT-6 Astraby OpenAI
    15devs
    $1.35 / 1M
  5. 5.Sonnet 5by Anthropic
    11devs
    $0.30 / 1M
  1. 6.Codex Auto Reviewby OpenAI
    11devs
    $0.58 / 1M
  2. 7.Fable 5.1by Anthropic
    10devs
    $0.84 / 1M
  3. 8.GPT 6 1 SOLby OpenAI
    8devs
    no list price
  4. 9.Opus 4.8by Anthropic
    7devs
    $0.91 / 1M
  5. 10.GPT-5.6 Lunaby OpenAI
    7devs
    $0.032 / 1M

Trailing 30 days. Effective price and cost per turn re-price each model's real token mix (input, output, reasoning and cache tokens) at today's list rates; spend is the cost recorded at sync, as everywhere else on this page. A model needs at least 2 developers and 25Mtokens in the window to be mapped. Per-turn figures leave out Cursor, which reports a whole agent request as a single turn. A model used only through Cursor, like Cursor's own Auto and Composer, has no turns to divide by and isn't mapped.

Effective Price

What a million tokens actually costs once caching is counted: each model's real trailing-30-day token mix, priced at its list rates. The top 12 by spend.

98% of these models' prompt tokens were cache reads.

  • Effective price
  • List price, input → output
  1. GPT-6 Astra98% cache hits
    7.4× under$1.35
    list $10 in · $50 out
  2. Opus 598% cache hits
    7.1× under$0.70
    list $5 in · $25 out
  3. Opus 5.598% cache hits
    12× under$0.34
    list $4 in · $20 out
  4. GPT-5.6 Sol98% cache hits
    7.5× under$0.53
    list $4 in · $20 out
  5. Fable 5.197% cache hits
    12× under$0.84
    list $10 in · $50 out
  6. Auto93% cache hits
    3.5× under$0.36
    list $1.25 in · $6 out
  7. Codex Auto Review86% cache hits
    4.3× under$0.58
    list $2.50 in · $15 out
  8. GPT-6 Sol98% cache hits
    7.4× under$0.27
    list $2 in · $10 out
  9. Sonnet 597% cache hits
    6.6× under$0.30
    list $2 in · $10 out
  10. Opus 4.894% cache hits
    5.5× under$0.91
    list $5 in · $25 out
  11. Composer 2.5 Fast92% cache hits
    2.0× under$1.47
    list $3 in · $15 out
  12. Grok 4.695% cache hits
    3.3× under$0.60
    list $2 in · $6 out

Effective price = the model's 30-day input, output, reasoning and cache tokens, each billed at its own list rate, divided by the total. “Under” compares it with the list input price. Shows the 12 models with the most spend among those with at least 50M tokens in the window.

Weekly Rhythm

Which days the network codes on: tokens by day of the week over the last 12 whole weeks, split by tool.

Weekends carry 24% of the network's tokens. An even spread would give them 29%. Codex does 71% of weekend tokens, against 45% on weekdays.

Avg tokens per day
0
2.5B
5B
7.5B
10B
  1. Monday: 8.2B tokens on an average day, 18% of the week. Codex 51%, Claude Code 43%, Cursor 3.0%, Pi 1.9%, OpenCode 0.1%, Others <0.1%.
  2. Tuesday: 7.3B tokens on an average day, 16% of the week. Claude Code 49%, Codex 42%, Cursor 4.6%, Pi 3.7%, OpenCode 0.2%, Others <0.1%.
  3. Wednesday: 5.8B tokens on an average day, 13% of the week. Claude Code 57%, Codex 32%, Cursor 5.9%, Pi 4.5%, OpenCode 0.6%, Others <0.1%.
  4. Thursday: 5.8B tokens on an average day, 12% of the week. Claude Code 55%, Codex 35%, Cursor 4.9%, Pi 4.6%, OpenCode 0.6%, Others <0.1%.
  5. Friday: 8.3B tokens on an average day, 18% of the week. Codex 58%, Claude Code 34%, Cursor 5.5%, Pi 2.5%, OpenCode 0.4%, Others <0.1%.
  6. Saturday: 5.1B tokens on an average day, 11% of the week. Codex 66%, Claude Code 26%, Cursor 4.6%, Pi 3.2%, OpenCode <0.1%, Others <0.1%.
  7. Sunday: 6.1B tokens on an average day, 13% of the week. Codex 74%, Claude Code 22%, Cursor 2.3%, Pi 1.6%, OpenCode <0.1%, Others <0.1%.
Share of tokens, Jul 14 – Oct 5, 2026
Weekend share · by toolNetwork average 24%
  1. 1Codex287B tokens
    33%
    21 devs
  2. 2Antigravity130M tokens
    27%
    5 devs
  3. 3Cursor24.5B tokens
    19%
    7 devs
  1. 4Pi17.1B tokens
    18%
    3 devs
  2. 5Claude Code230B tokens
    14%
    29 devs
  3. 6OpenCode1.5B tokens
    0.9%
    4 devs

Days are each developer's own calendar days (Cursor's cloud usage is dated in UTC), counted over the 12 whole weeks ending yesterday (Jul 14 – Oct 5, 2026), so every weekday appears exactly 12times and a half-synced today can't dent a column. The chart names, and the weekend list ranks, only what at least 3 developers used in the window. The rest folds into Others, which always stands for at least as many; the tick marks the network average.

Context Length

Crunching individual sessions: how much each model call has to read, prompt size per turn, across sessions started in the last 60 days.

Cursor and Grok Build are left out: they report each request or prompt with all its agent steps rolled into one number, so their “prompt” isn't a single model call.

Top Tools

The coding tools developers run their tokens through: what each one is, how many people use it, and the model it leans on.

  1. 1.Claude Code
    19.6B
    down 19%

    Anthropic's agentic coding CLI

    19 developers · mostly Opus 5.5 (85%)

  2. 2.Codex
    19B
    up 221%

    OpenAI's coding agent, CLI and IDE

    12 developers · mostly GPT-6 Astra (56%)

  1. 3.Cursor
    337M
    down 50%

    AI code editor forked from VS Code

    3 developers · mostly Auto (69%)

  2. 4.Antigravity
    102K
    down 98%

    Google's agent-first IDE and CLI

    1 developer · mostly Gemini 3.6 Flash (100%)

No usage this week:

Tokens over the trailing 7 days, across every model a tool ran, against the 7 days before. “New” marks a tool whose first usage falls inside the period; “Back”, one returning after a quiet stretch. Cursor's editor and cloud usage count as one tool, as do Antigravity's app and CLI. Each dot is the tool's colour in Market Share by tool; grey means it folds into Others there.

How these rankings are measured

What the numbers are, where they come from, and what they can't tell you.

Where the numbers come from

Every figure is real usage that developers synced to TokenMaxxer. It comes from the logs Claude Code, Codex, Cursor and the rest already write on their machines, or, for Cursor, from its cloud usage API when they connect it. Tokens count input, output, reasoning, cache reads and cache writes. No prompt or code ever reaches TokenMaxxer. Project names are synced but stay private unless their owner chooses to show them, and nothing here is broken down by project.

Who is counted

Only developers listed on the public leaderboard, and listing is opt-in. It's the same population as the leaderboard and the model, tool and lab pages, so totals reconcile with them, and anyone who leaves the leaderboard leaves every ranking at the same moment.

Time windows

This week and this month are the trailing 7 and 30 days ending today, compared with the 7 or 30 days before them. Weekly charts use the same trailing weeks, so the last column is always this week. Trending ranks week-over-week growth among models with at least 25M tokens this week. A model is new when its first usage (on the network, or inside the tool you picked) falls in the period, and back when it was used before but not in the comparison window.

Why the charts open on six months

Top Models shows the last 13, 26 or 52 weeks and opens on 26; Market Share always covers the trailing 26weeks. Older weeks are lopsided: Claude Code deletes old transcripts from disk while Cursor's cloud sync backfills years, so the far end of a year-long chart mostly shows which tools keep history, not what developers used.

Weekly rhythm

The weekday chart covers 12 whole weeks ending yesterday (Jul 14 – Oct 5), so every weekday is counted the same number of times and a half-synced today can't dent one column; the absolute view shows an average day. Days are each developer's own calendar day. The network's weekend share is Saturday and Sunday tokens over all tokens. An even week would give 2 in 7, about 29%. A tool, model or lab needs at least 3 developers to be named in the chart or listed by weekend share, and Others always stands for at least as many.

Turns and sessions

A turn is one model call (one response from the model), not one prompt you typed; an agent working through a task makes dozens. A session is one conversation in one tool, as the tool records it. Session figures count sessions that started in the last 60 days and went quiet at least 6 hours ago, so one still running is never counted short. Each session is credited to the model behind most of its cost.

Sample floors

A figure built on a handful of sessions describes one person's afternoon, so it is withheld rather than printed thin. A session bucket needs 5 sessions, a model 10, a tool 25 and a context row 500 turns, each from at least 2 developers. The price map lists models with 2+ developers and 25M+ tokens over 30 days; effective price needs 50M tokens.

Prices

Spend is tokens priced at each model's public API list rate, whatever plan the developer actually pays for, as recorded when the usage synced. Effective price and cost per turn instead re-price each model's last 30 days of tokens at today's list rates, so a model priced after its usage arrived doesn't look free. Unpriced sessions are left out of cost figures rather than counted as $0.

Cursor and Grok Build

Cursor reports usage per request and Grok Build per prompt, with every agent step they ran rolled into one number. Their tokens, spend and developers count everywhere, but they sit out of cost per session, session length, context length and the per-turn figures on the price map, where one request would pass for one turn, so a model used only through them isn't mapped. Cursor's cloud sync dates usage by UTC day, which is close enough for weekday totals.

Freshness

Usage data runs through Oct 6, 2026. Rankings refresh every few minutes and session statistics hourly. Devices sync on their own schedule, so the most recent day can still grow as late syncs land.

What these rankings can't tell you

They measure adoption, not quality: a model that re-reads a large context burns more tokens without being better, and a cheap model can lead on volume because it is cheap. They describe developers who chose to share their usage, not the whole market, and on a network this size one heavy user can move a week.