OpenRouter usage leaders
Current source-day leaders for 2026-07-07: Hy3, Deepseek V4 Flash, Minimax M3, Mimo V2.5, Deepseek V4 Pro.
How to read OpenRouter token rankingsLeaderboard
AI model leaderboard: OpenRouter rankings, cost, context, and momentum
This is the model leaderboard for tokenmaxxing work. It combines source-labeled OpenRouter rankings, public Hugging Face momentum, and OpenRouter pricing/context boards. It does not claim global model usage or market share.
The tokenmaxxing leaderboard
Who spends the most on AI tokens
Company-level token and model spend, compiled from public receipts — reported figures, derived math, and clearly-labeled estimates. Never GPU capex. Every row links its sources; updated 2026-06-21.
| # | Company | Est. token spend | Confidence | Receipts |
|---|---|---|---|---|
| 1 | OpenAIThe supply side's own token bill: compute purchased from Azure to serve OpenAI's models | ~$11.5B/yr pace ($8.7B in 9 months of 2025) paid to Microsoft Azure for inferenceFrom leaked internal Microsoft financials; Microsoft and OpenAI declined to confirm. $3.7-3.8B in 2024 plus $8.65-8.7B in Jan-Sep 2025. | Reported | Where's Your Ed At (leaked Microsoft docs) (2025-11)The Register (2025-11)TechCrunch (2025-11) |
| 2 | SoftBank GroupGroup-wide OpenAI deployment (Cristal intelligence, ChatGPT Enterprise) — distinct from SoftBank's equity investment and Stargate capex | $3B/yr on OpenAI solutions across group companiesSelf-disclosed in a primary-source press release — the largest disclosed recurring enterprise commitment to a model provider. | Reported | SoftBank Group press release (2025-02)Data Center Dynamics (2025-02) |
| 3 | AppleAnnual licensing fee for a custom 1.2T-parameter Gemini running on Apple's Private Cloud Compute | ~$1B/yr to Google for the custom Gemini behind SiriBloomberg-sourced; never officially confirmed. An Anthropic alternative would reportedly have cost ~$1.5B/yr. | Reported | Bloomberg (Mark Gurman) (2025-11)CNBC (2026-01) |
| 4 | Meta (internal usage)Internal employee Claude consumption, surfaced by the employee-built 'Claudeonomics' leaderboard Meta shut down | 60T Claude tokens/30 days across 85,000+ employees — implies ~$0.7-3.6B/yrMath: 60T tokens/month at $1-5 per M tokens blended = $60-300M/month. Fortune's own derivation put the TOP USER alone above $1.4M/month at Opus pricing. Blended rate is assumed; Meta's actual contract pricing unknown. | Derived | Fortune (2026-04)The Information (first report) (2026-04) |
| 5 | Cursor (Anysphere)Claude inference powering Cursor (mid-2025 run-rate); excludes OpenAI spend and AWS bills | ~$600-900M/yr to Anthropic — its largest customerMath: The Information put Anthropic's top two customers at $1.2-1.4B combined annualized, Cursor the larger — so >$600M. One investor: 'Cursor is spending 100% of its revenue on Anthropic.' Cursor's in-house Composer model (Nov 2025) likely reduced the share since. | Derived | VentureBeat (citing The Information) (2025-08)Where's Your Ed At (citing The Information) (2025-08) |
| 6 | Unnamed enterpriseOne month of Claude usage after failing to set limits on employee licenses — consultant-disclosed, company never named | $500M on Claude in a single monthAnonymous and secondhand — uncorroborated by Anthropic, so we label it an estimate despite wide coverage. The internet's speculation about which company has no receipt and is not repeated here. | Estimate | Axios (2026-05)Tom's Hardware (2026-05)TechCrunch (2026-06) |
| 7 | SalesforceAnthropic (Claude) tokens, driven almost entirely by AI coding agents | ~$300M on Anthropic in 2026 — Benioff's own projectionCEO projection on the All-In podcast, not an audited figure. Distinct from Salesforce's separate equity investment in Anthropic. Benioff also called for routing simple tasks to cheaper models. | Reported | The Next Web (All-In podcast) (2026-05)Quartz via Yahoo Finance (2026-05) |
| 8 | GitHub Copilot (Microsoft)Claude spend backing Copilot (mid-2025); Microsoft reports no Copilot inference line item | ~$300-600M/yr to Anthropic — its #2 customerMath: remainder of the $1.2-1.4B top-two figure after Cursor's larger share. In 2023, Copilot famously lost >$20/user/month on inference (WSJ); it moved to usage-based billing June 1, 2026. | Derived | GitHub Blog (official) (2026-04)VentureBeat (citing The Information) (2025-08) |
| 9 | PerplexityCommitted cloud + model access (OpenAI, Anthropic, xAI via Foundry) — broader than a pure token bill; plus 'hundreds of millions' to AWS | $750M over 3 years to Microsoft Azure/Foundry (~$250M/yr) for cloud + frontier-model accessIn 2024 Perplexity spent 164% of its revenue across AWS, Anthropic and OpenAI (The Information) — ~$8M on model APIs at ~$34M revenue. The 2026 commitment is a different magnitude. | Reported | Bloomberg (2026-01)Business Standard (2026-01) |
| 10 | ReplitAnthropic/OpenAI inference behind Replit Agent — derived from reported gross margins | ~$95-115M/yr implied model-inference costs (late 2025)Math: ~$150M ARR (Sep 2025) at 23-36% blended gross margin implies $96-115M/yr COGS, overwhelmingly third-party inference. CEO Masad acknowledged pricing was 'out of whack' vs model costs; repriced Feb 2026. | Derived | The Information (2025-08) |
| 11 | Windsurf (Cognition)Third-party model inference; floor estimate from 'negative gross margin' reporting | >$82M/yr floor implied pre-acquisition (negative gross margins at $82M ARR)Math: negative gross margin means COGS exceeded the $82M ARR at acquisition (Jul 2025) — a floor, not a point estimate. Cognition disputes 'wildly negative margins' characterizations post-merger. | Derived | Charles Rubenfeld (citing The Information) (2025-09) |
| 12 | LovableAnthropic/OpenAI/Google token costs behind the app-building product — derived from leaked gross margins | ~$65M/yr (mid-2025) rising toward ~$200M/yr (early 2026) in model costsMath: 35% gross margin means ~65% of revenue goes to COGS, predominantly model fees. At $100M ARR (Jul 2025): ~$65M/yr. At $400M ARR (Feb 2026) even at improved ~50% margin: ~$200M/yr. | Derived | Sifted (leaked pitch deck) (2025-12)Bloomberg (CEO Osika) (2025-08) |
| 13 | Bolt (StackBlitz)Claude inference resold as product ('selling inference is our revenue model') | ~$45-90M/yr implied Anthropic spend at $40M ARR (Mar 2025)Math: a $20/month plan covers ~10M tokens that cost $22.50-45 at Sonnet rates if fully used — 112-225% of price. Marked estimate because plan-utilization is assumed. Amodei called Bolt the fastest-growing customer Anthropic had seen. | Estimate | Contrary Research (2025-04)Lenny's Podcast (Eric Simons) (2025-01) |
| 14 | T-Mobile USOpenAI models powering the IntentCX customer-service platform | ~$100M to OpenAI over three years (~$33M/yr) for IntentCXAnnounced Sept 2024, runs through ~2027. OpenAI reportedly called it one of its largest enterprise contracts at signing. Annualized at one third. | Reported | The Information (2024-09)Sherwood News (2024-09) |
| 15 | IntuitOpenAI frontier-model access for agents across TurboTax, QuickBooks, Credit Karma, Mailchimp | $100M+ multi-year OpenAI deal (~$33M+/yr assumed over 3 years)Deal length undisclosed; annualization assumes at least 3 years (documented assumption). Includes Intuit apps inside ChatGPT. | Reported | Intuit press release (2025-11)Bloomberg (2025-11) |
| 16 | UberClaude Code + Cursor tokens across ~5,000 engineers (adoption 32%→84%); total budget never disclosed | Entire 2026 AI coding budget burned in 4 months; ~$25-100M/yr implied org-wideMath: ~4,200 active users at $500-2,000/month. Uber imposed a $1,500/employee/month cap per tool; COO Macdonald: the link from token spend to shipped product 'is not there yet.' One internal demo burned ~$1,200 in two hours. | Derived | Fortune (2026-05)TechCrunch (2026-06) |
| 17 | OpenClaw project (Peter Steinberger)OpenAI API tokens for ~100 autonomous Codex agents on one open-source project | $1.3M OpenAI bill in 30 days — 603B tokens across ~100 coding agents (~$15.7M/yr pace)Steinberger works at OpenAI, which covers the bill — subsidized, not market spend. He noted Fast Mode pricing inflated it; ~$300K/month at standard rates. | Reported | Tom's Hardware (2026-05)The Decoder (2026-05) |
| 18 | SemiAnalysisClaude tokens for research and agent workflows, self-disclosed | $10.95M/yr Claude token run-rate — at ~30 employeesRoughly $365K/yr per employee on tokens — the most extreme per-head figure on the board, self-disclosed by the research firm. | Reported | SemiAnalysis (2026-05) |
The market, for scale
- Median vs average US business AI spend: $2,246/mo median vs $140,842/mo average — the skew IS the tokenmaxxing story (P99: $831K/mo) (Ramp)
- Token spend growth: Business token usage +1,001% and dollar spend +497%, Jan 2025 → Apr 2026 (Ramp)
- Enterprise genAI market: $37B total 2025 enterprise spend (3.2x YoY); $12.5B of it on foundation-model APIs (Menlo Ventures)
- Average Global-2000 LLM budget: ~$7M in 2025, heading to ~$11.6M in 2026 (+65%) (a16z)
- The trillion-token club: 30 companies have each processed 1T+ OpenAI API tokens — incl. Salesforce, Shopify, Canva, Duolingo, Notion, T-Mobile, Ramp (OpenAI DevDay (via Extrapolator))
- Token questions on earnings calls: ~300 companies fielding token questions on April–May 2026 earnings calls, up from ~93 a year earlier (WIRED)
- Satya Nadella on tokenmaxxing at Microsoft: Microsoft CEO Nadella said employee tokenmaxxing happens 'a lot' and called it 'addictive' — but directed staff to match the model to the task rather than default to the biggest one (Business Insider)
- Box CEO on enterprise AI budgeting: Aaron Levie (Box): 'nobody has budgeted' for token costs that 'just showed up overnight' once engineers adopted Claude Code at scale (Semafor)
Receipts without invoices
- Mercor: Token spend on internal agents now exceeds total employee payroll (~300 staff) (Business Insider (20VC podcast))
- Amazon: Shut down KiroRank, an internal leaderboard ranking employees by AI token usage — 'don't use AI just to use AI' (Business Insider)
- Microsoft: Canceled most internal Claude Code licenses in Experiences & Devices after token billing consumed the annual AI budget ahead of schedule (The Next Web (via The Verge))
- Notion: AI model costs compressed profit margins by ~10 percentage points (Where's Your Ed At)
- Unnamed startup: 'One of my engineers spent $40,000 on tokens last month' — CTO via Faros AI; per-developer token use up ~18.6x in nine months (Jellyfish) (TechCrunch)
Routed in public
OpenRouter's open books
OpenRouter publishes its app leaderboard token volumes — the only first-party, daily-visible spend signal on the open web. Dollar ranges are derived from those volumes; the math is in each note.
| # | Company | Est. token spend | Confidence | Receipts |
|---|---|---|---|---|
| 1 | OpenRouter (entire marketplace)All inference routed through the marketplace by 8M users and apps | ~$50-150M/month of inference at 100T tokens/monthMath: 100T tokens/month at $0.50-1.50/M blended. Anchor: OpenRouter disclosed >$100M annualized inference spend in May 2025 at roughly 1/12th today's volume. | Derived | Business Wire (Series B) (2026-05)TechCrunch (2026-05) |
| 2 | Hermes Agent (Nous Research)All-time cumulative tokens routed via OpenRouter, per-app page; USD is a derived estimate, not a disclosed spend figure | $5.3M–$53.3M all-time cumulative spend estimate — #1 app on OpenRouter, 26.63T tokens all-time (as of 2026-07-07)Math: 26.63T all-time tokens at $0.2/M–$2/M = $5.3M–$53.3M all-time cumulative spend estimate. The OpenRouter /apps index is client-rendered and no longer curl-able; the per-app page exposes totalTokens (all-time cumulative) + rank — those are the only unambiguous, reliably-fetchable figures. Do NOT interpret figureLowUsd/figureHighUsd as a monthly or annual run-rate. | Derived | OpenRouter app page (hermes-agent) (2026-07) |
| 3 | Kilo CodeAll-time cumulative tokens routed via OpenRouter's public app leaderboard | $3.8M–$15.1M all-time cumulative spend estimate — #3 app on OpenRouter, 7.57T tokens all-time (as of 2026-07-07)Math: 7.57T all-time tokens at $0.5/M–$2/M = $3.8M–$15.1M all-time cumulative spend estimate. The OpenRouter /apps index is client-rendered and no longer curl-able; the per-app page exposes totalTokens (all-time cumulative) + rank — those are the only unambiguous, reliably-fetchable figures. Do NOT interpret figureLowUsd/figureHighUsd as a monthly or annual run-rate. | Derived | OpenRouter app page (kilo-code) (2026-07) |
| 4 | OpenClaw (all users)All-time cumulative tokens routed via OpenRouter; spend distributed across thousands of self-hosted users | $2.5M–$10.1M all-time cumulative spend estimate — #4 app on OpenRouter, 5.06T tokens all-time (as of 2026-07-07)Math: 5.06T all-time tokens at $0.5/M–$2/M = $2.5M–$10.1M all-time cumulative spend estimate. The OpenRouter /apps index is client-rendered and no longer curl-able; the per-app page exposes totalTokens (all-time cumulative) + rank — those are the only unambiguous, reliably-fetchable figures. Do NOT interpret figureLowUsd/figureHighUsd as a monthly or annual run-rate. | Derived | OpenRouter app page (openclaw) (2026-07) |
Estimates of company-level spend on LLM tokens and AI model usage, compiled from public reporting on 2026-06-10. 'Reported' = a figure stated by the company or a named outlet. 'Derived' = computed from disclosed token volumes, margins, or revenue splits — the math is in each row's note. 'Estimate' = credible reporting with an assumed utilization or rate, or anonymous sourcing. Scoped to model/API/token spend, never GPU capex or datacenter buildouts (JPMorgan's well-known $2B/yr AI figure is excluded for exactly that reason). Rows are ranked by their CONSERVATIVE (low) annualized estimate, so derived math can never outrank primary-source figures; one-time and multi-year totals are annualized conservatively where noted. Every row links its receipts.
Source-labeled rankings
Use the leaderboard as a routing and cost signal.
Cost and context checks
Pair ranking movement with model price, context window, and task fit before treating any leaderboard row as an operating decision.
No fake market-share claims
Router rankings can show public traffic inside that source. They are not proof of global model usage unless the source says so.
OpenRouter rankings, explained
Just want the usage rankings with context? The dedicated page covers what the numbers mean, how to read the publish lag, and why token volume differs from request count.
OpenRouter rankingsPublic usage
Tokens spent by top OpenRouter models
OpenRouter exposes public rankings on its site. The token columns below only appear when the source is live or backed by a last-known-good snapshot; the 30-day figure is a labeled run-rate estimate.
OpenRouter rankings: liveToken usage from OpenRouter's live model rankings, through 2026-07-07. Daily tokens are the source-day total; the 30-day column is a real measured total across the trailing 30 days from OpenRouter's data.Last checked 7/8/2026
| Rank | Model | Source day | Tokens/day | 30-day total (measured) |
|---|---|---|---|---|
| #1 | Hy3Tencent | 2026-07-07 | 858.4B | 926.9B |
| #2 | Deepseek V4 FlashDeepseek | 2026-07-07 | 806.2B | 20.9T |
| #3 | Minimax M3Minimax | 2026-07-07 | 665.9B | 17.3T |
| #4 | Mimo V2.5Xiaomi | 2026-07-07 | 665.6B | 17.7T |
| #5 | Deepseek V4 ProDeepseek | 2026-07-07 | 530.6B | 9.9T |
| #6 | Glm 5.2Z Ai | 2026-07-07 | 504.5B | 7T |
| #7 | Claude 4.7 OpusAnthropic | 2026-07-07 | 398B | 10.2T |
| #8 | Claude 4.8 OpusAnthropic | 2026-07-07 | 313.3B | 7.5T |
| #9 | Step 3.7 FlashStepfun | 2026-07-07 | 192.5B | 5.2T |
| #10 | Gpt 5.5Openai | 2026-07-07 | 175.1B | 4T |
| #11 | Claude 4.6 SonnetAnthropic | 2026-07-07 | 157.8B | 7T |
| #12 | Gemini 3 Flash PreviewGoogle | 2026-07-07 | 152.1B | 4.1T |
#1
Hy3Tencent
- Source day
- 2026-07-07
- Day
- 858.4B
- 30d total
- 926.9B
#2
Deepseek V4 FlashDeepseek
- Source day
- 2026-07-07
- Day
- 806.2B
- 30d total
- 20.9T
#3
Minimax M3Minimax
- Source day
- 2026-07-07
- Day
- 665.9B
- 30d total
- 17.3T
#4
Mimo V2.5Xiaomi
- Source day
- 2026-07-07
- Day
- 665.6B
- 30d total
- 17.7T
#5
Deepseek V4 ProDeepseek
- Source day
- 2026-07-07
- Day
- 530.6B
- 30d total
- 9.9T
#6
Glm 5.2Z Ai
- Source day
- 2026-07-07
- Day
- 504.5B
- 30d total
- 7T
#7
Claude 4.7 OpusAnthropic
- Source day
- 2026-07-07
- Day
- 398B
- 30d total
- 10.2T
#8
Claude 4.8 OpusAnthropic
- Source day
- 2026-07-07
- Day
- 313.3B
- 30d total
- 7.5T
#9
Step 3.7 FlashStepfun
- Source day
- 2026-07-07
- Day
- 192.5B
- 30d total
- 5.2T
#10
Gpt 5.5Openai
- Source day
- 2026-07-07
- Day
- 175.1B
- 30d total
- 4T
#11
Claude 4.6 SonnetAnthropic
- Source day
- 2026-07-07
- Day
- 157.8B
- 30d total
- 7T
#12
Gemini 3 Flash PreviewGoogle
- Source day
- 2026-07-07
- Day
- 152.1B
- 30d total
- 4.1T
Top models
Public momentum leaders
#1
Qwen/Qwen3-0.6B
Qwen
28.1M1.4K
transformerssafetensorsqwen3
View source#2
Qwen/Qwen3-8B
Qwen
16.9M1.2K
transformerssafetensorsqwen3
View source#3
facebook/opt-125m
14M273
transformerspytorchtf
View sourceWeekly briefing
Rankings shift every week. Read the movement, not the noise.
Models trade places on OpenRouter constantly. Get the week's ranking moves and the notable receipts in your inbox.
OpenRouterPricing and context windows
Hugging FaceDownloads, likes, and model tags
No fake usage claimsNo global token burn unless sourced or opt-in
Workhorse board
What proven workhorses cost
Ranked by demonstrated ability to do work — the models real OpenRouter traffic routes to on the latest complete day — joined to their current prices. Cheap-but-unused never makes this board.
| Model | Routed / day | Input | Output | Context |
|---|---|---|---|---|
| Hy3Tencent | 858.4B tokens | $0.20/M | $0.80/M | 262.1K |
| Deepseek V4 FlashDeepseek | 806.2B tokens | $0.09/M | $0.18/M | 1M |
| Minimax M3Minimax | 665.9B tokens | $0.30/M | $1.20/M | 1M |
| Mimo V2.5Xiaomi | 665.6B tokens | $0.105/M | $0.28/M | 1M |
| Deepseek V4 ProDeepseek | 530.6B tokens | $0.435/M | $0.87/M | 1M |
| Glm 5.2Z Ai | 504.5B tokens | $0.93/M | $3.00/M | 1M |
| Claude 4.7 OpusAnthropic | 398B tokens | $5.00/M | $25.00/M | 1M |
| Claude 4.8 OpusAnthropic | 313.3B tokens | $5.00/M | $25.00/M | 1M |
| Step 3.7 FlashStepfun | 192.5B tokens | $0.20/M | $1.15/M | 256K |
| Gpt 5.5Openai | 175.1B tokens | $5.00/M | $30.00/M | 1.1M |
| Claude 4.6 SonnetAnthropic | 157.8B tokens | $3.00/M | $15.00/M | 1M |
| Gemini 3 Flash PreviewGoogle | 152.1B tokens | $0.50/M | $3.00/M | 1M |
Context board
Largest published context windows
| Model | Context | Input | Output | Full context |
|---|---|---|---|---|
| Meta: Llama 4 ScoutMeta Llama | 10M | $0.10/M | $0.30/M | $1.00 |
| xAI: Grok 4.20 Multi-AgentX Ai | 2M | $1.25/M | $2.50/M | $2.50 |
| xAI: Grok 4.20X Ai | 2M | $1.25/M | $2.50/M | $2.50 |
| OpenAI GPT Latest~openai | 1.1M | $5.00/M | $30.00/M | $5.25 |
| OpenAI: GPT-5.5 ProOpenai | 1.1M | $30.00/M | $180.00/M | $31.50 |
| OpenAI: GPT-5.5Openai | 1.1M | $5.00/M | $30.00/M | $5.25 |
| OpenAI: GPT-5.4 ProOpenai | 1.1M | $30.00/M | $180.00/M | $31.50 |
| OpenAI: GPT-5.4Openai | 1.1M | $2.50/M | $15.00/M | $2.63 |
| Google: Gemini 3.1 Pro Preview Custom ToolsGoogle | 1M | $2.00/M | $12.00/M | $2.10 |
| Z.ai: GLM 5.2Z Ai | 1M | $0.93/M | $3.00/M | $0.9752 |
| MiniMax: MiniMax M3Minimax | 1M | $0.30/M | $1.20/M | $0.3146 |
| Google: Gemini 3.5 FlashGoogle | 1M | $1.50/M | $9.00/M | $1.57 |
Momentum board
Open model momentum
| Model | Downloads | Likes | Tags |
|---|---|---|---|
| Qwen/Qwen3-0.6BQwen | 28.1M | 1.4K | transformers, safetensors |
| Qwen/Qwen3-8BQwen | 16.9M | 1.2K | transformers, safetensors |
| facebook/opt-125mFacebook | 14M | 273 | transformers, pytorch |
| openai-community/gpt2Openai Community | 13.4M | 3.3K | transformers, pytorch |
| Qwen/Qwen2.5-7B-InstructQwen | 12.7M | 1.4K | transformers, safetensors |
| Qwen/Qwen2.5-1.5B-InstructQwen | 11.5M | 761 | transformers, safetensors |
| trl-internal-testing/tiny-Qwen2ForCausalLM-2.5Trl Internal Testing | 11M | 11 | transformers, safetensors |
| Qwen/Qwen3-Embedding-0.6BQwen | 10.6M | 1.1K | sentence-transformers, safetensors |
| Qwen/Qwen3-4BQwen | 9.7M | 648 | transformers, safetensors |
| meta-llama/Llama-3.2-1B-InstructMeta Llama | 9.3M | 1.5K | transformers, safetensors |
| meta-llama/Llama-3.1-8B-InstructMeta Llama | 9M | 6.2K | transformers, safetensors |
| deepseek-ai/DeepSeek-R1Deepseek Ai | 8.6M | 13.4K | transformers, safetensors |