If you still bet on MMLU and HumanEval leaderboards while ignoring four-figure monthly API invoices, the OpenRouter weekly rankings sorted by rolling 7-day token throughput deliver a sharper answer. For the week of May 18–24, 2026, global weekly volume hit 28.9 trillion tokens, with Chinese models leading the US for a fourth consecutive week. DeepSeek V4-Flash topped the chart at 3.43T, while Anthropic shows a classic falling token share but high dollar revenue premium paradox—with roughly 12% of tokens still generating about 46% of platform revenue. Programming now accounts for more than half of tracked usage per the a16z joint report. This guide is for developers and tech leads running multi-model routing. It covers data sources, the Top 10 weekly table, the dual truth of vendor share, the inverse-benchmark finding, and a six-step billing-driven selection checklist you can run every Monday.
OpenRouter is one of the largest neutral AI model API aggregators: 300+ models, 60+ providers, over 8 million users, and roughly 100 trillion tokens processed per month. Its leaderboard ranks by recent weekly token totals (input + output). Money spent and traffic routed cannot be polished for a launch keynote. Compared with fixed lab eval sets, real call volume better reflects how developers vote with their wallets in Agent workflows, batch programming, and multi-turn tool chains.
The conclusion is straightforward: benchmarks measure peak capability; billing measures habit. A 0.3-point MMLU gain may not move next month's invoice. But when DeepSeek Flash costs roughly one-fiftieth of Opus per token, Agent loops reroute immediately. That is why technical leads who own gateway policy increasingly treat OpenRouter's weekly export as a first-class signal—on par with latency dashboards and error-rate SLOs.
Vendor benchmarks are still useful. They tell you what a model can do in isolation under controlled prompts. They do not tell you what your organization will actually call on Tuesday at 2 a.m. when a cron-driven Agent retries a failed shell command for the forty-third time. At that moment, price per million tokens, cache hit rate, and provider uptime dominate. The weekly leaderboard encodes those tradeoffs at ecosystem scale.
Benchmarks test ceilings; bills test habits: Leaderboard scores do not predict your gateway routing. Token price and latency do.
Free routes distort paid intent: Zero-dollar models like Owl Alpha spike weekly rankings because developers prioritize "good enough" over "best." Selection that ignores free tiers overstates closed flagship share.
Programming is now the largest single use case: The OpenRouter and a16z 2025 AI Usage Report (based on ~100T tokens of anonymized metadata) shows programming tasks rising from about 11% in early 2025 to over 50%. Top-ranked models skew toward coding and Agent-friendly tiers.
Stability and latency beat peak reasoning: Production Agents care about API response time and tool-call success rates more than a perfect score on a single math puzzle.
Weekly windows catch breakouts: Rolling 7-day stats surfaced Hy3 Preview still growing +16% week-over-week after its free tier ended—faster signal than monthly averages.
Investors and media already track token metrics: OpenRouter's valuation sits around 26× price-to-sales. The leaderboard has graduated from a technical curiosity to a commercial barometer. Ignoring it means ignoring where real money flows.
"It is not who is smartest, but who gets called most—weekly token volume is the thermometer of AI adoption, developer trust, and market acceptance."
The table below summarizes OpenRouter public data for May 18–24, 2026 (7-day rolling token throughput, consistent with openrouter.ai/rankings). Platform weekly volume was about 2.4 trillion a year earlier; at 28.9 trillion today, that is roughly a 12× increase. AI applications have entered a scale-out phase—not a pilot phase where a single flagship model carries every workload.
The geographic split matters as much as the headline total. China-model weekly volume at 9.223T is nearly double US-model volume at 4.93T, and China's +19.89% WoW growth outpaced the global +7.4% average. US models still grew in absolute terms (+16.27%), but their share of the pie shrank because the overall market expanded faster on the low-cost side. For anyone forecasting capacity or negotiating enterprise API contracts, that asymmetry is the story—not whether Claude or GPT still tops a static eval chart.
| Metric | Data | WoW change | Reading |
|---|---|---|---|
| Global weekly volume | 28.9 trillion tokens | +7.4% (fifth consecutive weekly rise) | Total expansion outpaces single-model share shifts |
| China model weekly volume | 9.223 trillion tokens | +19.89% | Growth well above the global average |
| US model weekly volume | 4.93 trillion tokens | +16.27% | Absolute growth, but share compressed |
| China vs US | China #1 for four straight weeks | China share ~45%+ | China model traffic was under 2% in early 2025 |
Citable hard numbers: (1) Global weekly volume 28.9T, WoW +7.4%, fifth consecutive weekly increase. (2) China model weekly volume 9.223T, WoW +19.89%. (3) US model weekly volume 4.93T, WoW +16.27%. (4) Platform monthly token scale is on the order of 100T (OpenRouter public figures). Update tail ranks from the live site when publishing.
Five straight weeks of global growth suggest demand is broad-based—not a single-model viral event. That makes the leaderboard a safer input for capacity planning than one-off launch spikes. If your product roadmap assumes "we will default to the benchmark winner," these macro rows are the reality check.
Methodology note: Weekly rankings use a rolling 7-day window, not calendar weeks. Model and vendor market-share views are on the same page. Dollar revenue share and token share are shown separately; Section 04 explains why they diverge.
The top ten are sorted by weekly token volume. Three DeepSeek models placed in the upper tier; the series totaled about 5.74T tokens (WoW +25.9%), ranking first among vendors by token volume for two consecutive weeks. Kimi K2.6 had been sixth the prior week but dropped out of the top ten—weekly rankings rotate fast when a breakout cools.
DeepSeek-V4-Flash at 3.43T (+66% WoW) is the clearest signal in the table. A two-thirds weekly jump at already-trillion scale means production traffic is consolidating on a single ultra-cheap route—not experimental side projects. Tencent Hy3 Preview holding 3.07T after promotional pricing ended suggests retention, not just a free-tier spike. Claude Sonnet 4.6 remains the highest-ranked US closed model, which fits its role as a dependable enterprise coding tier with million-token context.
Owl Alpha at rank five is the free-tier wildcard: 1.15T tokens at $0 list price proves that Agent-specialized anonymous routes can compete with paid flagships on volume alone. Operations teams should treat Owl and similar promos as sandbox lanes—powerful for toolchain validation, risky for regulated data without explicit policy guardrails.
| Rank | Model | Vendor | Weekly tokens | WoW | Notes |
|---|---|---|---|---|---|
| 1 | DeepSeek-V4-Flash | DeepSeek (China) | 3.43T | +66% | Agent workflow default; ultra-low price |
| 2 | Tencent Hy3 Preview | Tencent (China) | 3.07T | +16% | Still growing after free tier ended |
| 3 | Claude Sonnet 4.6 | Anthropic (US) | 1.35T | — | 1M context; enterprise coding workhorse |
| 4 | DeepSeek-V3.2 | DeepSeek (China) | 1.31T | — | Low-cost long tail; roleplay active |
| 5 | Owl Alpha (anonymous) | OpenRouter | 1.15T | +29% | Free Agent-specialized; 1M context |
| 6 | Gemini 3 Flash Preview | Google (US) | 1.06T | — | Multimodal; academic and medical |
| 7 | DeepSeek-V4-Pro | DeepSeek (China) | 1.00T | — | Matrix flagship (series total 5.74T) |
| 8 | MiniMax M2.7 | MiniMax (China) | 806B | — | Long-context value pick |
| 9 | Grok 4.1 Fast | xAI (US) | 721B | — | 2M context; strong in legal |
| 10 | Step 3.5 Flash | StepFun (China) | 673B | — | Fast and cheap; batch workloads |
[High value · low volume] Anthropic Claude Opus → enterprise complex reasoning, strong willingness to pay [Mid value · mid volume] Google Gemini Flash → multimodal, academic and search ecosystem [Ultra-low · high volume] DeepSeek / MiniMax / StepFun → Agents, programming, batch jobs
The three-tier structure above is how finance and platform teams should read the same table engineers see as a model picker. Opus-class SKUs earn margin on hard problems; Gemini Flash sits in the multimodal middle; DeepSeek, MiniMax, and StepFun absorb the long tail of Agent iterations. Grok 4.1 Fast and MiniMax M2.7 show that long-context niches (legal, research) still sustain billion-token weeks without topping the chart.
Ranks 1–2 and 5 align with industry reporting from late May 2026. Ranks 3–4, 6, and 8–10 cross-check against the public OpenRouter leaderboard and contemporaneous analyst notes. V4-Pro weekly volume can be inferred from the series total of 5.74T minus V4-Flash and V3.2. Always pull the latest week from the live site before citing exact positions.
Cross-validation: Top-two and Owl Alpha figures were reported by mainstream financial media on 2026-05-25. Mid-table rows were reconciled against OpenRouter public rankings. Rankings shift weekly; treat this snapshot as a decision input, not a permanent scoreboard.
Token rankings alone understate Anthropic's monetization. Revenue share alone overstates its traffic dominance. OpenRouter shows token share and dollar revenue share side by side, exposing how pricing creates real market tiers.
| Dimension | Anthropic | DeepSeek family | Reading |
|---|---|---|---|
| Token share | ~12% (was ~25% a year ago) | Series weekly 5.74T; #1 vendor by tokens | Traffic leadership shifting to low-cost open lines |
| Dollar revenue share | ~46% | Ultra-low unit price; revenue share far below token share | Enterprises still pay premium for Claude |
| Flagship SKU | Claude Opus 4.6 ~$25M monthly revenue scale | V4-Flash driving mass Agent calls | Opus token volume is orders of magnitude below DeepSeek |
| China model timeline | Under 2% in early 2025 → first surpassed US in Feb 2026 → ~45%+ by May 2026 | Open source plus ultra-low pricing reshaped global routing | |
This is the Anthropic premium paradox: roughly 12% of tokens but about 46% of revenue. Developers route bulk Agent and programming work to Flash-tier models while enterprises keep paying for Opus and Sonnet on hard reasoning tasks. Anthropic's token share fell from about 25% a year ago to 12%—a dramatic traffic shift—yet dollar share near 46% shows pricing power on the workloads that still matter for margin.
DeepSeek's inverse profile explains the other side. The family drove 5.74T weekly tokens—more than the entire US cohort in aggregate terms for several SKUs—but unit economics keep revenue share far below token share. That is not a failure; it is the deliberate strategy of open, ultra-low-cost lines capturing Agent scale. V4-Flash is the volume engine; V4-Pro and V3.2 absorb adjacent workloads without forcing every call through the most expensive tier.
The OpenRouter and a16z 2025 AI Usage Report adds a counterintuitive finding: benchmark scores and actual market share are nearly inversely related. Teams optimize for inference cost, API stability, and Agent fit—not leaderboard max scores. That aligns with programming exceeding half of all traffic and Flash models dominating the weekly top ten. When the largest use case is code generation, diff application, and terminal tooling, "best on MMLU" is a weak proxy for "best on my invoice."
China's trajectory reinforces the point. Under 2% of platform traffic in early 2025, Chinese models first surpassed US volume in February 2026, and by May held roughly 45%+ share. Open weights plus aggressive API pricing moved faster than Western benchmark narratives. For routing policy, treat token share as your volume signal and revenue share as your margin signal. A gateway that sends every loop to Opus because it "wins benchmarks" will burn budget while the market has already moved to layered tiers.
Turn the leaderboard from news into ops. Run these steps weekly and pair them with the OpenRouter trends selection guide and OpenClaw multi-model routing docs. The goal is not to chase rank changes reactively—it is to keep your gateway rules, eval harness, and cost model aligned with where billions of tokens already flow.
Start by exporting your own OpenRouter or direct-vendor invoice alongside the public leaderboard. If your internal token mix diverges sharply from global Top 10—for example, you are still 80% on a single Sonnet route while the ecosystem is 50%+ programming on Flash tiers—you have a rebalancing opportunity before finance asks why Q3 API spend doubled.
Open openrouter.ai/rankings every Monday: Log global weekly total, China vs US share, and Top 10 moves. Paste the four macro rows from Section 02 into an internal weekly report.
Split your invoice: tokens vs dollars: If most tokens hit Flash-tier models but most spend sits on Claude, routing is already tiered—write that explicitly in gateway rules so Opus is not used for bulk completion.
Map scenarios to three tiers: Agent and batch jobs → DeepSeek-V4-Flash; enterprise complex reasoning → Claude Opus/Sonnet; multimodal → Gemini Flash.
Watch new top-ten entrants: Moves like Hy3 Preview and Owl Alpha often precede the next breakout. Use free tiers for non-sensitive prototypes before committing production keys. Document why a model entered your allowlist—price, context length, or tool-call reliability.
Calibrate evals to programming >50%: Selection meetings should weight SWE-bench, Terminal-Bench, and real failure rates over MMLU deltas. If your team does not ship code, swap in domain-specific regression suites—but drop vanity leaderboard slides.
Evaluate hybrid compute: When monthly API spend exceeds a high-spec Mac rental, move long-session CLI Agents and Ollama prefill to an SSH-dedicated node. OpenRouter keeps peak elasticity for Opus-class calls and burst traffic. See rental rates for M4 Pro and M4 Max tiers.
Step six is where infrastructure meets routing economics. When monthly API spend crosses the cost of a dedicated M4 Pro or M4 Max Mac Mini rental, shifting Ollama prefill, long-context caching, and repetitive Agent loops to local inference on an SSH node often flattens the bill curve. OpenRouter remains valuable for closed flagships, burst capacity, and models you cannot self-host—but it should not be the only layer in the stack.
A sleeping laptop or a cheap Linux VPS cannot sustain 12-hour Agent loops or run xcodebuild and notarytool. Thermal throttling on a personal MacBook stretches iOS builds; Windows and Linux hosts break Keychain signing paths. Pairing weekly leaderboard review with a fixed execution environment beats chasing a single "best model" every Monday.
Teams that need stable SSH long sessions, Keychain isolation, and predictable bandwidth for iOS CI/CD and AI Agent automation should write OpenRouter routing in the gateway and place heavy loads on a dedicated cloud Mac rather than routing every token through public APIs. The weekly rankings tell you which models the world calls; your execution layer decides whether those calls run reliably at 3 a.m.
NodeMini Mac Mini cloud rental fits that execution layer: swap API keys or model endpoints while SSH nodes and CI labels stay fixed. When DeepSeek Flash wins another +66% week, you change a gateway rule—not rebuild a laptop environment. See the help center for access setup and compute ordering for instant provisioning.
Benchmarks compare single-skill ceilings. OpenRouter sorts by rolling 7-day token throughput, reflecting real paid and free calls. For budget forecasting, market share, and Agent batch selection, billing data is usually more honest. Use both, but for different jobs.
Claude is priced far above DeepSeek and comparable open lines. Enterprise users pay for complex reasoning, but mass Agent and programming traffic flows to low-cost models. That creates high value · low volume and ultra-low · high volume tiers running in parallel. Route with both token and dollar share in view.
OpenRouter handles elastic multi-model routing and weekly tracking. Sensitive repo prefill and long-session CLI Agents can live on an SSH-reachable dedicated cloud Mac—see SSH session isolation and rental rates. Closed flagships stay on API; the local node cuts external bill dependency.