Model signals
Model Signals
Real usage patterns from OpenRouter's public rankings and app leaderboards: which models carry the most traffic, which apps and agents send it, and what it actually costs.
Source: OpenRouter (openrouter.ai/rankings), as of 12 August 2026. Live rankings · Live app rankings · Models API
Method
- Every number on this page traces to a public OpenRouter page or its public, keyless Models API — no proprietary probe methodology, no third-party dataset.
- The rankings and app leaderboards below are a dated, manual snapshot, not a live automated pipeline — OpenRouter's terms prohibit automated scraping of the rankings UI. Pricing is refreshable live via the public Models API.
- Read this as "what people actually route through OpenRouter," not a quality or safety benchmark — token volume reflects usage and price sensitivity, not correctness.
Top models by weekly token volume
| # | Model | Author | Weekly tokens | Change |
|---|---|---|---|---|
| 1 | DeepSeek V4 Flash 0731 | deepseek | 9.39T | +302% |
| 2 | Hy3 | tencent | 8.94T | +83% |
| 3 | DeepSeek V4 Flash 0423 | deepseek | 5.72T | +19% |
| 4 | MiMo-V2.5 | xiaomi | 5.31T | +2% |
| 5 | GPT-5.6 Luna | openai | 4.67T | +90% |
| 6 | GLM 5.2 | z-ai | 3.64T | +27% |
| 7 | DeepSeek V4 Pro | deepseek | 2.58T | +17% |
| 8 | Nemotron 3 Ultra (free) | nvidia | 2.36T | +0% |
| 9 | Gemini 3.6 Flash | 2.32T | +434% | |
| 10 | Laguna S 2.1 (free) | poolside | 1.85T | +58% |
Use-case categories (share of spend)
OpenRouter groups real request volume into task categories, ranked by share of spend rather than raw count.
- Classification
- Content Writing
- Q&A & Knowledge
- Research & Reports
- Summarization
- Workflow Execution
- Code Generation
- File I/O
- Shell Execution
- Debugging
- Code Review
- Frontend & UI
Top apps and agents, by tokens routed
| # | App | Category | Tokens |
|---|---|---|---|
| 1 | Hermes Agent Open-source self-improving agent by Nous Research, persistent memory across sessions | Personal Agents, CLI Agents | 1.45T |
| 2 | Claude Code Anthropic's agentic coding tool | CLI Agents | 356B |
| 3 | Kilo Code Open-source coding agent for VS Code, JetBrains, and CLI | CLI Agents, IDE Extensions | 238B |
| 4 | Cline Open-source IDE coding agent | IDE Extensions, CLI Agents | 207B |
| 5 | OpenClaw Open-source agent connecting to messaging apps and taking real actions | Personal Agents, CLI Agents | 161B |
| 6 | Framer No-code AI website builder | — | 120B |
| 7 | pi Coding agent | CLI Agents | 119B |
| 8 | Zazen (Freebuff fork) freebuff.com | — | 110B |
| 9 | Oh-My-Pi omp.sh | CLI Agents | 86.9B |
| 10 | Nous Research API Nous Research open-source AI research API | General Chat | 72B |
Fastest-growing this week
| App | Tokens | Change |
|---|---|---|
| Hermes Agent | 10T | +26% |
| Framer | 759B | +580% |
| Cline | 1.41T | +82% |
| Claude Code | 2.3T | +26% |
| Zazen (Freebuff fork) | 454B | >999% |
| HighLevel | 459B | >999% |
| Nous Research API | 511B | +310% |
| Oh-My-Pi | 444B | +188% |
Pricing and context (live catalog)
Pulled from OpenRouter's public, keyless Models API (/api/v1/models) — the same catalog used for real API integration, not a scrape of the rankings UI.
| Model | Prompt $/M | Completion $/M | Context |
|---|---|---|---|
| Claude Opus 5 | $5.00 | $25.00 | 1,000,000 |
| Claude Sonnet 5 | $2.00 | $10.00 | 1,000,000 |
| GPT-5.6 Sol | $5.00 | $30.00 | 1,050,000 |
| GPT-5.6 Terra | $1.00 | $6.00 | 1,050,000 |
| GPT-5.6 Luna | $0.10 | $0.60 | 1,050,000 |
| Gemini 3.1 Pro | $2.00 | $12.00 | 1,048,576 |
| Gemini 3.6 Flash | $1.50 | $7.50 | 1,048,576 |
| DeepSeek V4 Flash | $0.08 | $0.18 | 1,048,576 |
| DeepSeek V4 Pro | $0.63 | $1.26 | 1,048,576 |
| Grok 4.5 | $2.00 | $6.00 | 500,000 |
| Kimi K3 | $3.00 | $15.00 | 1,048,576 |
Other public sources worth cross-checking
Blind pairwise human-preference Elo rankings across text, vision, web-dev, and agent tasks. Measures preference, not correctness -- has documented biases toward verbosity and formatting.
Artificial AnalysisComposite Intelligence Index across 9 independent evals (GDPval-AA, Terminal-Bench, GPQA Diamond, and others), plus speed, latency, and cost-per-task tracked across roughly 600 models.
Hugging Face HubDownload and like counts for open-weight models specifically. A legitimate usage proxy for the open-weight tier; does not cover closed frontier models.