OpenRouter Open vs. Closed Rankings

A single-file dashboard tracking the daily market share of open-weight vs. closed models in OpenRouter's top-50 token-usage rankings (January 2025 onward), with filters for U.S.- and Chinese-developed models and a by-origin view.

Open index.html in a browser — no server or build tooling required at view time.

Data

File	Contents
data/rankings-<year>.json	Raw responses from OpenRouter's rankings-daily dataset: top 50 public models per day by total tokens (prompt + completion), plus one aggregated other row per day
data/model-classification.json	Per-model classification: developer, origin country (US / China / Other), and weights status (open / closed / unknown)
data/org-origins.json	Org prefix → origin/developer lookup used to auto-classify new models
index.html	The dashboard, with the compact joined dataset embedded by the build step

Automated refresh (GitHub Actions)

.github/workflows/refresh.yml runs daily at 06:30 UTC (and on manual dispatch): fetch → auto-classify new models → rebuild index.html → commit if changed.

API usage is two keyed requests per day (one per calendar-year window), well inside the 30/minute and 500/day limits. Viewing the dashboard costs zero API calls — the dataset is embedded in index.html at build time. The committed data/rankings-*.json files double as an archive independent of the API: OpenRouter's dataset starts at 2025-01-01 (earlier start dates are clamped), and each daily run re-downloads the full official history, so gaps self-heal.

One-time setup after pushing to GitHub:

1. Add a repository secret named OPENROUTER_API_KEY (Settings → Secrets and variables → Actions). Any valid OpenRouter key works.
2. Optionally enable GitHub Pages (Settings → Pages → deploy from branch main, root) to serve the dashboard at a public URL.

How new models are classified

scripts/classify.mjs appends entries for any model that enters the top 50 without touching existing ones — no LLM or manual step in the loop:

• Weights: from the live OpenRouter catalog — hugging_face_id set → open, absent → closed. openrouter/* stealth models and models already delisted from the catalog → unknown, flagged for review.
• Origin: org-prefix lookup in data/org-origins.json (a new qwen/* model is China, a new anthropic/* model is U.S.). Unseen orgs → Unknown, flagged.

Flagged entries appear in the Actions job summary with REVIEW markers in their note field. To correct any entry (a stealth reveal, a post-launch weight release, a new org), edit data/model-classification.json and/or add the org to data/org-origins.json — the classifier never overwrites manual edits. Auto entries carry "confidence": "auto".

Manual refresh

$env:OPENROUTER_API_KEY = "sk-or-..."   # any valid OpenRouter key
node scripts/fetch-data.mjs              # re-downloads data/rankings-*.json
node scripts/classify.mjs                # classifies any new top-50 entrants
node scripts/build.mjs                   # re-embeds the dataset into index.html

Classification method

• Weights: "open" if the model's weights are publicly downloadable as of the generated date (any license), assessed via the OpenRouter catalog's hugging_face_id, Epoch AI's benchmarked-models dataset, and targeted web checks. Openness is applied retroactively (e.g., Grok-2 counts as open across its history; weights released August 2025).
• Origin: country of the developing organization. Community finetunes inherit the base model developer's country; OpenRouter's cloaked stealth models (openrouter/*) are attributed per public reveals, or marked unknown if never revealed.
• Permaslug snapshots (-YYYYMMDD) and :free/:beta-style variants are merged into their base model. Shares are computed within the daily top 50, which covered ~93% of all OpenRouter tokens as of July 2026.