OpenRouter inference provider market share

Who actually serves the tokens — hosts, not model authors. Token shares cover each provider's complete OpenRouter traffic; revenue estimates and model views cover the top 30 models.

Open models onlyThird-party hosts only

Loading…

Provider leaderboard

For the selected range. Δ share compares against the prior equal-length period. Click a row to drill down.

Model head-to-head

Pick a model to watch hosts win and lose routing share on it. Includes all providers regardless of the open-models filter.

No models synced yet.

Methodology

Token totals and shares cover each provider's complete OpenRouter traffic(scraped from OpenRouter's per-provider pages, including their long-tail “Others” volume). Model-level views and revenue estimates cover the top 30models by volume plus each provider's own top models.
“Open models only” subtracts tracked closed-model tokens from each provider's complete total. Closed models (Claude, GPT, Gemini, Grok) are all top-30 tracked, so the subtraction captures nearly all closed traffic.
Tokens are prompt + completion combined, as reported by OpenRouter's per-provider usage charts. A per-provider output-only series is not published.
“Open models only” (default on) treats models authored by as closed, so first-party traffic doesn't drown out the competitive open-weight hosting market.
“Third-party hosts only” (default off) additionally hides providers that author models themselves (DeepSeek, Xiaomi, MiniMax, OpenAI, …) and first-party clouds (Google Vertex, Azure, Bedrock), isolating the neutral hosting market.
Regional and variant endpoints (e.g. google-vertex/us-east5) roll up to their base provider. Providers under 1% of range volume group into “Long tail”.
The current (partial) UTC day is excluded; per-provider history depth varies (~2 months max).
Revenue is an estimateof gross billings for OpenRouter-routed traffic, over each provider's complete volume. Every model named on a provider's page is priced exactly: tokens × that provider's listed prices, split into input / output / cached-input using the model's daily mix (assumed uniform across its providers). The residual “Others” long tail and delisted models (~10% of total volume, varies by provider) are imputed at that provider's own effective $/token over its priced volume — long-tail models tend to be cheaper, so this leans high. Prices are snapshotted daily going forward; days before the first snapshot use the earliest known price. BYOK and negotiated rates are invisible. Treat as ±30%.