What developers actually route through AnyRouter

AnyRouter's catalog covers 170+ models across 28+ providers, addressed by a single provider/model string — anthropic/claude-opus-4-8, z-ai/glm-5.2, google/gemini-2.5-pro. Models are organized by the organization that owns them, not by which upstream serves them, and that ownership is lopsided: a handful of labs account for most of the catalog, and a long tail of smaller labs each contribute a model or two.

That split matters because a model can also come in two SKU classes on the Cloudflare side alone — hosted first-party (@cf/owner/model) or partner-hosted (owner/model, no @cf prefix) — before you even get to a third-party upstream like DeepInfra or a direct provider API. Breadth in the catalog doesn't mean breadth in what gets called.

Calling any model is one request against a single, OpenAI-compatible endpoint:

curl https://anyrouter.dev/api/v1/chat/completions \
  -H "Authorization: Bearer sk-ar-v1-..." \
  -H "Content-Type: application/json" \
  -d '{"model": "z-ai/glm-5.2", "messages": [{"role": "user", "content": "Hello"}]}'

But traffic through the free routes — no credit charged, not a BYOK call — doesn't spread evenly across the catalog. Two models account for the overwhelming majority of free requests, all-time through July 11, 2026:

Request count alone hides what's actually happening. GLM-5.2 leads on requests and dwarfs everything else on tokens served — north of 200M free tokens — which is the signature of a model running real, sustained work, not one-off tests. gpt-4o-mini leads on requests too, but with a fraction of the tokens per call: it's the classifier-and-glue model, cheap enough to sit inside a loop where each call is a sentence or two.

Those are two different jobs wearing the same 'free model' label, and the request count on its own doesn't tell you which one a given model is doing — the tokens-per-request ratio does.

A meaningful share of production traffic doesn't come from a developer typing a prompt — it comes from a coding agent making the call on their behalf. Claude Code, Codex, and Cursor all speak different dialects against the same catalog:

Because every dialect terminates at the same catalog, a request from Claude Code and a request typed by hand can resolve to the same model — the only difference visible in usage analytics is the app attribution header the client sent, not a separate code path.

The free capacity behind those request counts isn't purchased in bulk — it's a shared key pool. Members donate spare quota from a provider's own free tier, quotas sum into one pool, and donors earn a share back plus a free plan tier. The more capacity gets donated, the more the free-tier numbers above can grow without a paywall behind them.

• A wide catalog and a concentrated traffic pattern aren't a contradiction — reach and popularity are different things
• Request count and token count together identify what job a model is actually doing, not just how often it's called
• Model ownership in the catalog skews toward a handful of labs, but the smaller entries still get called — just less often
• Free-tier volume scales with donated capacity, not with a fixed budget