Free models are tried first
Among the models that fit the task, ones with remaining free quota go first.
AI model gateway
Free models first. Automatic failover. No per-provider signup.
FreeModel is an AI model gateway that gives you a single API key for dozens of model providers, routes each task to the best available model, tries free models first, and falls back automatically when a provider fails or runs out of quota.
Among the models that fit the task, ones with remaining free quota go first.
A rate limit or an exhausted quota moves the request to the next candidate.
Five providers no longer means five accounts and five billing pages.
The live list is public and needs no key:
GET /v1/models · which ones are free →
# Anthropic-compatible clients: Claude Code, and others
export ANTHROPIC_BASE_URL="https://freemodel.online/api/gateway"
export ANTHROPIC_API_KEY="sk-your-key"
# OpenAI-compatible clients: Codex CLI, Cursor, Cline, SDKs
export OPENAI_BASE_URL="https://freemodel.online/v1"
A desktop app that runs several AI agents at once on Windows, with no setup. It ships pointing at FreeModel, so you never register with a provider or paste a key.
Both are free. No account needed to install.
A direct connection to one provider is the shortest path, and it is also the most fragile one. Everything that can change on the provider's side — a quota running out, a key being rotated, a model being retired, a price moving — arrives in your application as a failure, and each of those failures has a different cause and a different fix. A gateway exists to absorb that class of event before your code sees it, and to record what happened, so the next occurrence is a lookup rather than an investigation.
FreeModel adds exactly one hop and takes those events out of your error path. Of the 477 listed models, 66% accept text or multimodal input and 34% generate images, so a text workload is comparing inside the larger half of the catalogue. Among the 317 text models, 95% have tool calling enabled and 73% support reasoning, which is why an agent can be pointed at almost any of them without checking capabilities first; 29% accept image input, so a request carrying a screenshot narrows the pool without emptying it. The router filters that pool to what the request actually needs, spends free quota wherever any remains, and moves on when a provider refuses.
What you give up is knowing which model answered. A route name is a statement of intent, not a model id, and only 8% of the listed models are combination routes — the 38 entries that take this trade explicitly. If you need to reproduce a result exactly, or to name a model for compliance reasons, send a specific id instead and let the gateway do nothing but forward it.
ANTHROPIC_BASE_URL to https://freemodel.online/api/gateway
and ANTHROPIC_API_KEY to your key. For an OpenAI-shaped client, set
OPENAI_BASE_URL to https://freemodel.online/v1.auto/best-coding for
code, auto/best-reasoning for reasoning, auto/best-free when
the request must not cost anything.ANTHROPIC_BASE_URL or
OPENAI_BASE_URL — and which model name you send. Everything above that line
stays as it was, and switching back is the same edit in reverse.auto/best-coding picks the best available model for code this minute, and
may pick a different one next month. When you need the identical model every time, send
a specific id and the router gets out of the way.