9 billion free tokens

• Almost every AI lab gives away a free tier: a few million tokens a month, a few thousand requests a day. On its own each one is small. Stacked together, they add up to a lot.
• OmniRoute and FreeLLMAPI are open-source (MIT) gateways that do the stacking for you. You add your own provider accounts; they give you one endpoint, route each request to a model that still has quota, and fall over to the next one when a provider rate-limits you.
• OmniRoute estimates ~1.62 billion free tokens a month across the free tiers it catalogues; FreeLLMAPI lists ~7.4 billion. Together that's the ~9 billion headline (our arithmetic, and an upper bound: section 7).
• Both can sit behind Claude Code. You keep Claude Code's interface; the work is done by free models such as Gemini, Mistral, GLM, Qwen or DeepSeek, not by Anthropic's Claude.
• OmniRoute also compresses what you send: 15–95% fewer tokens depending on the mode, about 89% on average for its stacked mode, by its own maths.

An app that talks to a language model sends text to an API and gets text back. Every provider has its own API, its own keys and its own limits. Most of them, though, also offer an OpenAI-compatible endpoint, so the same request shape works almost everywhere.

A gateway (or router) is a small server you run yourself, usually on your own laptop. Your tools talk only to the gateway; the gateway talks to the providers.

How Claude Code fits in

Claude Code normally sends its requests to Anthropic. It can be pointed at a different server instead, and both gateways support that: FreeLLMAPI speaks Anthropic's Messages API (/v1/messages) over its router, and OmniRoute launches Claude Code through itself with one command. The agent, the terminal UI and your workflow stay the same; the model answering is whichever free model the gateway picks.

Problem 1: usage limits

Coding agents burn through tokens fast: every file they read, every command output and every long conversation goes back into the model. On a subscription you hit the usage limit mid-task; on a paid API you pay for every token.

Problem 2: free tiers are a mess to use by hand

The free capacity exists, but it's scattered. FreeLLMAPI's README describes the pain as thirty-four different SDKs, thirty-four rate limits, and thirty-four places a request can fail. OmniRoute's lists "10 dashboards, dead API keys and surprise bills". And the landscape changes weekly: providers launch models, retire them and change quotas without notice.

OmniRoute is a "free MIT AI gateway: one endpoint, 375 providers (150+ free), 1,400+ models", with 80.6k stars on GitHub (October 2026). It runs locally with a dashboard at localhost:20128 and an OpenAI-compatible API at localhost:20128/v1.

The compression, in numbers

This is the "95% fewer tokens" in the reel. OmniRoute offers presets, each with its own savings (its own figures):

Bars show the upper end of each range. The stacked mode is the one behind "up to 95%".

combined = 1 − (1 − RTK) × (1 − Caveman_input)
average  = 1 − (1 − 0.80) × (1 − 0.46) = 89.2%
range    = 78.4 – 94.6%

An example from its README, Standard mode: a 69-token explanation of a React re-render becomes "New object ref each render. Inline object prop = new ref = re-render. Wrap in useMemo.", 19 tokens, 72% fewer.

FreeLLMAPI describes itself as "7.4 billion tokens per month. 34 free LLM providers. 635 free model endpoints. All behind one /v1 endpoint", with 40.7k stars on GitHub (October 2026). It runs locally at localhost:3001, with a dashboard, and is built for one person.

A few of its 34 providers: Google, Groq, Cerebras, Mistral, OpenRouter, Cloudflare, Cohere, Z.ai, NVIDIA, Hugging Face and more.

It speaks OpenAI's formats (chat, responses, embeddings, images, audio) and also Anthropic's Messages API, which is what lets Claude Code and the official Anthropic SDKs run against the free pool. Prompt compression exists here too, but it's opt-in.