The 7.4B-Token Free Lunch: One Open Source Router Turns 34 Free API Tiers Into a Single Endpoint
The 7.4B-Token Free Lunch: One Open Source Router Turns 34 Free API Tiers Into a Single Endpoint
Check your password manager. How many LLM API keys are sitting in there? One for OpenAI, one for Google, maybe one for Anthropic — each attached to a free tier that sounds generous until you actually hit it. A few hundred thousand tokens here, a few thousand requests a day there. Individually, every free tier is a toy.
Somebody stacked 34 toys into a tower. freellmapi, an open-source project on GitHub, aggregates the free tiers of 34 LLM providers — Google, Groq, Cerebras, Mistral, Cloudflare, Zhipu's Z.ai, NVIDIA, even ModelScope's Qwen3 and DeepSeek V4 — behind a single OpenAI-compatible /v1 endpoint. Combined, those tiers add up to roughly 7.4 billion tokens of free inference per month, spread across 635 free model endpoints and 474 model families.
Here's the part I find telling: this morning's digest said 28 providers and 4B tokens. By the time I opened the README, it said 34 and 7.4B. The project is outrunning the news about it. That alone is a signal.
The hard part was never the free tokens
Free quotas have never been scarce. What's scarce is the will to wrangle 34 SDKs, 34 rate-limit policies, and 34 different ways for a request to fail.
freellmapi shoves all that grunt work into a router layer:
- Smart routing: each request picks the best available free model; when a provider rate-limits you, it fails over to the next one automatically
- Per-key accounting: tracks usage per key so you never blow past any single provider's free ceiling
- Encrypted key storage: your keys sit encrypted on the server
- Self-updating catalog: the model list syncs from a signed feed — new models, quota changes, and compatibility fixes land without a
git pull
It's not even locked to the OpenAI ecosystem. A custom provider slot accepts any OpenAI-compatible endpoint — local Ollama, llama.cpp, vLLM, whatever you're already running. Deployment is Docker one-liner territory, with macOS and Windows desktop apps on top. Point any OpenAI client library at it and you're done.
Think of it as a payment router that draws from 34 credit cards with free credit limits, automatically swiping whichever one isn't maxed out. The thing it saves you isn't money, exactly — it's the mental overhead of managing 34 accounts.
Why I say it's turning "free" into a business
freellmapi isn't a pure charity project. Created in April 2026, it's already at 19.7k stars with a premium tier: $19/year for the live catalog.
Free installs get a monthly snapshot — a new model reaches your router 30 days after it joins the live feed. Premium routers get it the same day. The model catalog itself is its own signed distribution channel.
That's a clever move: the free tier is the funnel, premium sells speed-of-information. But it raises an honest question — the free quotas belong to the providers, and the aggregator's business model sits on top of someone else's free policy. The day Google tightens its free tier, the "7.4B" shrinks. When your business runs on other people's free stuff, the risk is never in your hands.
Four things to check before you free-ride
The author says it's for personal experimentation only. That's not modesty — it's a disclaimer. In practice:
- Free tier means rate limits. Millions of tokens a month sounds huge until you divide it across 34 providers, where a single one might allow a few thousand requests a day. Fine for batch jobs. Forget production traffic.
- Encrypted keys ≠ privacy. Your requests still go to the providers. Encryption only protects you if the server gets dumped. Don't feed it sensitive data.
- The aggregator is a single point of failure. Router dies, all 34 providers die together. You traded 34 small problems for one bigger one.
- Free policies change without notice. Quota tweaks and model retirements sync through the feed — but your batch scheduling has to keep up.
Speaking of aggregation: it's not just APIs
Same day, another story: Anthropic launched a community plugin marketplace for Claude, and a community-curated Agent skill library has crossed 1,000 skills, compatible with mainstream tools like Claude Code, Codex, and Gemini CLI.
Same trend, different layer: in 2026, developers don't build wheels anymore — they aggregate ready-made ones. API aggregation on one level, skill-marketplace aggregation on another. Convenient, but note the marketplace is currently a read-only mirror with submissions going through an official channel. Don't expect one-click installs yet.
What you can do today
Don't wire freellmapi into production. Do these two things instead:
- Run a few days of batch jobs through it and measure the real throughput and stability of the free tiers you actually care about. The README says it best: each free tier is a toy on its own — but which toys fit your workload, only you can test.
- Go browse that 1,000-skill library, pick the three closest to your daily work, and actually try one tomorrow while coding. Trying three beats bookmarking a hundred.
Shortcuts to saving money and time exist — but they all have the same fine print: verify at small scale before you trust them at scale.
Sources
✨ Drafted by DeepSeek, polished by Claude.