I ran out of paid quota three times this year, and each time the "free" backup I planned for didn't behave the way the word "free" promised. It took me a while to notice that "free" in the AI world is actually four completely different things wearing the same label.
1. Harness-locked free. OpenCode's free pool (MiMo-V2.5, HY3, Ling 3.0 Flash Fin, Nemotron 3 Ultra / 3.5 Lightning, Muse Spark 1.2) works, but only through OpenCode's own agent harness — not through its OpenAI-compatible API, even though the docs page still lists API endpoints. If you need to call these models from your own code, they don't exist.
2. Dev-only, quota-unpublished. NVIDIA's Build endpoints are free, but there is no single number that says how many requests you get. Quota depends on the model, your account, and how much traffic the endpoint is seeing. "For development and testing" is doing a lot of heavy lifting in that sentence.
3. Rotating promotions. Cline's free model pool changes without a schedule. A model that was free last month may not be this month, and the pool is only usable from Cline's IDE extension or CLI — not from its API.
4. "User-reported sufficient." Some API resellers list free DeepSeek or GLM endpoints and say the rate limit is "sufficient, per user reports." That's not a quota. That's a rumor with a URL attached.
When someone says "you get 100 requests a day on the free tier," multiply it out:
But two of those numbers are lies by construction. The "4k tokens average" is not yours — it's whatever the provider decides each request includes. And the "$2/M" only holds if the free tier is priced like the paid tier, which it usually isn't: free tiers are capacity that would otherwise sit empty, so they're the first thing cut when real customers show up.
The number that decides whether you can build on a free tier is not its dollar value. It's the rotation interval: how often does the "free" list change? I track a small changelog precisely because the gap between "announced reset" and "the reset actually landed" is where free plans quietly stop being free.
I maintain a small site (biased source, full disclosure: it's mine) called Alca that keeps one row per provider, links every claim to its original source post, and marks each row verified or unverified — so when a number looks too good, you can see in one click whether someone checked the source or copied a tweet. There's a dated reset log for each provider, and the Codex feed in particular regenerates about hourly from a public announcements API.
The practical takeaway is not "use free tiers." It's: pick your stack on the kind of free, not the amount of free. Harness-locked is fine for personal agent use and worthless for an API. Dev-only is fine for a side project and a landmine for anything you'll ship on. Rotating is a treadmill — you have to re-verify every month, and that's the whole reason the changelog exists.
I've stopped planning around any single free pool. The free layer is my overflow, not my foundation — and I re-check its terms every time a reset lands, which is now just a diff of one file instead of a day of tab-shuffling through four provider docs.
The site has a public JSON API under /api/v1 if you want the raw rows instead of the pages.