TL;DR (July 2026): There is a real tier of AI APIs that stays free - no expiring trial, no credit card, no clock. A viral X post (368 likes) pointing at a curated GitHub repo of permanently-free, OpenAI-SDK-compatible endpoints put the list back in front of everyone. The two anchors are OpenRouter, which serves DeepSeek R1, Llama 3.3 70B, and Gemma 3 at zero cost, and Google's Gemini API, which keeps a free experimentation tier you can call without entering a card. The catch is the part the "free" label hides: every one of these tiers has a hard ceiling - requests per minute, requests per day, or tokens per minute - and the moment your traffic outgrows it you are silently moved onto a paid meter. "Free" is not "unmetered." It is a smaller meter you are not watching yet.
For a prototype, a side project, or an internal tool that runs a few hundred calls a day, you do not need to pay for inference at all in 2026. The permanently-free tier is genuinely useful, genuinely capable, and - because most of it speaks the OpenAI SDK format - a two-line base-URL swap away from your existing code. That is the good news, and it is real. The trap is treating "free" as if it means "infinite." It does not. It means "free up to a ceiling you did not read, after which the behavior changes." Knowing exactly where that ceiling is, and watching your usage climb toward it, is the difference between a free tier that works and a 3 a.m. outage.
The free APIs that actually stay free
The important distinction is between a trial (free credits that expire, usually behind a card) and a free tier (a standing allowance with no card and no clock). The repo that went viral is a list of the second kind. Here is the honest shape of what stays free, and the kind of ceiling each one enforces.
| Provider | Free models | Card required? | Cost | Ceiling you hit first |
|---|---|---|---|---|
| OpenRouter (free tier) | DeepSeek R1, Llama 3.3 70B, Gemma 3 | No | $0 | Requests per minute and per day |
| Google Gemini API | Free experimentation tier | No | $0 | Requests per minute, tokens per minute, requests per day |
Free-tier models and limits move constantly and are set per provider; treat the live numbers in the provider docs and the community repo as the source of truth, not this snapshot. OpenRouter's own free-LLM-APIs comparison and Google's Gemini API pricing page publish the current ceilings.
Two properties make this tier more than a toy. First, the models are not leftovers: DeepSeek R1 is a capable reasoning model and Llama 3.3 70B is a solid general workhorse, so "free" here does not mean "bad." Second, most of these endpoints are OpenAI-SDK-compatible, which is why the repo spread so fast - you point your existing client at a new base URL and API key and your code does not otherwise change. That compatibility is the whole ballgame for adoption, and it is also why so many teams end up leaning on a free tier far longer than they planned.
The ceiling you'll actually hit first
Free tiers do not fail by sending you a bill. They fail by refusing a request, and they refuse along whichever axis you saturate first. There are three, and they behave differently:
- Requests per minute (RPM). The one bursty workloads hit first. A batch job, a page that fans out several calls on load, or a retry loop can blow through a per-minute request cap in a single spike while your daily total is nowhere near the limit.
- Requests per day (RPD). The one steady, low-rate workloads hit. You never spike, but a cron job that runs all day quietly accretes until it trips the daily cap - usually in the afternoon, for something that started fine in the morning.
- Tokens per minute (TPM). The one long-context workloads hit. You are well under the request caps, but a few large prompts saturate the token budget. This is the sneakiest ceiling because two callers making the "same" number of requests can have wildly different token footprints.
The practical problem is that you rarely know which ceiling you are closest to until you cross it, because free tiers do not show you a running total. You find out from a 429, in production, after it has already broken something. And once you graduate to a paid key to escape the free ceiling, the failure mode inverts: instead of a refused request, you get a charge. The first paid step on Gemini - Tier 1 for Gemini 2.5 Flash-Lite - lifts you to 4,000,000 input tokens per minute, which is a lot of headroom and a lot of potential spend if a loop misbehaves. The free ceiling protected you from a runaway bill by refusing the call. The paid tier removes that protection. That is exactly the moment you need a meter you can see.
What changed in 2026
The free tier got better and the paid floor got cheaper, which together make the "graduate from free" moment softer and easier to sleep through.
- The open-weight models on the free tier are now genuinely capable. DeepSeek R1, Llama 3.3 70B, and Gemma 3 at $0 would have been a respectable paid tier a year ago. The gap between "free model" and "the model you would have paid for" narrowed sharply.
- Consumer subscriptions got cheaper too. Google cut its AI Plus plan from $7.99 to $4.99 per month on June 8, 2026 and doubled its storage allowance to 400GB, while the standard assistant tier across ChatGPT, Claude, Google AI Pro, and Perplexity holds at about $20/month. Cheaper consumer plans blur the line between "I am on the free API" and "I am paying," which is precisely the line your finance team cares about.
- The paid API floor collapsed. When you do outgrow free, the cheapest capable API models are startlingly cheap - which is good for your bill and bad for your discipline, because a near-free per-token price makes it feel like you are still on the free tier when you are not.
| Model (paid API) | Input $/1M | Output $/1M |
|---|---|---|
| DeepSeek V4 Flash (cheapest) | $0.14 | $0.28 |
| GLM-5.2 | $1.40 | $4.40 |
| Gemini 3.1 Pro | $2.00 | $12.00 |
| Claude Opus 4.8 | $5.00 | $25.00 |
| GPT-5.5 | $5.00 | $30.00 |
Approximate July 2026 API list prices per 1M tokens; see morphllm.com/llm-api and tldl.io. The spread from cheapest to most expensive is roughly 600x, and which model you route to swings a monthly bill by about 10x - which is why the meter matters more than the sticker price.
"Free" is a meter you're not watching yet
Here is the reframe that makes free tiers safe to depend on. A free tier is not the absence of a meter - it is a meter with a hard cap and no dashboard. The provider is counting your requests and tokens precisely; they just do not show you the running total, and they enforce the limit by refusing you rather than billing you. So the discipline is identical to the paid case: know your ceiling, and watch your usage climb toward it, so the graduation to a paid key is a decision you make on a Tuesday afternoon - not an incident you discover from a 429 at midnight.
That is the whole reason to instrument a free tier before you think you need to. If you meter your own calls - counting requests per minute, requests per day, and tokens per minute against the published ceiling - you get three things the provider will not give you: a warning before you hit the wall, a clean signal of exactly when you outgrew free, and a running cost the day you flip to a paid key. This is the same cost-tooling discipline that governs a $50,000 API bill, applied at the $0 end where it is cheap to build the habit. When you do move up, the tier and spend-cap structure is the next thing to understand, and a hard spend cap with a kill switch is what replaces the free ceiling's accidental protection against a runaway loop. If the reason you are on the free tier is cost, the cheapest paid models - the DeepSeek, GLM, and Flash-Lite class - are where you land next, and metering is how you confirm the cheap model is actually cheap for your workload rather than just cheap on the price page.
The honest take
The permanently-free AI API tier is one of the best deals in software right now, and you should use it. But "free" is a claim about price, not about limits, and it quietly trains a bad habit: running on a meter you cannot see. The teams that get burned are not the ones who used a free tier - they are the ones who scaled a product on one, hit an invisible ceiling in production, panic-upgraded to a paid key with no cap, and discovered the cost only when the invoice arrived. Treat the free tier the way you will have to treat the paid one: know your ceiling, watch your usage, and decide when to graduate. The free tier is a gift. The 429, or the surprise bill right after it, is the tuition you pay for not metering it. For the mechanics of what that 429 actually costs you when you retry it blindly, see whether a rejected request still counts against your limit - because it does.