Most “best free tier” lists rank by headline number. That is the wrong axis for a hobby project, because free tiers fail on shape long before they fail on volume: a limit of 5 requests per minute is generous for a nightly batch job and useless for a chatbot, even when the daily token allowance is enormous.
So this guide sorts by what you are actually building. Every figure is from the provider’s own docs and dated on its offer page.
Match the tier to the workload
| If you’re building… | Start with | Why |
|---|---|---|
| A chat bot (Discord, Telegram, web) | Groq | 30 req/min on the documented gpt-oss example |
| A general side project / daily driver | Gemini free tier | Capable models such as Gemini 3.8 Flash and 2.5 Pro |
| A batch or scheduled job | Cloudflare | 10,000 Neurons/day, resetting at 00:00 UTC |
| Prototyping across many models | OpenRouter | One key, a whole :free catalogue |
| Search, RAG retrieval, reranking | Cohere | Rerank and embed inside 1,000 calls/month |
| Something already deployed on Cloudflare | Workers AI | 10,000 Neurons/day, inference next to your Worker |
Chat bots: Groq, then Gemini
A bot’s problem is requests per minute, not tokens per day. Ten people typing at once is ten requests in a few seconds, and a per-minute cap turns that into visible errors.
Groq’s Free plan publishes 30 requests per minute on
openai/gpt-oss-120b, openai/gpt-oss-20b and qwen/qwen3.8-27b, a documented per-minute allowance, with a daily cap of 1,000 requests per
model. (openai/gpt-oss-safeguard-20b is tighter at 3 a minute, and Groq’s
groq/compound systems were decommissioned on 21 September 2026.) For a personal
bot, 1,000 requests a day is the number to design against.
The Gemini free tier is the other sensible default and the better all-rounder for quality — with two caveats: Google does not publish numeric free-tier limits, so you check yours in AI Studio, and not every Gemini model is on the free tier (Gemini 3.1 Pro Preview, for one, is not).
If you would rather compare the raw numbers than the use cases, see free LLM API rate limits compared.
Batch jobs: Cloudflare Workers AI
Cloudflare Workers AI gives 10,000 Neurons per day, resetting at 00:00 UTC, as a recurring allowance for work that runs on a schedule rather than on demand. Estimate whether your nightly summariser, document pass or evaluation fits: model choice and token counts determine consumption. Two caveats: Neurons are a normalised compute unit, so your real capacity depends on which model you pick, and eight large third-party models (Moonshot’s Kimi, Z.ai’s GLM and DeepSeek V4 variants) require a paid billing method and will not run on the free allowance at all.
Cerebras used to hold this slot on the strength of 1M tokens a day per model. As of September 2026 it requires a verified payment method and its $5 in credits expire after 30 days, so it is a trial rather than a free tier — see the Cerebras offer page for the current terms.
Prototyping: OpenRouter
When you do not yet know which model you want, OpenRouter
is the cheapest way to find out: one key, one API, and a catalogue of :free
model variants at 50 requests a day (20 a minute), rising to 1,000 a day once
you have bought $10 of credits. Failed attempts count toward the daily cap. Treat
the free catalogue as rotating — it is excellent for comparison, less good as a
production dependency. Our OpenRouter guide
covers the mechanics.
RAG and search: Cohere
For the retrieval half of a project, Cohere’s trial keys are underrated: 10 rerank requests a minute inside an overall cap of 1,000 API calls a month. Embedding a personal knowledge base is a few dozen calls, so the monthly cap is rarely the binding constraint — and reranking meaningfully improves small-corpus RAG. Check the current trial Embed limit in your dashboard.
Already on Cloudflare? Use Workers AI
If your side project is a Worker, Workers AI gives 10,000 Neurons a day (resetting 00:00 UTC) with inference running next to your code. Neurons are Cloudflare’s compute unit rather than tokens, so the token-equivalent varies by model.
What not to build a hobby project on
Hugging Face’s monthly credit used to be the smoke-testing option here. It is gone for free accounts: Hugging Face’s docs have listed no free Inference Providers credit since 7 October 2026.
Cloud trial credits (Google Cloud $300, AWS up to $200, Azure $200) are a different instrument entirely: they run on a clock (90 days for Google, a 6-month free plan at AWS, 30 days at Azure) and are wasted on steady low usage. Save them for a burst — a fine-tune, a big evaluation, a launch week.
The setup that actually works
Hold several keys, not one. Registering with five providers costs nothing, and it buys you real resilience: when one tier’s daily cap trips, you fail over instead of going dark. Keep each key in an environment variable, and read how to combine free AI tiers legitimately first — horizontal stacking across providers is fine; multiplying accounts at one provider is not.
For the current dated numbers on every tier here, see the free AI API tiers pillar or the full comparison of what each free tier gives.