Free LLM APIs for Indie Developers (2026)

Short answer: Start with Google's Gemini API free tier for general prototyping, Cloudflare Workers AI if you want a free allocation you can run a small live feature on, and Groq, SambaNova or OpenRouter's :free models when you want open-weight models fast. Do not plan a launch around GitHub Models (retired), Mistral's Free mode (no API access) or Hugging Face free credits (none for free accounts). Pay as soon as rate-limit errors would reach real users or your prompts contain customer data.

Every limit and quote below comes from the provider's own docs or pricing page, checked on 15 Nov 2026. Free tiers change often, so treat this as a snapshot and check the linked page before you build on one.

TL;DR#

• Gemini API: free input and output tokens on many Flash, Flash-Lite and embedding models, with per-project limits shown in AI Studio. Free-tier data may be used to improve Google's products (Gemini pricing, terms).
• Cloudflare Workers AI: 10,000 neurons per day free; $0.011 per 1,000 neurons above that on Workers Paid (Workers AI pricing).
• OpenRouter :free models: 20 requests per minute, 50 per day, or 1,000 per day after buying at least 10 credits (OpenRouter limits).
• SambaNova Free Tier: 20 requests per minute, 20 per day and 200,000 tokens per day on listed models, no payment method needed (SambaNova rate limits).
• Gone or not free: GitHub Models retired on 30 July 2026; Mistral Free mode "doesn't include API access"; Hugging Face free users get no monthly credits.

Disclosure: ZeroRaised is a launch site for bootstrapped products. We have no commercial relationship with any provider in this post and earn nothing if you sign up. Facts come from the linked pages as checked on 15 Nov 2026; analysis marked "Our read" is ours.

How this post fits with our other AI guides#

This is a list of free tiers and their small print. If you are choosing a paid provider on quality, latency and price, read how to choose an LLM API. If you want one key for many providers, our OpenRouter alternatives compares gateways. For a full look at training and retention policies, see do AI APIs train on your data; here we only flag what each free tier says. And if "free" means running a model on your own machine, see how to run an LLM locally.

Free LLM API tiers compared#

"Not stated" means we could not find the point on the provider's own page today, not that the answer is no.

Each free tier in detail#

Google Gemini API: the widest free tier, with data strings attached#

Google's pricing page lists a Free Tier with "Free input & output tokens" on a long list of models, including gemini-3.8-flash, gemini-3.6-flash, gemini-3.5-flash-lite, gemini-3.1-flash-lite and the gemini-embedding-2 and gemini-embedding-001 embedding models. Its top models are excluded: the page says "Gemini 4 Carbon is not available on the Free tier" and the same for Gemini 4 Argon. Grounding with Google Search is listed as not available on the free tier for most models (Gemini pricing).

Limits are not one public table. The rate-limits page says limits "can be viewed in Google AI Studio", are measured in RPM, input TPM and RPD, are "applied per project, not per API key", and that "quotas reset at midnight Pacific time" (Gemini rate limits). Moving to Tier 1 means you "Set up and link an active billing account".

Watch two things:

• Model churn. The pricing page already lists gemini-2.5-pro routing to gemini-3.8-flash and gemini-2.5-flash routing to gemini-3.5-flash-lite on 1 December 2026. Hard-coded model IDs will change under you.
• Data use. The terms say that on Unpaid Services Google "uses the content you submit to the Services and any generated responses to provide, improve, and develop Google products and services and machine learning technologies", add that "human reviewers may read, annotate, and process your API input and output" and say plainly: "Do not submit sensitive, confidential, or personal information to the Unpaid Services." They also say that if you make apps available to users in the European Economic Area, Switzerland or the UK, you "may use only Paid Services" for them (Gemini API terms).

Our read: brilliant for prototypes, evals on synthetic data and personal tools. For a public app with UK or EU users, plan on the paid tier from day one.

Groq: a Free plan for fast open-weight models#

GroqCloud's rate-limits page has separate Free Plan and Developer Plan views and tells you to check "the limits page in your account settings" for your exact numbers (Groq rate limits). The public table shows Developer plan base limits, so we have not quoted Free plan figures. Models listed include openai/gpt-oss-120b, openai/gpt-oss-20b, qwen/qwen3.8-27b and Whisper speech models. Upgrading needs "a valid payment method (credit card, US bank account, or SEPA debit account)" and "takes effect immediately" (Groq billing FAQs).

Our read: worth a free key when you want open-weight models behind a fast hosted API, such as for chat UIs and agents that make many short calls. Check your Free plan numbers in the console before you rely on them.

Cloudflare Workers AI: the free tier built for running things#

Both Workers Free and Workers Paid include "10,000 Neurons per day". On Workers Free there is no overage, you must upgrade; on Workers Paid you "will be charged at $0.011 / 1,000 Neurons for any usage above the free allocation", and limits reset at 00:00 UTC each day (Workers AI pricing). Neurons map to tokens per model. For example, @cf/meta/llama-3.1-8b-instruct is listed at $0.282 per million input tokens and $0.827 per million output tokens.

The default text generation limit is "300 requests per minute, unless the model requires the Workers Paid plan", and text embeddings get "3000 requests per minute" (Workers AI limits).

Our read: because the free allocation lives inside a normal hosting plan with a clear overage price, this is the free tier we would most happily ship a small production feature on, such as summaries, tagging or embeddings for search.

OpenRouter :free models: many models, low ceilings#

Free variants (model IDs ending in :free) are limited to 20 requests per minute. Per day, the limit is 50 requests if you have bought less than 10 credits and 1,000 if you have bought at least 10 (OpenRouter limits; the FAQ gives the same 50 and 1,000 figures). A negative credit balance can block free models too. On data, OpenRouter's privacy docs say "There are separate settings for paid and free models" and that if you opt out of training, it "will not route to providers that train" (OpenRouter logging), which can shrink the free models available to you.

Our read: buying 10 credits once is the cheapest upgrade on this list, because it lifts the free cap twentyfold.

SambaNova Cloud: generous per minute, tight per day#

SambaNova's Free Tier is "Applied when there is no payment method linked with your account". Listed models such as DeepSeek-V3.1, Meta-Llama-3.3-70B-Instruct and gpt-oss-120b get 20 requests per minute, 20 requests per day and 200,000 tokens per day (SambaNova rate limits). Twenty calls a day is enough for evals and testing, not for users.

Cerebras: a $5 trial credit, not a free plan#

Cerebras now offers credit rather than an ongoing free tier: "Add a valid payment method to receive a one-time $5 promotional credit", which "expires 30 days after activation". It says "You will not be charged or automatically enrolled in paid usage", and access pauses when the credit runs out (Cerebras pricing). While on the credit, gpt-oss-120b is capped at 5 requests per minute and 1M tokens per day (Cerebras rate limits). The pricing page also carries a footnote: "For development, evaluation, and experimentation; not intended for production use."

Cohere: trial keys for evaluation#

Cohere describes trial keys as "evaluation keys (free but limited in usage)". They are "limited to 1,000 API calls a month", with chat at 20 requests per minute, text embeddings at 2,000 inputs per minute and rerank at 10 requests per minute (Cohere rate limits). To "serve Cohere in production", Cohere's docs walk you through upgrading to a Production key (Cohere going live). Our read: worth a trial key to test Cohere's embed and rerank models in a search or RAG prototype.

NVIDIA build.nvidia.com: prototyping only#

Members of the free NVIDIA Developer Program get "free access to NIM API endpoints for prototyping". That route "is for prototyping, research, development and testing purposes only", and "Using NIM in production requires an NVIDIA AI Enterprise license" (NVIDIA NIM FAQ).

Closed or paid-only routes#

• GitHub Models: "As of July 30, 2026, GitHub Models has been fully retired." GitHub points to Azure AI Foundry and GitHub Copilot instead (GitHub docs).
• Mistral: "Free mode includes a monthly usage budget for trying Vibe and Studio. It doesn't include API access." The API needs billing turned on (Mistral usage and limits).
• Hugging Face Inference Providers: free users get "None" in monthly credits and can use the service by buying credits; PRO users get $2.00 per month (Hugging Face pricing).

If you have AI credits from a startup programme, those are a different thing from a free tier; our explainer on what AI credits are and our guide to Anthropic credits for startups cover them.

Which free tier for which job#

Our read, based on the limits above:

How to stack free tiers with a fallback#

1. Use one OpenAI-style client where you can. Hugging Face's pricing docs, for example, show the OpenAI Python client pointed at https://router.huggingface.co/v1. Check each provider's docs for an OpenAI-compatible endpoint; where one exists, switching is mostly a base URL, key and model ID.
2. Keep model IDs in config, not code. Gemini alone is rerouting two model IDs on 1 December 2026.
3. Order providers by priority. For example: Workers AI first, Groq second, OpenRouter :free third.
4. Fall back on HTTP 429 and 402. Catch rate-limit and payment errors, then retry the next provider with a short backoff.
5. Log which provider answered. Outputs differ between models; you want to know which one produced a bad answer.
6. Never route sensitive prompts to a tier that trains. Tag requests that contain user data and send them only to paid endpoints.

Our read: stacking is fine for a prototype, but three providers means three sets of quirks. Once you have users, one paid provider with one free fallback is simpler.

When to start paying: a worked example#

This example is ours, using Cloudflare's published per-token prices for @cf/meta/llama-3.1-8b-instruct.

Say your side project summarises 2,000 support tickets a day, each with 1,000 input tokens and 300 output tokens:

• Input: 2M tokens × $0.282 per million = about $0.56 per day
• Output: 0.6M tokens × $0.827 per million = about $0.50 per day
• Total about $1.06 per day, minus the 10,000 free neurons (worth $0.11 at $0.011 per 1,000) = about $0.95 per day, or roughly $28 a month, plus your Workers Paid plan subscription.

At 200 tickets a day, the same workload costs about $0.11 per day, which is almost exactly the free allocation. That is the pattern for most indie apps: free tiers cover you until real usage arrives, and then the bill is small. Our read: start paying when any of these is true:

• A user would see a rate-limit error.
• Your prompts contain customer or personal data.
• You serve UK, EU or Swiss users through Gemini.
• You spend more than an hour a month managing quotas.

To keep the paid bill small afterwards, see how to reduce LLM costs.

Risks of building on a free tier#

• Limits change without much notice. Groq and Gemini publish account-specific limits in their consoles rather than in fixed public tables.
• Free tiers disappear. GitHub Models retired, Mistral's Free mode dropped API access and Hugging Face free accounts get no credits.
• Model IDs expire. Gemini's reroutes on 1 December 2026 are an example.
• Terms may forbid production. NVIDIA's free access is explicitly for prototyping, and Cerebras's footnote says "not intended for production use".
• Data terms. Gemini's free tier may be used to improve Google's products and read by human reviewers.

Frequently asked questions#

What is the best free LLM API for an indie developer?#

For most prototypes, Google's Gemini API free tier, because it covers many current Flash, Flash-Lite and embedding models at no cost. If you want open-weight models, Groq's Free plan, SambaNova's Free Tier and Cloudflare Workers AI's 10,000 free neurons per day are good options. OpenRouter's :free models are useful for trying many models behind one key, but 50 requests per day is too low for production unless you have bought at least 10 credits.

Does Google train on Gemini API free tier data?#

Google's Gemini API terms say that for Unpaid Services, Google uses submitted content and responses to provide, improve and develop its products, that human reviewers may read, annotate and process API input and output, and that you should not submit sensitive, confidential or personal information. On Paid Services Google says it does not use prompts or responses to improve its products.

Can I use a free LLM API in production?#

Sometimes, but read each provider's terms. NVIDIA's free developer access is for prototyping, research, development and testing only. A footnote on Cerebras's pricing page describes an offering as not intended for production use. Gemini's terms say apps for users in the EEA, Switzerland or the UK may use only Paid Services. Cloudflare Workers AI's free allocation sits inside normal Workers plans, so it is the easiest free tier to run a small live feature on.

Is GitHub Models still free?#

No. GitHub's documentation says that as of 30 July 2026 GitHub Models has been fully retired, and points to Azure AI Foundry and GitHub Copilot as alternatives. If your prototype used the GitHub Models endpoint, it needs a new provider.

When should I start paying for an LLM API?#

Our read: when a real user would notice a rate-limit error, when you need prompts kept out of training, when you serve users in the UK or EU on Gemini, or when the bill for a paid tier would be a few dollars a month. At small volumes paid inference is often cheaper than the engineering time spent juggling free quotas.

Ship it#

Built something on a free tier? Work through our product launch checklist for bootstrapped makers, then launch it on ZeroRaised, the launch directory for products built without VC money. For more guides like this, join the ZeroRaised newsletter.

Sources#

All last checked 15 Nov 2026.

• Gemini API pricing
• Gemini API rate limits
• Gemini API additional terms of service
• Groq rate limits
• Groq billing FAQs
• Cloudflare Workers AI pricing
• Cloudflare Workers AI limits
• OpenRouter API limits
• OpenRouter FAQ
• OpenRouter privacy and logging
• SambaNova rate limits policy
• Cerebras pricing
• Cerebras rate limits
• Cohere rate limits
• Cohere going live
• NVIDIA NIM API FAQ
• GitHub Models documentation
• Mistral usage and limits
• Hugging Face Inference Providers pricing