Free AI API Keys (2026): Genuinely Free LLM APIs Compared

The best free AI APIs in 2026 are Google's Gemini API free tier, Groq, Mistral's free mode, OpenRouter's :free models and Cloudflare Workers AI. All of them give you a real API key without a credit card and are rate-limited instead of credit-capped.

OpenAI's API has no free tier. Anthropic gives new API users only a small starter credit. For speech and embeddings, Deepgram ($200), AssemblyAI ($50), Voyage AI and Jina AI offer large free allowances. Every number below was checked against the provider's own pricing, rate-limit or docs page on October 17, 2026.

What "free AI API key" actually means

Providers use the word "free" for three quite different things, and mixing them up is the most common reason a project breaks a month in:

• Rate-limited free tiers never run out of money, but cap how many requests or tokens you can send per minute or per day. Gemini, Groq, Mistral, OpenRouter's :free models, Cloudflare Workers AI and Cohere trial keys work this way. They reset (daily or monthly) and you can keep using them indefinitely as long as the provider keeps the tier.
• One-time signup credits are a dollar balance you spend down. Deepgram's $200, AssemblyAI's $50, Cerebras's $5 and Anthropic's small starter credit fall in this group. Once the balance is gone, you pay per use. Some expire (Cerebras: 30 days), and Deepgram is unclear (its pricing page says the credit lasts until used, but its docs say promotional credits expire one year from signup).
• Token allowances are a fixed number of free tokens per account, mostly for embeddings and reranking. Voyage AI gives 200 million free tokens on its current models; Jina AI gives an auto-generated key with free tokens.

A fourth category, larger cloud trials and startup credits (Google Cloud's $300, AWS Activate, Microsoft for Startups), usually needs a card or an application. They are covered at the end of this guide and in how to get free AI credits.

Free LLM APIs compared

Google Gemini API

Gemini is the most generous free tier from a frontier lab. Google's pricing page lists "Free of charge" input and output for Gemini 3.8 Flash, 3.6 Flash, 3.5 Flash-Lite, 3.1 Flash-Lite, 3 Flash Preview, 2.5 Pro, 2.5 Flash, 2.5 Flash-Lite, several Live, TTS and Transcribe models, Gemini Embedding 2 and Gemma 4. Not free: Gemini 4 Argon, Gemini 3.1 Pro Preview, most image models (Nano Banana, Nano Banana 2, Nano Banana Pro), Veo 3.1 and Lyria. One image model, Gemini 3.1 Flash Lite Image, is listed as free.

Google no longer publishes a numeric free-tier rate-limit table. The rate-limits page says limits vary by model and are shown for your project in AI Studio; daily request quotas reset at midnight Pacific time. Google Search grounding is mostly a paid-tier feature on Gemini 3 and newer models. Full details: Gemini API free tier, Gemini API pricing and our Gemini API key guide.

Groq

Groq runs open models on its own LPU hardware and is known for very fast output. Its console's free-plan limits list gpt-oss-120b, gpt-oss-20b and Qwen3.8-27B at 30 requests per minute, 1,000 requests per day, 8,000 tokens per minute and 200,000 tokens per day, plus Whisper speech-to-text at 20 RPM and 2,000 requests per day. Limits apply to your organization, not to each key, and Groq says the authoritative numbers are on your account's limits page. When you exceed a limit, the API returns HTTP 429 with a retry-after header.

Mistral

Mistral's docs say API access in Studio "is enabled by default with no credit card required. Usage and rate limits apply." Free mode includes a monthly usage allowance, but Mistral shows the actual numbers only on the Limits page inside your console, so figures quoted on blogs can't be confirmed. If you see a 402 Payment Required error, Mistral's docs say it means there is no payment method on the account, which matters once you go beyond what free mode includes.

OpenRouter :free models

OpenRouter is a router in front of many providers, with one OpenAI-compatible endpoint. Models whose ID ends in :free cost nothing. Its limits page sets 20 requests per minute for free models, and 50 requests per day if you have bought fewer than 10 credits, or 1,000 per day after buying at least 10 (the docs say the higher cap actually starts at 9 to absorb rounding and fees). On October 17, 2026 its public models list had 16 free models, including google/gemma-4-31b-it:free, nvidia/nemotron-3-super-120b-a12b:free, nvidia/nemotron-3-ultra-550b-a55b:free and cohere/north-mini-code:free. A negative account balance can block even free models with a 402 error.

Cloudflare Workers AI

Workers AI includes 10,000 Neurons per day on both the Workers Free and Workers Paid plans. On Workers Free, requests stop once the daily allocation is used; on Workers Paid, usage above it costs $0.011 per 1,000 Neurons. Limits reset daily at 00:00 UTC. A Neuron is Cloudflare's unit of GPU compute, so how many requests 10,000 Neurons buys depends on the model and prompt size.

Cohere, NVIDIA, Anthropic and Cerebras

• Cohere trial keys are capped at 1,000 API calls a month, with 20 requests per minute for chat models (Command A, Command R and others), 2,000 inputs per minute for Embed and 10 requests per minute for Rerank. Cohere's pricing page says trial keys can't be used for production or commercial purposes.
• NVIDIA says members of its free Developer Program get free access to NIM API endpoints on build.nvidia.com "for prototyping", for the duration of their membership, and only for prototyping, research, development and testing. NVIDIA doesn't publish rate limits or a credit amount.
• Anthropic says new users "receive a small amount of free credits to test the API". It doesn't publish the amount. After that, you pay per token; the cheapest current model, Claude Haiku 5.5, costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. See Claude API pricing and how to get an Anthropic API key.
• Cerebras gives new accounts $5 in credits after you add a verified payment method. The credits expire 30 days after they're granted, and Cerebras says it has no tier that renews for free. The trial runs gpt-oss-120b and qwen-3.8-27b at 5 requests per minute.

Free APIs for speech, voice and embeddings

Two details matter here. Voyage's free tokens don't apply to its Batch API, and its pricing page lists slightly different model names in its text and its table, so confirm in your dashboard which models draw from the free allowance. Jina's Reader page says every new API key comes with 10M free tokens, and you can buy more once they're used.

Big providers with no free API tier

• OpenAI: the official API pricing page lists no free credits or trial. A few small items are free, such as the omni-moderation-latest model and the first 1 GB of file search storage. The cheapest text model, gpt-5-nano, costs $0.05 per million input tokens and $0.40 per million output tokens. See OpenAI API pricing.
• Anthropic: new users get a small free credit to test the API, then pay per token. See Claude API pricing.
• xAI (Grok API): the official docs mention prepaid credits and promo codes but no signup credit. See xAI Grok API credits.
• Together AI: its billing docs say it doesn't currently offer free trials, and the first purchase must be at least $15 ($5 after that). See Together AI signup credits.
• Hugging Face Inference Providers: free users get no monthly credits and must buy credits to use it. PRO users get $2/month in general compute credits. See Hugging Face inference credits.
• GitHub Models: retired on July 30, 2026. GitHub points model users to Microsoft Foundry instead. See GitHub Models.
• DeepSeek: no published free signup credit, but very low pay-as-you-go prices. See DeepSeek API pricing.

Which free API should you use?

How to get a free AI API key in 5 minutes (Gemini example)

1. Sign in to Google AI Studio (aistudio.google.com) with a Google account.
2. Open the API Keys page and click Create API key. AI Studio sets up a Google Cloud project for you. If you already use Google Cloud, you may first need to import your projects from the Projects page.
3. Copy the key and store it in an environment variable named GEMINI_API_KEY (or GOOGLE_API_KEY; if both are set, GOOGLE_API_KEY wins).
4. Call a free-tier model and check your limits on the AI Studio rate-limit page.

export GEMINI_API_KEY="your-key-here"

curl "https://generativelanguage.googleapis.com/v1beta/models/gemini-3.8-flash:generateContent" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"contents":[{"parts":[{"text":"Hello"}]}]}'

The flow is similar elsewhere: Groq and OpenRouter create keys from their consoles' API keys pages, Mistral creates them in Studio (the full key is shown only once, so save it straight away), Cloudflare uses an account API token with Workers AI permission, and Jina generates a key automatically when you first visit its site.

Use one client for all of them

Most free providers offer an OpenAI-compatible endpoint, so you can use the official OpenAI SDK and just change the base URL, key and model name. That makes it easy to switch providers when a free tier changes. Official base URLs:

from openai import OpenAI
import os

client = OpenAI(
    base_url="https://api.groq.com/openai/v1",
    api_key=os.environ["GROQ_API_KEY"],
)

reply = client.chat.completions.create(
    model="openai/gpt-oss-120b",
    messages=[{"role": "user", "content": "Hello"}],
)
print(reply.choices[0].message.content)

To fall back between providers, catch HTTP 429 (rate limit) and 402 (payment required) errors and retry the same prompt on the next provider. Groq and OpenRouter return rate-limit headers you can read to wait the right amount of time instead of retrying blindly.

Rate limits explained

• RPM / RPD: requests per minute and per day. A chat app with one user rarely hits RPM limits, but a script processing a spreadsheet hits them within seconds.
• TPM / TPD: tokens per minute and per day. This is often the real ceiling. Groq's 8,000 TPM on gpt-oss-120b means one long prompt with a big document can use a whole minute's budget.
• Reset times differ: Gemini's daily quotas reset at midnight Pacific time, Cloudflare's at 00:00 UTC, and Cohere's trial cap is monthly.
• Limits are usually per project or organization, not per key. Creating more keys in the same account doesn't raise them, and creating extra accounts to get around limits is generally against provider terms.

Tips for staying on free tiers

• Check data terms. Google's pricing page says free-tier content is used to improve its products, while paid-tier content is not. OpenRouter's free models may also log prompts depending on the provider. Don't send private customer data through free tiers.
• Combine providers. Use an OpenAI-compatible client and fall back from Groq to OpenRouter to Gemini when you hit a rate limit.
• Cache and batch. Store answers to repeated prompts and send fewer, larger requests where a provider's limits are per request rather than per token.
• Read the license. Cohere trial keys can't be used for production or commercial work, NVIDIA's free access is for prototyping, and the ElevenLabs Free plan doesn't include a commercial license.
• Watch for card-gated credits. Cerebras's $5 and Google Cloud's $300 trial require a payment method.
• Keep keys secret. Google's own advice: never commit keys to Git, never hardcode them in web or mobile apps, and route client-side calls through a backend.
• Never use "free API keys" posted publicly. Keys leaked on GitHub or forums are usually stolen. Using them can break the provider's terms and get them revoked mid-project.

Free-tier limits and model lineups change often. Each provider in this guide has a deal page with the numbers we last verified. Check the provider's own console before you build anything that depends on them.

Troubleshooting common errors

Need more than a free tier?

Startup programs give much larger credits: see Google for Startups Cloud, Microsoft for Startups, AWS Activate and Claude for Startups. New Google Cloud accounts also get a $300 free trial for most Google Cloud services (not the Gemini API or partner models); it lasts 90 days and isn't available if you've ever paid for Google Cloud, Maps Platform or Firebase. Students should also check the GitHub Student Developer Pack.