Gemini API Free Tier (April 2026): Every Rate Limit, Quota, and Gotcha

Two Ways to Access Gemini for Free

Google offers Gemini through two platforms, each with different free tier mechanics. Understanding the distinction is important because they serve different use cases, and mixing them up leads to unexpected bills or unnecessary limitations.

Google AI Studio (aistudio.google.com)

This is the primary free tier path. You get an API key with zero configuration. No Google Cloud project. No billing account. It works immediately after signing in with a Google account.

AI Studio gives you access to every current Gemini model: 2.5 Flash, 2.5 Pro (experimental), 2.0 Flash, and 1.5 Pro. Each model has different rate limits on the free tier, and we cover those in detail below. The key advantage here is simplicity. You get a key, you make API calls, and you pay nothing until you choose to upgrade.

The tradeoff: Google may use your free-tier inputs and outputs to improve their models. If you are building something with sensitive data, this matters.

Vertex AI ($300 Credit for New Accounts)

Vertex AI is Google Cloud's full ML platform. New GCP accounts get $300 in free credits valid for 90 days. That covers Gemini API calls, but also any other Google Cloud service you use during that window.

After the credits expire, you pay standard rates. There is no ongoing free tier for Vertex AI. The advantage over AI Studio is higher rate limits, no data-sharing clause, enterprise features (grounding, custom fine-tuning, data residency), and SLAs. For most developers starting out, AI Studio is the right choice. Vertex AI makes sense when you need production guarantees or are already in the Google Cloud ecosystem.

Free Tier Rate Limits by Model

These are the limits that actually matter day-to-day. Every model has three constraints: requests per minute (RPM), tokens per minute (TPM), and requests per day (RPD). You will hit the RPM limit long before TPM in most use cases.

A few things stand out. Gemini 2.5 Pro is capped at 50 requests per day on the free tier. That is enough to test and prototype but not enough for any real workload. Gemini 2.0 Flash and the newer Gemini 3.1 Flash-Lite are the most generous on requests per minute. As of May 2026, Google's free tier covers only Flash and Flash-Lite models. The Pro models moved behind billing. So if your project runs on the free tier, you are running on Flash.

Gemini 3 Flash Free Tier Limits

Gemini 3 Flash arrived in early 2026 as Google's recommended free-tier model, and it replaced 2.5 Flash as the default for most new projects. On the free tier it runs 10 requests per minute, 250,000 tokens per minute, and 1,500 requests per day. The 1M-token context window carries over from the 2.5 generation, so you can still feed it long documents without paying.

Two numbers matter here. The 10 RPM ceiling is the one most people hit first. The 1,500 RPD cap is the one that ends a busy day early. At a steady 1 request per minute you run for 25 hours of clock time before the daily cap resets, which means a low-traffic app fits comfortably. A burst-heavy app does not.

Gemini 3.1 Flash-Lite sits below 3 Flash on quality but gives you 15 RPM. If your task is classification, extraction, or routing rather than long-form reasoning, Flash-Lite buys you 50% more throughput on the same free tier. People building automation agents often run Flash-Lite for the cheap, high-volume steps and call 3 Flash only when a step needs the stronger model.

One caveat worth repeating: the limits Google publishes are starting points. AI Studio shows the live cap for your specific project, and it varies by region, account age, and whether billing is attached. Spinning up extra API keys under the same project does not add quota. The limit is per project, not per key.

Running PicoClaw and Other Automation Agents on the Free Tier

The query that brings most people to this page in 2026 is some version of "can I run my agent on Gemini for free, all day, without a bill." The short answer: yes, within the daily cap, and Gemini's free tier is the most agent-friendly of the major providers.

PicoClaw is the tool that made this popular. It is a tiny CLI agent (under 10MB of RAM, sub-second startup) that runs file operations, code execution, scheduling, and web search, and you point it at any LLM by dropping a key into ~/.picoclaw/config.json. People deploy it on a Raspberry Pi or a cheap VPS and let it run 24/7 against the Gemini free tier. The same pattern works for OpenClaw, Aider in unattended mode, and homegrown cron-driven scripts.

What you can actually do for free with an always-on agent:

• Scheduled jobs. A cron task that runs every 10 minutes makes 144 calls a day. That leaves you over 1,300 calls of headroom under the 1,500 RPD cap on Gemini 3 Flash. Summarizing an inbox, watching a folder, or checking a feed all fit easily.
• A personal coding agent. Interactive coding sessions burn requests in bursts. You will hit the 10 RPM wall during heavy editing, not the daily cap. Exponential backoff turns that wall into a 1-2 second pause instead of a failed run.
• A Telegram or Discord bot. PicoClaw's gateway mode connects to chat platforms. For a bot serving you and a few friends, free-tier limits are invisible. For a public bot, you will blow past 1,500 RPD fast.
• Multi-step pipelines. Here is the trap. A single "task" in an agent can fan out into 8 or 10 model calls (plan, search, read, reason, act, verify). Ten tasks an hour at 8 calls each is 80 RPM of demand against a 10 RPM ceiling. You need queuing.

The honest limit for automation: the daily cap is your real budget, not the per-minute one. Picture an agent that averages 4 model calls per task. The 1,500 RPD ceiling on Gemini 3 Flash gives you about 375 agent tasks a day before the free tier cuts you off. That is plenty for a personal assistant, a monitoring job, or a side project. It is not enough for an agent serving real users.

Three tactics keep a free-tier agent alive longer. Route cheap steps (classification, routing, short extraction) to Gemini 3.1 Flash-Lite to spread load across two separate quotas. Cache anything deterministic so repeat questions never hit the API. And queue agent steps so a burst of fan-out calls drains under 10 RPM instead of slamming the limit and triggering 429s. For where the free tier stops making sense, see our breakdown of LLM pricing per million tokens and how Gemini compares to Claude API pricing once you go paid.

What the Free Tier Includes Beyond Text

Gemini is multimodal, and the free tier includes all modalities. This is a significant differentiator from OpenAI and Anthropic, where vision capabilities are often limited or cost extra.

• Image understanding: Send images alongside text prompts. Useful for OCR, chart reading, visual Q&A, and image classification. No separate image API needed.
• Video understanding: Upload video clips up to 2 hours in length (with Gemini 1.5 Pro or newer). The model can answer questions about video content, extract information from frames, and summarize visual sequences.
• Audio processing: Send audio files for transcription, translation, and content understanding. Supports common formats including MP3, WAV, and FLAC.
• Document analysis: Upload PDFs directly. The model processes text and images within the document without needing a separate parsing step.
• Code execution: Gemini 2.0 Flash and 2.5 models can execute Python code in a sandboxed environment and return results. This runs inside Google's infrastructure at no additional cost on the free tier.

The rate limits above apply across all modalities. A multimodal request (text + image) counts as one request toward your RPM and RPD limits. Token counts for images and video are calculated based on resolution and duration, which can consume your TPM faster than text-only requests.

Gemini App vs. Gemini API: Different Products

This is where people get confused. The Gemini app (gemini.google.com) and the Gemini API are separate products with separate free tiers.

Gemini app (free): Unlimited text conversations with Gemini. Includes image generation (via Imagen 3). No API access. Think of this as Google's equivalent to free ChatGPT. It uses Gemini 2.0 Flash by default.

Gemini Advanced ($19.99/month via Google One AI Premium): Access to Gemini 2.5 Pro in the app. 2TB Google storage. Gemini in Gmail, Docs, Sheets, and other Workspace apps. NotebookLM Plus. Still no API access included.

Gemini API (Google AI Studio free tier): What this guide covers. Programmatic access to all Gemini models. Completely separate from the Gemini app subscription. You can use both the free app and the free API simultaneously.

A common mistake: people pay for Gemini Advanced thinking it gives them API access. It does not. The API free tier through Google AI Studio is separate and does not require any subscription.

Free Tier vs. Pay-as-You-Go Pricing

When you do outgrow the free tier, here is what you will pay. These rates apply through Google AI Studio with a billing account enabled.

Gemini 2.0 Flash at $0.075 per million input tokens is the cheapest major LLM API on the market. For comparison, GPT-4o mini costs $0.15/$0.60 (twice the input price) and Claude Haiku 4.5 costs $1/$5 (over 13x the input price). If cost is your primary concern and you need a capable model, Gemini 2.0 Flash is difficult to beat even after you leave the free tier.

The jump from free to paid also increases your RPM by 100-200x. That alone justifies upgrading for any production workload.

Practical Tips for Maximizing the Free Tier

If you want to stay on the free tier as long as possible, here are strategies that work.

1. Use Gemini 2.0 Flash for most tasks. It has the highest rate limits (15 RPM, 1M TPM) and is fast. Reserve 2.5 Flash or 2.5 Pro for tasks that genuinely need better reasoning.
2. Cache results aggressively. If your app asks the same types of questions repeatedly, cache the API responses locally. A simple key-value cache (Redis, SQLite, even a dictionary) can eliminate 30-60% of redundant calls.
3. Batch your requests within rate windows. Instead of making calls as they come in, queue requests and send them in controlled bursts that stay under 15 RPM.
4. Use the Gemini app for manual testing. The web app has no rate limits for conversational use. Use it for prompt development and iteration, then switch to the API for production calls.
5. Implement exponential backoff. When you hit rate limits, the API returns 429 errors. A simple retry with exponential backoff (1s, 2s, 4s, 8s) handles transient limit hits gracefully without losing requests.

How Gemini Free Tier Compares to Other Providers

Here is an honest comparison of what you get for free across the major AI API providers as of April 2026.

Google's position is clear: they want developers building on Gemini, and they are willing to subsidize the onboarding. The combination of no expiration, no credit card, multimodal support, and access to competitive models makes the Gemini free tier the strongest starting point for developers in 2026.

For a detailed breakdown of all providers, see our complete AI API free tier comparison.

Common Mistakes to Avoid

After tracking developer experience with Gemini's free tier, these are the mistakes we see most often.

• Confusing AI Studio with Vertex AI. They have different URLs, different authentication, different rate limits, and different billing. Pick one and stick with it.
• Not handling 429 errors. The free tier will rate-limit you. If your code does not have retry logic, requests will silently fail during traffic spikes.
• Sending unnecessarily large prompts. Long system prompts eat into your TPM budget. Keep prompts concise. Use few-shot examples only when they measurably improve output quality.
• Ignoring the data-sharing clause. If you send customer data, PII, or proprietary information through the free tier, Google can use it for training. For production apps with real user data, upgrade to paid or use Vertex AI.
• Using 2.5 Pro when 2.0 Flash would work. The Pro model has a 50 RPD limit on free tier. Most tasks that developers throw at Pro can be handled by Flash with a better prompt. Test with Flash first.

Related Resources

Gemini Free Tier vs. Building with Context Caching

One feature worth knowing on the paid tier: context caching. If your application sends the same system prompt or large document with every request, caching lets you pay for that context once and reuse it across requests. Cached tokens cost 75% less than regular input tokens on Gemini 2.5 Flash.

On the free tier, context caching is not available. Every request sends the full context. For prototyping, that's fine. Once you're running real workloads where the same instructions repeat across hundreds of calls, caching pays for the upgrade quickly. A 10K-token system prompt repeated 1,000 times costs $0.38 uncached versus $0.09 cached on Gemini 2.0 Flash. At scale, the math tilts hard toward paid.

The other paid-tier feature worth knowing: batch requests. Gemini supports asynchronous batch processing for non-time-sensitive workloads at a 50% discount on token costs. If you're processing thousands of documents overnight, the combination of batch pricing and context caching makes Gemini the cheapest capable LLM API by a significant margin. This does not exist on the free tier. Free tier requests are synchronous only.

For reference: our LLM pricing comparison tracks current costs across all major providers. And for developers choosing between the free tier and the API: the decision usually comes down to whether you hit the RPM limits before you hit your budget limits.

Getting Started: Practical First Steps

If you haven't used the Gemini free tier yet, here's the fastest path to your first API call.

1. Go to aistudio.google.com. Sign in with any Google account. No special access needed.
2. Click "Get API Key." You can create a key attached to an existing Google Cloud project or create a new one. For free tier use, the project setup is automatic.
3. Make your first call. The API follows the same general pattern as OpenAI's SDK. The quickest way to start is with the Python SDK: pip install google-generativeai. The Google AI Studio interface also has a code export button that generates the equivalent API call for whatever you test in the UI.
4. Pick the right model. For most first projects, Gemini 2.0 Flash is the best starting point. It has the highest free tier limits (15 RPM, 1M TPM), fast response times, and strong enough quality for most tasks. Use 2.5 Flash if you need better reasoning.
5. Add exponential backoff immediately. Even in development, you'll hit rate limits when you're testing quickly. Build retry logic before you need it, not after a confusing string of 429 errors.

Related: ChatGPT free tier guide and which AI free tiers are worth building on in 2026.