Gemini API Free Tier: Models, Quotas and Agent Workloads

Gemini API has free access to selected models. Your usable capacity depends on the model and project limits shown in AI Studio. Check those limits before scheduling an agent: a task can make several API requests, and a quiet daily average can hide a burst that exceeds your quota.

By PE Collective. Updated October 28, 2026. This guide covers the Gemini Developer API; consumer Gemini subscriptions and Google Cloud services have separate terms.

Our data: eight reproducible workload calculations below, prepared October 28, 2026, plus a documentation check across six Google sources. The workloads are hypothetical planning examples. They are not measured throughput or promises about your account. This article was prepared with AI assistance; no model inference tests were performed.

Jump to free models, your project limits, workload calculations, 429 errors or data handling.

Which Gemini models have free API access?

Google's current pricing table lists free standard input and output for Gemini 3.8 Flash and Gemini 3.5 Flash-Lite. Gemini 4 Argon has no free tier. Check the exact model ID and service mode; a model family's name alone does not establish what is included. Source: Google's Gemini API pricing, checked October 28, 2026.

Standard API modelFree input/outputFeature limits to check
gemini-3.8-flashListed as freeFree caching listed; Search and Maps grounding unavailable free
gemini-3.5-flash-liteListed as freeCaching, Search and Maps grounding unavailable free
gemini-4-argonNot availableRequires paid access

These eligibility examples cover selected models. Feature availability differs even between the two Flash entries. A free text response therefore does not establish free access to every tool, modality or processing mode.

Older tutorials need a model check. Google's lifecycle page lists June 1, 2026 as the shutdown date for Gemini 2.0 Flash and Flash-Lite. It limits Gemini 2.5 access to previous active users and directs new projects to 3.5 Flash-Lite or 3.8 Flash. It does not describe 2.5 as universally deprecated. Verify your exact identifier against Google's model lifecycle table before copying an old configuration.

Where do you find your actual free-tier limits?

Use the active limits for your project in AI Studio. Google defines RPM as requests per minute, TPM as input tokens per minute, and RPD as requests per day. Limits apply per project, rather than per API key; daily quotas reset at midnight Pacific. Model, usage tier and account status affect the limits, and stated capacity is not guaranteed. See Google's rate-limit documentation and active-limit link.

Every Gemini API key belongs to a Google Cloud project. AI Studio can create a default project and key for a new user, which makes setup look simpler than it is underneath. Existing users may need to import a project. Google's API-key guide explains the relationship. Record which project your application uses before comparing a dashboard limit with an error.

  1. Record the exact model ID, project and service mode from your application configuration.
  2. Read the active quota for that same project and model. Save the observation date with your capacity plan.
  3. Check whether other applications share that project, then subtract their expected usage from your planning allowance.
  4. Measure requests and input tokens for representative tasks before increasing the schedule or concurrency.

Keep API keys out of logs and screenshots. A useful diagnostic record contains the project identifier, model, timestamp and quota category; it does not need the secret that authenticates the request.

How much quota does an automation agent need?

Count model calls per task, then calculate daily volume and peak-minute demand separately. Include retries and repeated planning steps. The following examples use invented workloads and, where needed, a hypothetical 1,500-request daily allowance. Replace that allowance with your active project limit; 1,500 is not a current Google entitlement claimed by this guide.

Planning questionInputs and formulaResult
One scheduled run every 10 minutes24 hours × 60 minutes ÷ 10144 runs/day
Four calls per scheduled run144 runs × 4 calls576 requests/day
Daily task ceiling before retries1,500 hypothetical requests ÷ 4 calls/task375 tasks/day
Ten tasks per hour, eight calls each10 × 8 ÷ 60 minutes1.33 requests/minute on average
Five four-call tasks launched in one minute5 tasks × 4 calls20 requests in that minute
Input-token demand for that burst20 requests × 2,000 input tokens40,000 input tokens in that minute
One request every minute for a full day24 hours × 60 minutes1,440 requests/day
Worst-case attempts with two retries per call576 initial calls × (1 + 2 retries)1,728 attempted requests/day

The hourly example is 80 requests per hour, or about 1.33 per minute if distributed evenly. If the same work arrives together, the average understates its burst. In the five-task example, all 20 calls fall inside one minute; a hypothetical limit below 20 would require a slower schedule even if the daily total fits.

The token example assumes each request sends 2,000 input tokens. Growing conversation history can change that input on later steps. Track the distribution of request sizes rather than copying the first request's token count across an entire run.

The final row is a maximum attempt count if every initial call uses both retries. It is not an expected failure rate or a statement that every failed attempt is charged identically. Use it to bound how much traffic your own retry policy could generate. Under the hypothetical daily allowance, that retry policy would exceed the planning budget.

For a first deployment, set a task limit and concurrency limit together. Save unfinished tasks durably, avoid duplicate execution when resuming them, and stop scheduling when your chosen allowance is exhausted. A low-frequency automation is easier to fit when it can wait until later; a user-facing action with a deadline needs a separate failure response.

What should you do when Gemini returns 429?

Read the error details and compare the named quota with your active limits before retrying. A short burst and an exhausted daily allowance require different scheduling decisions. Waiting a second does not create a fresh daily allowance; preserve the work and wait for the relevant reset or evaluate an authorized capacity change.

For transient errors, Google's troubleshooting guide recommends exponential backoff, random jitter and a maximum retry count. It also notes that official SDKs can already retry transient failures. Check that behavior before wrapping the client in another retry loop. Do not endlessly retry invalid requests or permission errors.

  1. Log the status, error category, model and timestamp without logging credentials or sensitive prompt text.
  2. Pause new work while checking whether the limit concerns requests, tokens or a different quota.
  3. Apply bounded backoff only where retrying can help. Stop after the configured attempt budget.
  4. Checkpoint the task and expose its pending or failed state to the caller; do not report success before receiving a valid result.

Test your application's failure handling with a local mock response before relying on it. Simulating a 429 lets you verify queue behavior without deliberately exhausting a live quota. Also test an unavailable model and an invalid request so the application does not treat every failure as congestion.

What data-handling conditions apply?

Google's Gemini API terms say unpaid-service inputs and responses may be used to improve products and may receive human review. They instruct users not to submit sensitive, confidential or personal information to unpaid services. The terms include exceptions: EEA, Swiss and UK users receive the paid-service data-use treatment, and active billing changes how covered services are classified.

For paid services, Google says prompts and responses are not used to improve its products, while limited logging still applies. Read the applicable terms for your account and intended data before using either tier. A price label alone is not a complete retention or privacy assessment.

Use synthetic or public test inputs while deciding whether the service fits your workflow. Keep any production-data approval separate from the question of whether the request count fits. This guide does not determine your organization's obligations or approve a particular dataset.

When does evaluating paid access make sense?

Evaluate it when an eligible free model cannot meet your required schedule, feature set or data-handling requirements. First collect a representative workload: task volume, calls per task, input size, output size, burst pattern and tolerated delay. This produces a useful cost estimate and exposes inefficient retries before you increase spending.

Confirm paid model access, rates and quotas for the intended project before enabling billing. Compare token costs, separately priced tools and any caching storage charges where applicable. Higher limits still require capacity planning. A free-tier bottleneck alone does not establish which paid model is suitable.

For the next step, enter your actual quota alongside the eight calculations above and replace the sample workload with your measurements. If you need a cost reference, continue to our LLM token-pricing comparison. Recheck Google's linked sources when changing a model or billing configuration.

Frequently Asked Questions

Is there one Gemini free-tier limit for everyone?

No. Check the active model and project limits in AI Studio. This guide deliberately does not present a universal requests-per-minute or requests-per-day allowance.

Does creating another API key increase my quota?

Keys in the same project share its limits. Check which project the application uses before comparing its errors with a quota dashboard.

Can PicoClaw or another scheduled agent fit within a free quota?

Check the agent configuration for its exact model, project and calls per task. In the example above, one run every ten minutes with four calls per run needs 576 requests per day before retries. Compatibility and quota access still need verification in your setup.

Does a low daily average prevent rate-limit errors?

No. Five tasks that each make four calls within one minute produce a 20-request burst. Check peak-minute requests and input tokens alongside the daily total.

Should I keep retrying after a daily quota is exhausted?

Save the pending task and wait for the relevant reset or evaluate an authorized capacity change. Short repeated retries cannot replenish a daily allowance.