Vertex AI Free Tier vs Gemini API Free Tier: Which to Choose

Ask whether Gemini has a free tier and you will get two answers. Both are correct, and they describe two different products. Google runs a free path on Vertex AI, its enterprise cloud, and a separate free path on the Gemini Developer API at ai.google.dev. Same model family, same vendor, different money, different quotas, and — most importantly — different data-use terms.

This guide breaks down both free tiers, does the arithmetic on when each one runs out, and gives you a decision rule you can apply in about thirty seconds. If you are choosing between them for a project that touches anything you would not publish, the choice is not really about price.

What is the difference between Vertex AI and the Gemini API?

Dimension	Vertex AI (cloud.google.com)	Gemini Developer API (ai.google.dev)
What you get free	Up to $300 Welcome credit, valid 90 days, plus a recurring grounding allowance	Per-model free token tiers (rate-limited, not credit-limited)
Shape of the free tier	A balance that depletes	A meter that resets
Data terms	Standard Google Cloud terms; free-trial content is not used to improve products	Free-tier content is used to improve Google's products
Identity and governance	Full Google Cloud IAM: projects, org policy, service accounts, audit logs	API key, no IAM surface
Grounding with Google Search	5,000 grounding queries per month at no charge, aggregated across Gemini 3 models, then billed per 1,000 queries	Not available on the free tier; a paid-tier feature
Intended user	Teams evaluating or shipping toward production	Individuals prototyping

The table is the article in miniature. The two rows that decide most real projects are shape and data terms. A balance and a meter fail in completely different ways — one runs out permanently, the other refills at midnight Pacific — and the data clause decides whether you are even allowed to point the thing at your payloads.

The Vertex AI free trial: $300 of credit on a 90-day clock

Google describes its free trial as a Free Trial billing account preloaded with $300 in Welcome credit, valid for 90 days for new customers (Google Cloud free trial↗). It is not a token grant. It is money, and it is spendable across Google Cloud services, including Vertex AI inference, embeddings, batch jobs and tuning.

Three consequences follow from "balance, not meter":

1. The credit is fungible. You can spend it on any eligible Vertex service, not just Gemini. Embeddings, batch prediction, model tuning and grounding all draw from the same $300.
2. The 90-day clock is the real ceiling. A workload that burns $200 slowly over 90 days has consumed its free tier just as surely as one that burns $200 in week three. If you have an evaluation to run, front-load it.
3. It expires rather than resets. There is no tomorrow-morning refill. When the credit is spent or the window closes, the account transitions to normal pay-as-you-go billing unless you shut it down.

The other half of Vertex's free surface is recurring rather than one-off: Google's Agent Platform pricing page lists 5,000 grounding queries per month at no charge, aggregated across all Gemini 3 models, with additional queries billed at $14 per 1,000 (Vertex / Agent Platform pricing↗). That is a standing monthly allowance for retrieval-augmented work — the kind of line item that quietly costs real money on other platforms. Two important footnotes come with it: the allowance is aggregated across models rather than per model, and Google warns that a single request may trigger more than one search query, and you are billed for each individual query performed.

The Gemini API free tier: token quotas and a data price

The Developer API free tier is a different animal. There is no dollar credit. Instead, Google applies rate limits to a set of eligible models, measured across three dimensions: requests per minute (RPM), input tokens per minute (TPM) and requests per day (RPD). Google's own documentation is clear about the mechanics (Gemini API rate limits↗):

• Limits are applied per project, not per API key. Minting a second key in the same project does not create a second quota.
• RPD quotas reset at midnight Pacific time, which is a genuinely useful property: unlike a credit balance, an exhausted free tier is restored on a schedule you can plan around.
• The Free usage tier qualifies with an active project or a free trial, and paid tiers unlock higher ceilings. Tier 1 arrives when you link a billing account (with a billing tier cap of $250), Tier 2 after paying $100 plus three days, and Tier 3 after paying $1,000 plus 30 days.
• Spend-based rate limits were added on top of request limits: Tier 1 is capped at $10 of spend per rolling 10 minutes, Tier 2 at $50, and Tier 3 at $200. The Free tier has no spend limit because it cannot spend.
• Specified rate limits are not guaranteed. Google states this explicitly, which means the exact free-tier ceilings for a given model should be read from the live AI Studio rate-limit view rather than from any article — including this one. We deliberately do not publish per-model free-tier numbers here, because they vary by model and by account and change without notice.

And then the clause that matters most. On the Gemini API pricing page↗, the Free tier lists "Content used to improve our products", while the Paid tier lists "Content not used to improve our products". That is the entire trade: the Developer API free tier is free in cash and priced in data. It is a rational choice for synthetic prompts, public documents and throwaway experiments. It is a disqualifier for customer data, internal code, health information or anything under an NDA.

Grounding is the other free-tier gap. Google's pricing page lists Grounding with Google Search as "Not available" on the Gemini API free tier. On the paid tier you get 5,000 free search requests per month shared across Gemini 3.x models, then $14 per 1,000 requests. So a free-tier prototype cannot simply bolt on live web search at $0 — that capability is either a Vertex line item or a paid Gemini API line item.

The dividing line is the data-use clause, not the price

Strip away the tables and the two free tiers differ on one axis: who is allowed to learn from what you send. Same Gemini models, same output quality — different legal and product posture.

That asymmetry has a practical consequence people miss. The free tier you want (no card, no expiry, no setup) is not the free tier you are allowed to use for sensitive data. Teams routinely prototype on the Developer API free tier with real fixtures, decide the model is good, and only then discover that those fixtures were governed by product-improvement terms. The fix is cheap if you catch it in week one and expensive if you catch it at launch.

A simple rule that survives contact with a real project:

• Public or synthetic data → Gemini API free tier. Zero cost, no clock, current-generation models, daily quota that resets.
• Real or regulated data → Vertex AI. The $300 credit funds a serious evaluation under Google Cloud terms, and you inherit IAM, org policy and audit logging from the first request.

How we verified this

We re-checked the numbers in this article on 2026-09-17 directly against Google's own surfaces: the Google Cloud free trial page↗ for the $300 / 90-day terms, the Gemini API rate-limits documentation↗ for limit dimensions, per-project scoping, the midnight Pacific reset and the usage-tier ladder, the Gemini API pricing page↗ for the data-use clause and grounding availability, and the Agent Platform pricing page↗ for the monthly grounding allowance. We deliberately omit per-model free-tier RPM/TPM/RPD numbers because those change and because Google publishes them per account in AI Studio; treating a dated blog number as authoritative is how people mis-size their first production week.

Worked example: does a 90-day prototype fit inside $300?

Suppose your prototype reads 3 million input tokens and writes 500,000 output tokens per day, six days a week, for three months. Using the paid-tier rates published on the Gemini API pricing page for a small Flash-Lite-class model ($0.30 per million input tokens, $2.50 per million output tokens):

• Input: 3M tokens × $0.30 / 1M = $0.90 per day
• Output: 0.5M tokens × $2.50 / 1M = $1.25 per day
• Total: $2.15 per day → about $193.50 over 90 days

That fits inside $300 with roughly $106 of headroom, which is enough to absorb a heavier evaluation week, a batch re-run, or a few thousand grounding queries at $14 per 1,000.

Now run the same shape against a larger Flash model. Google currently lists $0.75 per million input tokens and $3.75 per million output tokens for its newest Flash model, with those promotional rates running through December 31, 2026:

• Input: 3M × $0.75 / 1M = $2.25 per day
• Output: 0.5M × $3.75 / 1M = $1.875 per day
• Total: $4.125 per day → $300 runs out at about day 73, not day 90

Same workload, one model tier higher, and the trial ends nearly three weeks early. This is why "we have $300" is not a plan; the clock and the model choice are both levers.

The grounding allowance adds a second piece of arithmetic. If your RAG evaluation issues 8,000 grounding queries in a month, the first 5,000 are covered by the monthly allowance and the remaining 3,000 bill at $14 per 1,000 — $42 for the month. If you choose the Gemini API route instead, grounding is not available on the free tier at all, so those 8,000 queries land on the paid tier from query one.

Which free tier should you actually use?

Choose the Gemini API free tier when: your prompts contain nothing sensitive, your volume fits inside per-minute and per-day quotas, you want to start in five minutes with one API key, and you want a free tier that resets instead of expiring.

Choose Vertex AI when: your data cannot be used for product improvement, you need IAM and audit trails, you want a $300 budget to run a real evaluation rather than a toy test, or you need the recurring monthly grounding allowance for RAG work.

Use both, in this order: prototype the plumbing on the Gemini API free tier with synthetic fixtures — that costs nothing and validates your code path. Then move the evaluation that touches real data onto Vertex AI with the trial credit, and delete the sensitive fixtures from the free-tier project. This is the pattern that gets you fast iteration without putting regulated data through a product-improvement clause.

Do not assume "free" means "same terms." The models are identical. The terms are not. The terms are the part that survives contact with a launch.

Limitations and caveats worth knowing before you commit

• The trial credit is for new customers. Existing billing accounts, or accounts that have previously had one, are not eligible for the Welcome credit.
• Card verification is required for the free trial because a billing account must be created; the credit shields you from charges during the window, but the account exists.
• Grounding queries bill per individual search query, not per user prompt. A single prompt can expand into several queries, so your grounding line item can be larger than your prompt count suggests.
• Free-tier rate limits are documented as not guaranteed. Build retry-with-backoff behaviour from day one; a free tier that 429s during a demo is a design problem, not a vendor problem.
• Vertex pricing is not identical to Gemini API pricing. The Flash-Lite and Flash rates used above come from the Gemini API pricing page. Vertex AI has its own price sheet; check the Vertex pricing page for the model you intend to run rather than reusing Developer API rates.
• Your free-tier model list is not the paid model list. Google restricts free-tier access to certain models, and restricts experimental and preview models more tightly than stable ones.

FAQ

Does upgrading to a paid Gemini API tier end the data-use clause? Yes. Google's pricing page distinguishes the Free tier ("Content used to improve our products") from the Paid tier ("Content not used to improve our products"), and context caching plus the Batch API become available once billing is linked.

Is there a way to get free Gemini on Vertex after the 90 days? The credit expires, but the 5,000 grounding queries per month allowance is a recurring monthly line item rather than part of the trial balance. Everything else moves to standard pay-as-you-go rates.

Can I use the Gemini API free tier and Vertex AI with the same project? They are billed and governed through different surfaces — a Gemini API key versus a Google Cloud project and billing account — so treat them as two integrations even when they run the same model. Do not assume a quota consumed in one affects the other.

What happens if I exceed the free tier instead of a credit balance? You receive a rate-limit error for the current window and the quota resets on schedule; only paid tiers incur spend-based limits, which are evaluated on a rolling 10-minute window and return an error of the resource-exhausted type.

Is the $300 credit spendable on models from other vendors on Vertex? Vertex hosts partner models alongside Gemini, and the credit is Google Cloud spend, so the practical answer is yes for eligible Vertex services. Verify per-model eligibility on the Vertex pricing page before assuming a specific partner model is covered.

Where this sits in the directory

A directory that lists "Google Gemini — free" as one row hides the distinction that decides which free tier you are actually on: credit versus meter, enterprise terms versus product-improvement terms, one-off versus recurring grounding. That is why Vertex AI and the Gemini Developer API are tracked as separate entries with their own free-tier shape, data-use clause and clock. When you compare Google against the rest of the market, keep the same discipline: a free tier is four things at once — an amount, a shape, a clock and a data clause.