How Does Google Gemini API Pricing Work?

Most people asking what the Gemini API costs are really asking two different questions at once. The first is "can I build this for free?" The second is "what happens to my bill when this actually gets used?"

Google's pricing page answers both, but it answers them in a table with dozens of rows, several model families, and footnotes about prices changing on specific dates. It is accurate and it is not especially readable.

This guide is the structural version. It covers what you get without paying, what changes the moment you attach a billing account, which two features cut a production bill roughly in half, and the parts of the pricing model that surprise teams after launch rather than before it. Specific per-token rates move, so treat any number here as an illustration and check the live page before you build a forecast on it.

What does the Gemini API actually cost?

There are three tiers on Google's pricing page: Free, Paid and Enterprise.

On the free tier, Google's own summary is that you get "limited access to certain models", "free input & output tokens" and "Google AI Studio access" — with the significant trade-off, stated on the same page, that your content is "used to improve our products".

On the paid tier you pay per million tokens, separately for input and output, at a rate that depends on which model you call. As an illustration of the shape of it rather than a quote to budget against, Google currently lists standard Gemini 3.8 Flash at $0.75 per million input tokens and $3.75 per million output tokens — and flags that this is promotional pricing through 31 December 2026, doubling on 1 January 2027.

That asymmetry is the single most important thing to internalise. Output costs several times more than input. A prompt-heavy, answer-light workload is cheap. A workload that generates long responses is not.

Not sure which of your pages AI assistants are actually citing? Get a free SEO audit and find out where you stand in AI search.

How is Gemini API pricing structured?

Everything is priced per million tokens

Tokens are chunks of text, roughly a few characters each. Input tokens are everything you send — system instructions, conversation history, documents you paste in. Output tokens are everything the model generates.

Both are metered, both are priced separately, and in most cases output is the expensive half. This is why the same model can cost wildly different amounts for two applications with identical request counts.

Model family is the biggest cost lever you control

Google's lineup is organised into families rather than a single model. Broadly: Flash models for general-purpose work at moderate cost, Flash-Lite models for the cheapest high-volume tasks, and Pro models for the hardest reasoning.

Alongside those sit specialised models with their own pricing: Live models for real-time streaming conversation, TTS models for speech synthesis, Image models for generation, and Transcribe models for audio.

Picking the smallest model that passes your evaluation is worth more than any other optimisation. The gap between a Lite model and a Pro model on the same workload is not ten percent, it is often an order of magnitude.

Reasoning output is still output

Models that do internal reasoning before answering generate tokens while doing it, and those tokens are billed as output. A model that "thinks" more produces a bigger bill for the same visible answer.

This catches teams who benchmark on quality, pick the model that reasons hardest, and then find their cost per request is several times what the headline rate implied. Measure cost per completed task, not cost per million tokens.

Some things are priced per request, not per token

Grounding with Google Search is the clearest example. Google lists 5,000 free search requests per month, shared across the Gemini 3.x models, and then charges per thousand requests after that. If you ground every answer in live search, that line can become a bigger share of your bill than the tokens.

Your content has to be findable before an assistant can cite it. Run a free SEO audit to see what is holding your pages back.

How do you move from the free tier to the paid tier?

1. Open Google AI Studio and go to the billing setup for your project.
2. Create a project and set up billing, or import an existing Google Cloud project.
3. Confirm the project is on the Paid Tier.
4. Use the API key tied to that project for paid calls.
5. Watch usage under Dashboard > Usage in AI Studio.

Google describes the upgrade path directly: "You can create a project and set up billing, or import an existing project, to upgrade to the Paid Tier in Google AI Studio." Billing runs through Cloud Billing accounts, which you can set up inside AI Studio rather than in the Cloud console.

You do not have to give up the free tier

This is the detail worth knowing before you start. Free and paid live at the project level, not the account level, and Google is explicit about the consequence: "You can switch between Paid Tier projects and Free Tier projects as needed by using the respective API keys linked to each type."

So the sensible setup is two keys. A free-tier key for prototyping, experiments and anything you do not mind being used to improve Google's products, and a paid-tier key for anything touching real customer data. Google also notes that AI Studio itself "remains free of charge unless users link a paid API key for access to paid features".

Billing accounts set the ceiling for every project under them

One line on the billing page has caught out more than one team: "All projects linked to a Cloud Billing account inherit the billing account's usage tier and associated rate limits and account caps."

That means rate limits are not a per-project property you can raise by spinning up a new project. Spread work across five projects on one billing account and they share the same tier.

What actually lowers a Gemini API bill?

Batch API

If your work is not interactive — overnight classification, bulk enrichment, evaluating a dataset — the Batch API is the easiest win available. Google lists it as a 50% cost reduction against standard pricing, and it is a paid-tier feature.

The trade is latency. You submit a job and collect results later instead of waiting on a response. For anything a user is not sitting and watching, that trade is almost free.

Context caching

Context caching, also paid-tier only, lets you pay once to keep a large shared prefix resident instead of re-sending it with every request.

This is transformative for a specific shape of application: one where every call carries the same long document, policy manual, codebase or system prompt. If 90% of your input tokens are identical across requests, caching attacks 90% of your input cost. If every request is genuinely different, caching does nothing and adds complexity.

Prompt discipline

The least glamorous lever and often the largest. Trimming a bloated system prompt, capping conversation history, and asking for structured output instead of prose all cut tokens on every single call for the rest of the application's life.

Setting a maximum output length is the highest-leverage version of this, because output is the expensive side of the meter.

Cheaper tokens will not fix invisible content. Start with a free SEO audit and fix the foundations first.

When does the free tier stop being enough?

The free tier is limited by rate, not just by model access. Usage is measured on three dimensions simultaneously — requests per minute, input tokens per minute and requests per day — and Google's documentation is blunt about how they interact: "Your usage is evaluated against each limit, and exceeding any of them will trigger a rate limit error."

Paid usage then sits in tiers that you do not apply for. They escalate automatically with spend: attaching billing moves a project from Free to Tier 1, which Google says "will typically take effect instantly". Tier 2 requires accumulated spend of $100 or more plus three days from your first payment, and Tier 3 requires $1,000 or more plus thirty days.

The practical consequence is that a brand-new paid project is not a high-limit project. If you are planning a launch, the limits you will have on launch day are the limits of a Tier 1 account, not the ones listed at the top of the table. Build and spend a little in advance rather than discovering this during a traffic spike.

What the pricing page will not tell you

Prices on this page are dated, deliberately

Several models currently carry promotional rates with an explicit end date and a scheduled increase afterwards. Any cost model you build needs a note of which prices are promotional and when they change, or your forecast quietly breaks at the turn of the year.

Free-tier data handling is a real decision, not a footnote

"Content used to improve our products" is a design constraint, not a disclaimer to skim. It rules the free tier out for customer data, anything under a confidentiality obligation, and most regulated workloads — regardless of how generous the token allowance is.

Token counts are not word counts

Estimating with words will understate your bill, often substantially for code, structured data or non-English text. Measure with real requests before committing to a number, and measure again after any prompt change.

Where you read your costs matters

AI Studio shows usage; the Cloud Billing console shows money. Google points Gemini API costs to "the Cost management pages in the Cloud Billing console." Teams that only ever look at the AI Studio dashboard tend to be the ones surprised by an invoice.

Gemini API pricing vs the alternatives

Versus Google's consumer subscriptions. A Gemini app subscription is a flat monthly fee for a person using a chat interface. The API is metered and built for software. They are not substitutes, and the subscription does not include API usage.

Versus other model APIs. Every major provider now prices per million tokens with input cheaper than output, so the headline numbers are genuinely comparable — but only if you compare like-for-like model classes and include reasoning tokens. A cheap-looking model that thinks at length can cost more per finished task than an expensive one that does not.

Versus running an open model yourself. Self-hosting trades a variable per-token bill for fixed infrastructure cost and engineering time. It wins at sustained high volume with predictable load, and loses badly at spiky or low volume.

Versus the free tier indefinitely. Tempting, and the right answer for learning and prototypes. It stops being viable the moment you have real users, because rate limits and data handling, not price, are what force the upgrade.

Final Thoughts

Gemini API pricing is simpler than its table suggests, once you hold three facts in your head. Output tokens cost several times more than input tokens. Model family matters more than any other choice you make. And the free tier is bounded by rate limits and data handling, not by generosity.

If you are estimating a budget, do it in this order: pick the smallest model that passes your evals, measure real token counts on real prompts, apply batch pricing to anything non-interactive, and apply caching only if your requests genuinely share a long prefix. That sequence will get you closer than any spreadsheet built off headline rates.

And check the live page before you commit. This is the most time-sensitive category of technical content there is — models get renamed, rates get revised, and promotional pricing expires on dates printed in the footnotes.

If you are comparing providers rather than just budgeting for one, our breakdown of how much Claude AI costs walks through the same structure for Anthropic's models, including where its API pricing diverges from its subscription plans.

Build on the best model you like — you still need to be findable. Claim your free SEO audit and see how your content performs in AI search.