Gemini Pricing: What Each Plan and Model Costs

Gemini pricing is three separate bills that share one name.

An app subscription. A per-token meter. A per-seat contract for companies. Google calls all of it Gemini, so a number copied from one bill is wrong for the other two.

Here is the short version, checked against Google's pages on October 16, 2026. The Gemini app costs nothing at the base level, or $4.99 a month for Google AI Plus, $19.99 for Google AI Pro, and $99.99 or $199.99 for Google AI Ultra. The Gemini API bills per token, from $0.10 per million input tokens on the cheapest model to $2 on 3.1 Pro. Gemini 4 Argon, the newest flagship, costs $2 in and $10 out per million tokens until November 30, then $4 and $20.

Those numbers are correct, and they still won't predict what you pay. The app plans measure your usage in multiples of a limit Google doesn't publish. The API charges for tokens you never see. And several prices below change before January 1.

Pick the bill that matches how you use Gemini, and ignore the other two.

Think of the app plan as a buffet and the API as an electric meter. The buffet costs the same whether you eat or not, and it has a plate limit nobody posts on the wall. The meter has no ceiling and runs every second the lights are on.

Most overspending on Gemini comes from standing in the wrong line.

Gemini pricing for the app runs from $0 to $199.99 a month

Google sells the app side as Google AI plans. The plan decides which models you can open, how much you can use them, and how much else comes in the box. These are the US prices on Google's plans page as of October 16, 2026:

Notice what the table lacks. There is no message count.

The footnote on that page says limits depend on the complexity of your prompt, the features you use and the length of your chat, and that your limit refreshes every 5 hours until you reach a weekly cap. The 2x, 4x and 5x figures are multiples of a number you can't see.

When you hit the cap, Google offers two exits. According to its I/O 2026 announcement, it shifts you to a smaller, faster model, or Pro and Ultra subscribers can buy pay-as-you-go AI credits for Antigravity and Flow now, with the Gemini app coming soon. If you lean on video or coding agents, $19.99 is a floor, not a ceiling.

That same announcement explains why Ultra has two price points. Google cut the 20x tier from $250 to $200 and added the $99.99 entry tier. Google's own AI plans page does the bundle math for you: the Ultra bundle at $99.99 against $146 a month when the pieces are bought separately, a saving of $46.

Two more details are worth knowing. Google's plans page currently advertises Pro at no charge for one year to college students. And Gemini 4 Argon is a paid-plan feature in the app: Google's subscriptions page tells anyone looking for it to pick AI Pro or AI Ultra. What Argon is and where it wins has its own post.

Google AI Pro is the plan to price everything else against

At $19.99, Pro is where Google puts the bundle. You get 5 TB of storage, Gemini inside Gmail and Docs, Chrome auto browse, a YouTube Premium Lite plan and access to Argon. Gemini and ChatGPT now cost almost the same, so the choice between them turns on what else comes in the box, not on the number.

Here is how the five tiers sort by what you do all day:

You might think $4.99 for Plus is the obvious steal. For the storage, it is. For work, 2x of a base limit you can't see is a small number, and the first time you run into it is in the middle of a task.

If Gemini is the only AI you pay for, buy Pro and stop reading plan pages.

That changes when you also pay for another assistant. Then the question isn't which Google plan fits you. It is whether a second or third subscription earns its seat, and that comes up near the end.

Gemini API pricing is per token, and the model sets the bill

A token is a chunk of text a few characters long. Google prices input and output separately for every model, and output costs several times more than input. These rates come from Google's Gemini API pricing page, last updated October 1, 2026, plus Google Cloud's pricing page for Argon. All of them are per million tokens on the paid tier:

The spread on input alone is 20 to 1, from $0.10 to $2. The newest Flash models are the odd ones: 3.8, 3.7 and 3.6 Flash share a rate that is half of 3.5 Flash on input and less than half on output, because Google is running an introductory price through December 31.

Newest is cheapest, for now.

Google also runs a developer tier with no per-token charge on several models, with one condition. Google's pricing page marks content on that tier as used to improve its products, and content on the paid tier as not used. Gemini 3.1 Pro Preview isn't on the no-charge tier at all. If your prompts contain client data, that line decides your tier before price does.

Per-million prices are hard to feel, so run one job through them. Say you send 10,000 support-reply requests a month, each with 2,000 tokens in and 600 tokens back. This is arithmetic on Google's published rates, not a measurement, and it leaves out thinking tokens, which come up shortly:

The same job costs $37.50 or $200 depending on a menu choice you make once. Pick the cheapest model that passes your own quality check, not the one with the best launch post.

One caution on the Argon rows. Google's footnote says the lower promotional rate is delivered as 50% credits back on net spend. Read that as a credit that arrives after the spend, not as a lower sticker price.

Gemini 4 Argon costs $2 and $10 until November 30, then it doubles

Argon is the model people ask about, so it gets its own paragraph of arithmetic. On Google Cloud's page, input is $2 per million tokens and output is $10 through November 30, 2026. Starting December 1 it is $4 and $20. Cached input moves from $0.10 to $0.20 on the same date.

If you are building on Argon, budget the December price. A prototype that costs $100 a month in November costs $200 in December for the same traffic, and nobody changes a line of code to cause it.

Caching softens it. Say every request carries a 50,000-token brief. Fresh, that is $0.10 of input in November and $0.20 in December. Cached, it is $0.005 and $0.01, a 95% cut either way, and the dollars you save double along with the price. Google's cache storage table doesn't list Argon, so confirm the hourly storage fee before you count on it.

The Argon date is one of three that matter, and the others involve Flash. Put them on a calendar before you commit a budget:

Anything you model at today's Flash price will be wrong by January. A year of spend planned in October at $0.75 and $3.75 is a plan built on two months of discount.

In the app, none of this touches you. Pro and Ultra include Argon with no per-token meter, which is the one place the buffet beats the meter. You pay for the limit, not the tokens.

Thinking tokens are on your invoice and nowhere in your prompt

Look at the output row on Google's pricing page for any Gemini 3 model. It reads "Output price (including thinking tokens)." Reasoning models write notes to themselves before they answer. You never see those notes, and you pay for them at the output rate.

Take the support-reply job again on Gemini 3.1 Pro Preview. Visible output was 600 tokens. Now suppose the model thinks for 3,000 tokens first. The 3,000 is an assumption to show the shape, not a measurement of any model.

Output becomes 3,600 tokens at $12 per million, which is $0.0432. Add $0.004 for input and the request costs $0.0472. Across 10,000 requests that is $472, against $112 when thinking is ignored.

Same prompt, same answer on screen, a bill four times larger.

The fix is to measure before you budget. Google's billing notes point to the thoughts field in the response metadata as the place to audit how many thinking tokens a request used. Use it like this:

1. Pull 20 prompts from your real work, not demo prompts.
2. Run each one on the model you plan to use.
3. Read the thinking-token count for every response.
4. Average the count, then multiply it by the model's output rate.
5. Add that to your visible-output cost before you price the job.

If the number is ugly, you have two moves. Pick a lighter model for the easy prompts, or reserve the Pro model for the prompts that need it. Models that let you lower the thinking setting cut output tokens directly.

Caching pays only if you reuse the document three times an hour

Context caching lets you store a large block of input and read it back at a fraction of the input price. For Gemini 3.1 Pro Preview, a cached read is $0.20 per million tokens against $2.00 for fresh input. That looks like a 90% discount, and it is, until you meet the storage meter.

Storage on that model is $4.50 per million tokens per hour. Work through a 100,000-token document:

• Fresh input costs $0.20 per request (0.1 million tokens at $2.00).
• A cached read costs $0.02 per request, a saving of $0.18.
• Storing it costs $0.45 per hour (0.1 million tokens at $4.50).

Divide $0.45 by $0.18 and you get 2.5. You need at least three requests an hour against that cache to come out ahead. This ignores what filling the cache costs, so treat three as a minimum.

Leave that cache up for a day with nobody asking and you have spent $10.80 on nothing.

Flash is kinder. On Gemini 3.8 Flash through December 31, the same document costs $0.075 fresh and $0.0075 cached, a saving of $0.0675. Storage is $0.50 per million tokens per hour, so $0.05 an hour for 100,000 tokens. One request an hour covers it. The Pro storage fee is nine times the Flash fee, and that ratio decides which model deserves a cache.

Caching fits an assistant that answers questions about the same manual all afternoon. It is a bad fit for a chatbot that sits idle overnight.

The 200K cliff reprices the whole prompt

Pro models charge more once a single prompt crosses 200,000 tokens. The part people miss is how. Google's pricing notes say that if a query's input context is longer than 200K tokens, all tokens, input and output, are charged at long-context rates. Not just the tokens above the line.

Run it on Gemini 3.1 Pro Preview. A 190,000-token prompt costs $0.38 in input at $2 per million. A 250,000-token prompt costs $1.00 at $4 per million. You added 60,000 tokens, about 32% more text, and input went up 163%. Output also moves from $12 to $18 per million.

It is a per-prompt rule, not a monthly average. A dashboard showing a comfortable mean can hide a few oversized requests billing at the higher rate. If you pack documents into context, count tokens before you send, and set a soft limit near 190,000.

The Flash models have no such step. Google Cloud's page lists Gemini 3.8 Flash at the same $0.75 input rate whether the prompt is under or over 200K tokens. If your job is long documents and a Flash model passes your quality check, the cliff never applies.

Batch and Flex halve the rate, Priority adds 80 percent

Google sells the same model at three speeds. Batch and Flex are half the standard rate, and Priority is 80 percent more. Using the rates on Google's pricing page:

If a job runs overnight and nobody is waiting on it, summaries, tagging, a weekly report, the half-price lane is an easy saving. If a person is waiting, you are in Standard or Priority. Opting into Priority for a whole production app without noticing the multiplier is how a $500 month becomes $900.

Search grounding is its own meter. Google includes 5,000 search requests a month at no charge, shared across all Gemini 3.x models, then charges $14 per 1,000. The catch is in Google's footnote: one request from your user may trigger more than one query to Google Search, and you are charged for each individual query.

Count queries, not prompts. If your app runs 20,000 grounded search queries in a month, 15,000 are billable, which is $210. If each user prompt fires three queries, that is only about 6,700 prompts before you pass the line.

Images, video and audio are priced by the piece

Generation models don't bill by the word. Images bill per image and video bills per second, using the rates on Google's pricing page:

A 10-second clip is $4.00 on Veo 3.1 Standard, $1.00 on Fast and $0.50 on Lite. That sounds fine until you remember nobody keeps the first take.

Say you generate 20 ten-second takes to find the three worth keeping. That is 200 seconds of video: $80 on Standard, $20 on Fast, $10 on Lite.

You pay for the rejects. The practical move is to find the shot on the cheap lane, then spend the Standard rate on the one you keep.

Resolution is the second multiplier. Nano Banana 2 prices an image at $0.045 for 0.5K, $0.067 for 1K, $0.101 for 2K and $0.151 for 4K, so a 4K image costs 2.25 times a 1K one. A hundred 4K images is $15.10 against $6.70 at 1K.

Video works the same way, but through tokens. Gemini Omni Flash bills video output at $17.50 per million tokens, and Google Cloud's page says a second of video uses 5,792 tokens at 720p, 8,688 at 1080p and 17,376 at 4K. That is about $0.10, $0.15 and $0.30 a second. Ask for 4K when 720p would have shipped and you pay triple.

Work out which of the three bills is yours

The answer depends on who is paying and what they are building. Here is how each reader maps to a bill:

The Enterprise prices come from Google's plans page. Business is built for small teams with no IT setup. Standard and Plus add higher usage limits, unlimited seats, and controls such as VPC-SC and customer-managed encryption keys.

A team of 12 on Business is $252 a month. The same 12 on Google AI Pro would be $239.88, so the $12.12 difference is what Google charges for company-data connectors and its promise that your data isn't used to train its models. Whether that is cheap depends on whether your lawyer asks.

For a developer, the table's third row is the hard one, because the number isn't a price. It is a formula. Start from the job table above, add your measured thinking tokens, apply the lane you can live with, and only then compare models.

If Gemini is one of your assistants, price the whole stack

Gemini rarely arrives alone. You might use one model to write, another to code, a third for research, and a fourth because a teammate swears by it. Each gets its own subscription, its own limits and its own memory of you, which is none.

Add up one subscription per assistant and the total lands near $100 a month, with every one of them starting each conversation blind to what the others were told. What you want is every model in one tab, with one thread that knows what the others said.

That is what Magai is. Gemini's Pro and Flash models sit in the model picker beside ChatGPT, Claude, Grok, DeepSeek and dozens more. Type /Gemini in a message, or pick it from the list, and you can switch models mid-conversation without losing the thread. The model changes. The thread's memory doesn't reset.

The plans are priced against that. Standard is $20 a month with every standard model and tool and $20 of monthly usage. Pro is $40 with every model including the premium ones and $40 of usage. Ultra is $200 with $200 of usage. Magai states the allowance as an amount of monthly usage, not a multiple of a baseline you can't see. Every plan has a 30 day, 100% money-back guarantee.

Image and video generation sit in the same workspace. Magai includes Nano Banana for images and Veo and Gemini Omni Flash for video, and a chat model can hand a task to an image or video model inside the same thread. If you are paying per second of Veo anyway, you can test the prompt where the rest of your work already lives.

Here is the fair limit on that pitch. Magai doesn't replace Google's bundle. If the 5 TB of storage, Gmail features and YouTube Premium Lite are why you'd buy Pro, buy Pro. And if you are shipping a product on the API, Google's meter is still your bill. Magai is for the people doing the work, not for the endpoint.

For everyone else, the math is the bundle math: one $20 plan covering several assistants instead of five at $20 each. How the models differ is in Claude vs Gemini, and Claude's plans are priced out in Claude pricing. When you are ready to test Gemini against the others on your own prompt, open Magai and put the same question to two models in one thread.

Buffet or meter. Pick one, then read the number that goes with it.