Gemini API Pricing: Models, Free Tier, Batch, Caching

The Gemini API bills per million tokens. The flagship Gemini 4 Argon costs $2 in and $10 out until 30 November 2026, then $4 and $20. Gemini 3.8 Flash is $0.75 and $3.75 until the end of the year, and Gemini 3.5 Flash-Lite is $0.30 and $2.50. A Free tier exists for most Flash models. Prices checked 2026-10-31.

Prices checked 2026-10-31

Everything below is from Google's Gemini Developer API pricing page, fetched on the check date. Prices are USD per million tokens (MTok), paid tier, standard processing. The page shows two prices for the newest models: a lower one that ends on a stated date, and the list price after.

Google also lists Gemini 3.1 Flash-Lite, 2.5 Flash-Lite, text-to-speech models, a live translation model and two transcription models. We cover the text and reasoning line here. For other vendors, see Claude API pricing, OpenAI API pricing and the cross-vendor LLM API prices table. For the models themselves, see Gemini models.

The date clauses matter

Two of Google's current prices carry a built-in rise.

• Gemini 4 Argon costs $2 in and $10 out through 30 November 2026. From 1 December 2026 the page lists $4 and $20. If you are budgeting for next year, use the December numbers, which are double.
• Gemini 3.8 Flash and 3.6 Flash cost $0.75 and $3.75 through 31 December 2026, and $1.50 and $7.50 from 1 January 2027.

Anything you launch this quarter on Argon pays the lower price for at most a month. Model your costs at the post-promotion rate, then treat the first month as a discount.

Standard, Batch, Flex and Priority

Google sells the same model at several service levels. Using Argon as the example, all with the same date clauses:

The pricing page also lists a Contributor tier for Argon, a private preview on an invite-only basis at $1 in and $5 out, where content is used to improve Google's products. We would not plan around it.

Free tier and paid tier

The pricing page describes two plans for the Developer API.

The last row is the one to read twice. On the free tier, Google says your content is used to improve its products. On paid, it is not. Anything private should go through a paid project.

Gemini 4 Argon and Gemini 3.1 Pro Preview are marked as not available on the free tier. The Flash and Flash-Lite models are free of charge there, so a small prototype can run at zero cost. Check the live rate limits in AI Studio. We did not find a table of free-tier rate limits on the pricing page.

What counts as a token

Google states plainly that "every token the model processes is counted and billed, including thinking tokens and internal tool system instructions and function declarations when using grounding." Thinking tokens are priced as output. A reasoning model can therefore produce an output bill several times the length of the visible answer. If you set a thinking budget on Gemini 2.5 Flash, you cap that cost.

Audio on 2.5 Flash is priced separately ($1.00 input versus $0.30 for text, image and video). Gemini 3.5 Flash-Lite lists one input price of $0.30 across text, image, video and audio.

Context caching

Gemini caches in two ways, and the pricing page shows both.

• Implicit caching is automatic. Argon supports implicit caching only, and explicit caching is not supported on it. The cached input price is $0.10, rising to $0.20 on 1 December.
• Explicit caching with storage is available on other models. Storage is charged per million tokens per hour: $0.50 for Gemini 3.8 Flash and 3.6 Flash until the end of 2026 and $1.00 after, $1.00 for Flash-Lite and 2.5 Flash, and $4.50 for the Pro models.

A cache read costs a tenth of standard input on Flash and Flash-Lite, and a twentieth on Argon,, so a long, repeated prefix gets cheap fast. The storage fee means explicit caching only pays off if you re-read the cache often enough.

Long prompts

Pro models price by prompt length. Gemini 3.1 Pro Preview is $2.00 in and $12.00 out for prompts up to 200K tokens, and $4.00 and $18.00 above. Gemini 2.5 Pro is $1.25 and $10.00, then $2.50 and $15.00. Argon, 3.8 Flash and 3.5 Flash-Lite list a single price. Argon's page describes long-decode generation up to 1M tokens, so its flat price covers very long outputs.

Grounding and tools

Search grounding is the main add-on charge. For Gemini 3 and newer models, Google gives 5,000 free search requests per month (shared across those models), then $14 per 1,000 requests, with a daily cap of 1,500 requests per day per project on the paid tier. Grounding with Google Maps works the same way: 5,000 free requests a month, then $14 per 1,000 queries. On older models the unit is a grounded prompt, $35 per 1,000 for Search on Gemini 2.5 Pro and Flash, after a free daily allowance.

One customer request can trigger several Google searches, and each search is charged. Watch this on agent loops.

A worked comparison

These are our arithmetic from the table, not a measurement. Take 50,000 input tokens and 2,000 output tokens per request at post-promotion prices.

At the current promotional price Argon would be $0.12 for the same request, level with Sonnet 5.5. After December it is double. That is the arithmetic the date clause hides.

How Gemini billing compares to the other API vendors

The API is not the Gemini app

People searching for Gemini pricing often want the consumer plans, which are a different bill. The Google AI plans page lists Google AI Pro and Google AI Ultra, with Ultra offering 5x higher Gemini usage limits than Pro. The page did not show a price when we fetched it, probably because it varies by country. Other writers have recorded Google AI Pro at $19.99. Those subscriptions do not buy API tokens. The API is billed in Google AI Studio and Google Cloud, and a subscription does not offset it.

Moving between vendors

If you are porting a workload from Claude or OpenAI, three things change the price more than the headline rate.

1. Thinking tokens. Gemini bills them as output. Compare outputs measured on your own prompts, not rate cards.
2. Caching style. Argon only does implicit caching, so you cannot pin a prefix. Models with explicit caching let you control it, at the cost of a storage fee.
3. Tool charges. Grounding is metered by search query, not by request, so an agent that searches three times per answer pays for three.

Keep a small evaluation set. Run it on two models, record the bill and the failures, and decide from the numbers.

Who should pick which

Our take

Google's pricing page is unusually dynamic. Between promotional rates, tier names, a contributor program and per-model grounding rules, it takes real care to forecast a bill. Treat each date clause as a price change that has already been announced.

The sensible plan is to prototype on a free Flash model, move to paid before you handle real user data, and benchmark Argon against a cheaper Flash on your own tasks before you pay for it. We did not run those tests. On paper, Flash at $1.50 and $7.50 is about 60 percent cheaper than Argon's December price, and Flash-Lite is far cheaper again.

If you want to use Gemini from a coding agent, note that Gemini CLI is now Code Assist enterprise only, and individuals moved to Antigravity CLI on 2026-06-18. See Antigravity and Gemini CLI for the tools, and AI coding agent pricing compared for plan prices.

Checklist

• Budget at the post-promotion price, not today's.
• Use a paid project for anything private.
• Count thinking tokens as output and set a thinking budget where it exists.
• Cap grounding: free requests are shared across all Gemini 3 and newer models.
• Put repeated prefixes in the cache and compare storage cost to read savings.
• Re-read the pricing page monthly. Several rates change on 1 December and 1 January.

Questions we hear

Is the free tier enough for a real product? The page describes it as for developers and small projects getting started, with lower limits and with content used to improve Google's products. A production app needs the paid tier for higher rate limits, caching, batch and privacy.

Does Gemini charge for failed requests? The page states that every token the model processes is billed. It does not discuss failed requests in the section we read.

Which Gemini model should an agent use? Google describes Gemini 3.8 Flash as built for long-horizon software engineering and autonomous agents, and Argon as its flagship for complex software engineering and multi-step agents. The cheaper one is the sensible default until your own tests say otherwise.

Can I get volume discounts? The enterprise column lists provisioned throughput and volume-based discounts through sales, via Google Cloud. The pricing page quotes none.

Are text-to-speech and live translation priced the same way? No. They have separate model entries on the page, and we did not summarize them. Check those entries if you build voice features.

Forecasting a quarter

Pick the model, take the post-promotion price, estimate tokens per request and requests per day, then add thinking tokens as output. For a worked case: 20,000 requests a day at 3,000 input and 800 output tokens on Gemini 3.8 Flash after 1 January costs 20,000 x (3,000 x $1.50 + 800 x $7.50) per million, which is $210 a day, about $6,300 a month before caching. This is our arithmetic. Caching a shared prefix and moving non-urgent work to batch are the two levers that cut it.