Gemini API Pricing 2026: Every Model per Million Tokens, and the Rates That Double in December and January

Published 27 October 2026. Every price below was read from Google's Gemini Developer API pricing page (last updated 27 October 2026 UTC) and Google's linked documentation on 27 October 2026, and every price is as of that date.

Gemini API pricing runs from $0.25 to $2.00 per million input tokens and from $1.50 to $12.00 per million output tokens on Google's current text models, as of 27 October 2026: Gemini 3.1 Flash-Lite at $0.25 and $1.50, Gemini 3.5 Flash-Lite at $0.30 and $2.50, Gemini 3.8 Flash at $0.75 and $3.75, Gemini 4 Argon at $2.00 and $10.00, and Gemini 3.1 Pro Preview at $2.00 and $12.00 (rising to $4.00 and $18.00 once a prompt passes 200,000 tokens). Two of those prices are introductory and already have an end date. Gemini 4 Argon doubles to $4.00 and $20.00 on 1 December 2026, and Gemini 3.8 Flash and 3.6 Flash double to $1.50 and $7.50 on 1 January 2027.

A rate card tells you the price of a token; your application decides how many tokens you buy, and in a chat product that number grows every turn because the whole conversation is sent again. In the worked example further down, a 40-turn conversation on Gemini 3.8 Flash costs $0.41 without caching and $0.11 with a warm implicit cache today, and $0.82 and $0.22 from January. On Gemini 4 Argon the same conversation costs $0.39 with caching, more than Claude Sonnet 5.5 or GPT-6.1 Sol at the same list price, because of a caching rule specific to Gemini 4 that none of the nine Gemini pricing pages we read before writing this one mentions.

Gemini API price per million tokens, every current text model

US dollars per million tokens on the paid Standard tier, from Google's pricing page, as of 27 October 2026. Argon and the two Flash models are shown at their introductory prices, which end on 30 November 2026 (Argon) and 31 December 2026 (3.8 and 3.6 Flash). Text, image and video input share one rate unless a separate audio rate is shown.

Every model in the table reads up to 1,048,576 input tokens, according to Google's model page for each, such as the pages for Gemini 3.8 Flash and Gemini 4 Argon. Argon is the exception in how that window works: input, thinking and output share one 1,048,576-token window, synchronous responses stop at 65,536 output tokens, and long decode with background execution lets a single response run up to 1,048,576 tokens minus the input.

Gemini 2.5 models are still on the price sheet, but Google says on its models page that it is "limiting access to the 2.5 models to users who have actively used them in the past", so a new project should not plan around them. For existing users their rates as of 27 October 2026 are $1.25 and $10.00 for Gemini 2.5 Pro (prompts up to 200K; $2.50 and $15.00 above), $0.30 and $2.50 for Gemini 2.5 Flash, and $0.10 and $0.40 for Gemini 2.5 Flash-Lite, which is still the lowest text rate on the page.

The rates that change on 1 December 2026 and 1 January 2027

Google's Gemini 4 Argon guide describes the current Argon rate as "an introductory price of $2/1M input tokens and $10/1M output tokens ($4/1M input and $20/1M output thereafter)". Nothing else on the page changes: Gemini 3.1 Pro Preview, both Flash-Lite models and the embedding model carry no end date.

A budget built on today's Flash or Argon rate is therefore a budget for the next five to ten weeks. Price any workload you expect to run into 2027 at the new rates, and treat the gap as a reason to set up caching and batching now, since both discounts are percentages and survive the increase.

Prompts over 200,000 tokens cost twice as much on 3.1 Pro

Gemini 3.1 Pro Preview and Gemini 2.5 Pro are the only text models with a prompt-size tier. On 3.1 Pro, input goes from $2.00 to $4.00 per million and output from $12.00 to $18.00 once the prompt is over 200,000 tokens. Google's Vertex AI pricing page states the rule outright: "If a query input context is longer than 200K tokens, all tokens (input and output) are charged at long context rates." The higher rate applies to every token in the request, including the ones below the threshold.

That makes the threshold a cliff. A 150,000-token prompt with a 2,000-token answer costs $0.324 on 3.1 Pro; a 250,000-token prompt with the same answer costs $1.036, which is 3.2 times the price for 1.67 times the tokens. Argon and the Flash and Flash-Lite models have no such tier. Until 1 December, Argon matches 3.1 Pro's input rate at any prompt length and undercuts its output rate, so a long-document job that does not need 3.1 Pro specifically is cheaper on Argon; after 1 December, Argon's $4.00 and $20.00 sit above even 3.1 Pro's long-context $4.00 and $18.00.

Thinking tokens are billed as output

Every output price on Google's page is labelled "including thinking tokens". The Argon section goes further: "Every token the model processes is counted and billed, including thinking tokens and internal tool system instructions and function declarations when using grounding." A request whose visible answer is 1,000 tokens and whose reasoning is 4,000 tokens bills 5,000 output tokens, which on Argon is $0.05 instead of $0.01.

On Argon you control this with thinking_level (low, medium or high, with medium as the default) and with max_output_tokens, which caps thinking and visible output together. Google's Argon guide recommends setting max_output_tokens to at least 32,768 because Argon "can spend more of the limit on thinking than earlier models", and suggests lowering the thinking level first, and leaving the cap alone, when you want fewer tokens. The token counting guide lists total_thought_tokens in the usage object, which is where reasoning spend shows up.

Context caching: implicit, explicit, and what storage costs

Gemini has two caching modes, and they bill differently. Implicit caching is on by default for Gemini 2.5 and newer models, needs no code, charges no storage, and passes on the discount automatically when a request starts with a prefix Google has already processed. Explicit caching creates a named cache object with a time to live (one hour by default, with no minimum or maximum), and bills two meters: the cached tokens at the reduced rate whenever a request uses them, and storage for every hour the cache exists.

Two caveats sit in Google's caching documentation. Implicit caching comes with "no cost saving guarantee", so a busy prefix usually hits and a quiet one may not. And a prompt shorter than the minimum is never cached, so a 3,000-token system prompt on 3.8 Flash pays full price on every call until the conversation behind it grows past 4,096 tokens. The usage field to watch is usage.total_cached_tokens.

When an explicit cache pays for its storage

An explicit cache earns its keep when the tokens it saves per hour outweigh its hourly storage. Take a 100,000-token document that requests reuse. On Gemini 3.1 Pro Preview, storage costs $0.45 an hour and each request that reads the cache instead of resending the document saves $0.18 ($0.20 instead of $0.02 for those tokens), so the cache pays for itself at three or more requests an hour. On Gemini 3.8 Flash, storage costs $0.05 an hour and each request saves $0.0675, so a single request an hour covers it, and the ratio holds after 1 January because storage and read prices both double. Delete a cache when the traffic stops; storage is billed for the whole time to live whether anyone reads it or not.

The 2,050-token rule on Gemini 4 Argon

Argon does not support explicit caching at all, and its implicit cache has a rule that matters for chat. Google's caching page says: "Gemini caches the full previous request only if the new request adds at least 2,050 tokens. Model output, thought, and function results all count toward this total. For smaller additions, Gemini re-processes the last 2,050 tokens of the previous request at the standard input rate."

An agent loop that appends a tool result of several thousand tokens per step clears that bar every time, which is why Google adds that "most agent turns exceed this threshold". A chat turn usually does not. A user message of 150 tokens and a reply of 350 add 500 tokens, so every turn pays full price on 2,550 tokens instead of 500, and at turn 40 of the example below that makes an Argon turn cost $0.0105 against $0.0069 on Claude Sonnet 5.5 at the same $2 and $10 list price.

Batch, Flex and Priority

Tier comparison from Google's Priority inference and Flex inference guides. On the pricing page itself every Priority rate is 1.8 times Standard: Argon at $3.60 and $18.00, Gemini 3.1 Pro at $3.60 and $21.60, 3.8 Flash at $1.35 and $6.75, and 3.5 Flash-Lite at $0.54 and $4.50.

Flex is the more useful of the two discounts for agent builders. It charges the Batch price but keeps a normal synchronous call, so a pipeline where each step needs the previous answer can still take the 50% discount, as long as it can wait minutes and tolerate an occasional shed request. Priority traffic that exceeds its limits is downgraded to Standard rather than failed, and its default rate limit is 0.3 times the Standard limit, according to the rate limits page. Batch jobs have their own limits: 100 concurrent batch requests, a 2GB input file and 20GB of file storage.

The Gemini API free tier in 2026

Google offers free input and output tokens on the Standard tier for Gemini 3.8 Flash, 3.6 Flash, 3.5 Flash-Lite, 3.1 Flash-Lite and 3 Flash Preview, as well as the Gemini 3.8 Live models, 3.5 Transcribe, 3.5 Live Translate, the 3.8 Flash TTS and Flash-Lite TTS models, Gemini Embedding 2 and Gemma 4. Gemini 4 Argon, Gemini 3.1 Pro Preview, Gemini Omni Flash and the Nano Banana image models have no free tier. On 3.8 Flash, context caching is free on the free tier too, while Batch and Flex are paid only.

Google no longer prints a table of free-tier rate limits. The rate limits page says limits are measured per project (not per API key) in requests per minute, tokens per minute and requests per day, that daily quotas reset at midnight Pacific time, and that your actual numbers are shown in Google AI Studio. Any page quoting a fixed free-tier RPM for 2026 models is quoting something Google does not publish. Grounding with Google Search is not available on the free tier for Gemini 3 and newer models, apart from testing in AI Studio.

You pay for the free tier with your data. Google's Gemini API terms say that on unpaid services Google uses the content you submit and the responses "to provide, improve, and develop Google products and services and machine learning technologies", and that "human reviewers may read, annotate, and process your API input and output". The terms carve out the European Economic Area, Switzerland and the United Kingdom, where the paid-service data terms apply to all services, and they require that you "use only Paid Services when making API Clients available to users" in those regions. A free key is a prototyping key; anything that touches customer data belongs on a paid project.

A Google AI Pro or Ultra subscription does not change any of this. Google's AI plans page says the plan benefits for developers "apply only within the Google AI Studio web interface" and that direct API use "is billed and managed separately".

What a multi-turn conversation costs, and how Gemini compares with Claude and OpenAI

Take a chat assistant with a 2,000-token system prompt, where each user message is 150 tokens and each reply is 350 tokens. Turn 1 sends 2,150 input tokens. Turn 40 sends 21,650, because it carries the 39 exchanges before it, and the whole 40-turn conversation consumes 476,000 input tokens to produce 14,000 output tokens. A 100-turn conversation consumes 2,690,000 input tokens. Thinking tokens are left out, so a reasoning model would add to every output figure below.

Rates for the comparison come from Anthropic's pricing page (Sonnet 5.5 at $2 input, $10 output, $2.50 five-minute cache write and $0.10 cache read; Opus 5.5 at $4, $20, $5 and $0.20) and OpenAI's pricing page (GPT-6.1 Sol at $2, $10, $2.50 thirty-minute cache write and $0.10 cached input; GPT-6.1 Terra at $1, $5, $1.25 and $0.05, short-context rates for prompts up to 272K tokens), all read on 27 October 2026.

The three providers cache differently, so the way "cache warm" is modelled matters. For Gemini it is the best case of implicit caching: from the first turn whose previous request was over 4,096 tokens (turn 6 here), the previous request's prompt is read at the cached rate and the new 500 tokens are billed at full input price, with Argon also paying full price on the 2,050 tokens its rule re-processes. For Claude and OpenAI it is automatic or implicit caching with a five-minute or thirty-minute cache, where each turn writes its new 500 tokens at 1.25 times the input price and reads everything before them; Claude caches from 512 tokens and OpenAI from 1,024 tokens, according to Anthropic's and OpenAI's caching guides, so both cache from turn 1.

Four things come out of the table. Without caching, Argon, Claude Sonnet 5.5 and GPT-6.1 Sol cost exactly the same, $1.09, because all three list at $2 and $10 today. With caching, Argon is the most expensive of the three at $0.39, because of the 2,050-token rule. Gemini 3.8 Flash is the cheapest model in the table on a warm cache at $0.11 until the end of the year, and at $0.22 from January it lands just above GPT-6.1 Terra at $0.12 and still below Claude Sonnet 5.5 at $0.24. And on 1 December Argon's cached conversation doubles to $0.77, more than Claude Opus 5.5 at $0.48.

A conversation's cost also grows faster than its length. Going from 40 to 100 turns multiplies turns by 2.5 and the uncached 3.8 Flash bill by more than five, because every turn carries every turn before it. At 10,000 forty-turn conversations a month, 3.8 Flash costs $4,095 without caching and $1,106 with it today, and $8,190 and $2,213 from January.

Caching is the first lever, and it has two limits: Gemini's implicit cache carries no cost-saving guarantee, and no cache helps a user who leaves and comes back after it has expired. The other lever is to stop resending the whole transcript. We build Maximem Synap, a memory layer that returns a short, ranked set of facts about the user, trimmed to a token budget you set (2,000 tokens by default), in place of the full history. When that beats caching, and when it does not, is worked through in how to reduce LLM token costs in long conversations, and the same arithmetic on Anthropic's rate card is in our Claude API pricing guide.

Grounding with Google Search and Google Maps

Prices as of 27 October 2026 from Google's pricing page.

On Gemini 3 and newer, Google bills per search query, not per request: "A customer-submitted request to Gemini may result in one or more queries to Google Search. You will be charged for each individual search query performed." If 10,000 grounded prompts a month each trigger two searches, that is 20,000 queries, 15,000 of them past the free allowance, which costs $210 a month on top of the tokens. Vertex AI states that "input tokens provided by Grounding with Google Search or Web Grounding for Enterprise are not charged", and on the Developer API, Argon bills the internal tool instructions that grounding adds. With dynamic retrieval, only requests whose response contains at least one grounding URL are charged for grounding.

Tools, embeddings, and the image, video and speech models

Tools. Code execution is billed at the model's normal token rates with no charge for runtime: generated code and results bill as output when created and as input when the model reads them back. URL context bills the fetched content as input tokens, and computer use bills as regular tokens. File Search charges $0.15 per million tokens for embeddings at indexing time, with storage and query-time embeddings free and retrieved document tokens billed as context, according to the File Search guide. The Deep Research agent, managed agents and the Antigravity agent bill model inference at list rates, and Google does not bill their environment compute during the preview.

Embeddings. Gemini Embedding 2 costs $0.20 per million text tokens ($0.10 in batch), $0.45 per million image tokens (about $0.00012 per image), $6.50 per million audio tokens (about $0.00016 per second) and $12.00 per million video tokens (about $0.00079 per frame), and it is free on the free tier.

Batch rates for the image models are half the Standard rates shown. Per-image and per-minute figures are Google's own conversions on the pricing page.

How Gemini counts tokens

Google's token guide puts a token at about four characters, and 100 tokens at about 60 to 80 English words. An image no larger than 384 pixels in both dimensions counts as 258 tokens, and a larger image is tiled into 768 by 768 pixel tiles at 258 tokens each. Audio counts as 32 tokens per second, and PDF pages are billed at the image token rate. Call count_tokens before sending a request to see its input size, and read total_input_tokens, total_output_tokens, total_thought_tokens and total_cached_tokens in the usage object afterwards.

Gemini API vs Vertex AI pricing

Vertex AI, which Google now sells as the Gemini Enterprise Agent Platform, serves the same models at the same rates on its global endpoint. The differences are on its pricing page:

• Regional endpoints cost 10% more. Non-global endpoints list Argon at $2.20 and $11.00 today and $4.40 and $22.00 from 1 December, and 3.8 Flash at $0.825 and $4.125.
• Argon's introductory price arrives as a credit. On Vertex the lower rate is "provided through 50% credits back on net spend on select models within a given period", so your invoice shows the full rate and the credit separately.
• Long completions bill even when they fail. From 13 November 2026, a Gemini 4 completion longer than 32K tokens is "charged for retrievable token usage even if the request ultimately fails due to client or server timeouts."
• Error responses are not charged. Apart from the long-completion rule above, Vertex charges only for requests that return a 200 response code.

Vertex also adds Flexible Savings Plans (spend-based committed use discounts), grounding on your own data at $2.50 per 1,000 prompts, and a 1.5 times price on tuned-model endpoints in non-global regions. The Developer API is the shorter path to a first production deployment; Vertex is where you go for regional data residency, enterprise controls and a committed discount.

How Gemini API billing works: prepay, usage tiers and spend limits

Tiers and caps from Google's billing and rate limits pages, as of 27 October 2026.

New accounts default to prepay: upgrading from the free tier means linking a billing account and adding at least $5 of credit. Two operational details deserve a place in your runbook. When the prepay balance reaches $0, "all API keys in all projects linked to that billing account will stop working simultaneously" and return HTTP 402 until you add credit, and because billing lags usage by about ten minutes, batch jobs and agents can run past the balance before they stop. Unused prepaid credit expires after 12 months and is not refundable. The spend-based limit is separate from the token limits: a Tier 1 project that spends more than $10 in a rolling ten-minute window gets 429 errors, which a handful of long-context Argon requests can reach.

Older models and shutdown dates

Google's deprecations page lists what to migrate off, as of 27 October 2026. Gemini 3.1 Flash-Lite shuts down on 7 May 2027, with 3.5 Flash-Lite as the replacement at $0.30 and $2.50 instead of $0.25 and $1.50. Gemini 3.1 Flash TTS Preview shuts down on 17 November 2026. The gemini-3.7-flash endpoint now routes to Gemini 3.8 Flash and gemini-3.5-flash routes to 3.6 Flash, so code still calling either is billed at the 3.8 or 3.6 Flash rate. Gemini 2.5 Flash Image was scheduled to shut down on 2 October 2026. Gemini 2.5 Pro, Flash and Flash-Lite have no shutdown date but are limited to existing users, and Gemini 3.1 Pro Preview and 3 Flash Preview have no shutdown date announced.

How to estimate your monthly Gemini bill, and cut it

A request costs (uncached input tokens times the input rate) plus (cached tokens times the cached rate) plus (output and thinking tokens times the output rate), plus grounding queries past the free allowance and any explicit cache storage. For a conversation, input per turn is the system prompt plus every earlier turn, so estimate per conversation as in the table above, then multiply by conversations per month, and price anything running past 1 December (Argon) or 1 January (3.8 and 3.6 Flash) at the new rates.

Six levers, roughly in order of how much they move a typical bill:

1. Pick the cheapest model that passes your own evals. 3.5 Flash-Lite costs 40% of 3.8 Flash on input today and a fifth of it from January; run your prompts on both before committing.
2. Keep a stable prefix and let implicit caching work. Put the system prompt and fixed context first and per-turn content last, and make sure the prefix passes 4,096 tokens on Argon, 3.8 Flash and 3.1 Pro if you want it cached at all.
3. Move anything that can wait to Batch or Flex for half price.
4. Cap thinking. Lower thinking_level before lowering max_output_tokens, and watch total_thought_tokens.
5. Stay under 200,000 tokens on 3.1 Pro, or use a model without the long-context tier.
6. Stop resending the whole conversation once chats run long, by summarising or retrieving only what the next turn needs.

Frequently asked questions

How much does the Gemini API cost per million tokens?

As of 27 October 2026: Gemini 4 Argon $2.00 input and $10.00 output, Gemini 3.1 Pro Preview $2.00 and $12.00 (or $4.00 and $18.00 above 200,000 tokens), Gemini 3.8 Flash and 3.6 Flash $0.75 and $3.75, Gemini 3.5 Flash-Lite $0.30 and $2.50, and Gemini 3.1 Flash-Lite $0.25 and $1.50. Batch and Flex halve those rates, and cached input costs 5% to 10% of the input rate.

Is the Gemini API free?

Partly. The free tier covers Gemini 3.8 Flash, 3.6 Flash, both Flash-Lite models, 3 Flash Preview, the Live, transcription and TTS models, and Gemini Embedding 2, within rate limits shown per project in Google AI Studio. Gemini 4 Argon and Gemini 3.1 Pro Preview are paid only. On the free tier Google may use your prompts and responses to improve its products, except in the EEA, Switzerland and the UK.

Is Gemini 3.8 Flash getting more expensive?

Every Gemini 3.8 Flash and 3.6 Flash rate doubles on 1 January 2027: input to $1.50, output to $7.50, cached input to $0.15 and cache storage to $1.00 per million tokens per hour. Gemini 4 Argon doubles earlier, on 1 December 2026, to $4.00 input and $20.00 output.

Does Gemini charge for thinking tokens?

Thinking tokens are billed at the output rate on every model, and Argon also bills internal tool instructions added by grounding. Control them with thinking_level and max_output_tokens, and check total_thought_tokens in the usage object.

How much does Gemini context caching cost?

Implicit caching is automatic and charges only the cached-input rate, $0.10 per million on Argon and $0.075 on 3.8 Flash as of 27 October 2026, with a 4,096-token minimum on current models. Explicit caching adds storage, from $0.50 per million tokens per hour on 3.8 Flash to $4.50 on 3.1 Pro, and is not available on Argon.

Is Gemini cheaper than Claude or GPT?

On the rate card, Gemini 3.8 Flash at $0.75 and $3.75 sits below Claude Sonnet 5.5 and GPT-6.1 Sol at $2 and $10, while Argon matches them until 1 December. On a 40-turn chat with caching, 3.8 Flash costs $0.11 against $0.24 for Sonnet 5.5 and GPT-6.1 Sol, but Argon costs $0.39 because its implicit cache re-processes 2,050 tokens on every short turn.

Is Vertex AI more expensive than the Gemini API?

Not on the global endpoint, where rates match. Regional (non-global) endpoints cost 10% more, and Vertex delivers Argon's introductory price as a 50% credit on net spend, with the full rate on the invoice.

What happens when my Gemini prepay balance runs out?

Every API key on that billing account stops working and returns HTTP 402 until you add credit. Because billing lags by about ten minutes, long batch jobs and agents can overshoot the balance first.

Sources

All retrieved 27 October 2026.

• Gemini Developer API pricing, Google (page last updated 27 October 2026 UTC)
• Google Gemini API documentation: context caching, explicit caching, the Gemini 4 Argon guide, Batch API, Flex inference, Priority inference, rate limits, billing, tokens, models, deprecations, File Search, Google AI plans and terms of service
• Gemini Enterprise Agent Platform (Vertex AI) generative AI pricing, Google Cloud
• Claude API pricing and prompt caching, Anthropic
• OpenAI API pricing and prompt caching guide, OpenAI

Worked examples are our own arithmetic on those rates.