Gemini API Pricing (October 2026): Gemini 4 Argon, 3.8 Flash, Pro, and Flash-Lite Per-Token Costs

The Full Gemini API Price List

The table below lists every current Gemini text model and its standard pay-as-you-go rate per million tokens. For tiered models, the rate shown is the base tier, which applies to requests up to 200k tokens of context. The over-200k surcharge, processing tiers, and context-caching rates are explained in the sections below. Audio input, where it differs, is noted.

Two patterns stand out. First, output costs 3x to 8x more than input across the lineup, so token-heavy generation tasks (long code, long documents) cost far more than read-heavy tasks (classification, extraction). Second, the gap between the cheapest and most expensive models is large: gemini-2.5-flash-lite output at $0.40/1M is 30x cheaper than gemini-3.1-pro output at $12.00/1M. Gemini 4 Argon output at $10/1M sits below gemini-3.1-pro-preview at $12/1M until December 1, when Argon doubles to $20/1M.

Gemini 4 Argon Pricing

gemini-4-argon is Google's flagship reasoning model, listed for long-horizon coding, math, and multi-step agents with generation up to 1M tokens. It is paid-tier only. Its prices are dated: the current rates run through November 30, 2026, and every Argon rate doubles on December 1.

Argon supports implicit caching only, with cached input at $0.10/1M ($0.20/1M from December 1). Explicit context caching is not supported. A job that sends 3M input and 0.9M output tokens a day costs $15.00 a day on Argon standard today and $30.00 a day from December 1. The same job on gemini-3.8-flash costs $5.63 a day.

Gemini API Free Tier: Which Models Are Free

The free tier charges nothing for input and output tokens and runs through Google AI Studio. Google uses free-tier content to improve its products; paid-tier content is not used. Rate limits are lower than on the paid tier, and context caching is limited.

The two models without a free tier are the two most expensive ones, Argon and 3.1 Pro Preview. Prototype on gemini-3.8-flash for free, then move to the paid tier when you need higher limits or need your prompts kept out of training.

Grounding with Google Search Pricing

Grounding lets Gemini run Google Search (or Google Maps) queries and cite the results. It is billed per search request on top of tokens, and it is a paid-tier feature on the 3.x and 4 models.

On Gemini 3 and newer, one customer request can trigger several search queries, and each query is billed. An agent that grounds 20,000 requests a month at two queries each uses 40,000 queries: 5,000 free, then 35,000 at $14 per 1,000, or $490 a month before tokens.

How Gemini Charges: Per-Token and Tiered

Every Gemini API request is billed on two counters: input tokens (your prompt, system instructions, files, and conversation history) and output tokens (the model's response). You are charged the model's per-million rate for each, prorated to the exact token count. A request that uses 12,000 input tokens and 800 output tokens on gemini-2.5-flash costs 12,000 / 1,000,000 multiplied by $0.30, plus 800 / 1,000,000 multiplied by $2.50, which is $0.0036 plus $0.002, for $0.0056 total.

The wrinkle that breaks naive cost math is the context tier. For gemini-3.1-pro and gemini-2.5-pro, the per-token rate is not fixed. It depends on the total context length of the request. Stay under 200k tokens and you pay the base rate. Cross 200k and the higher rate applies to the whole request.

Models marked "flat" in the table (every Flash and Flash-Lite model) charge one rate regardless of context length. This makes them simpler to budget for, which is another reason high-volume workloads gravitate to the Flash and Flash-Lite tiers.

Batch, Flex, and Priority Tiers

The standard rate is not the only rate. Gemini sells the same models at three other processing tiers: Batch and Flex at half price, Priority at roughly 1.8x.

The multipliers hold across the lineup: gemini-2.5-flash-lite bills $0.05/$0.20 on Batch and $0.18/$0.72 on Priority, and gemini-3.1-pro bills $1.00/$6.00 on Batch (base tier). The spread between Batch and Priority on the same model is 3.6x, so separating traffic that can wait from traffic that cannot is a direct cost lever before any model change.

The Over-200k Context Surcharge

The Pro models raise their price once a request crosses 200,000 tokens of context. The higher rate applies to the entire request, not only the tokens past 200k. A 201k-token request to gemini-3.1-pro is billed entirely at the above-200k rate.

For gemini-3.1-pro, crossing 200k doubles the input rate and raises output by 50%. For gemini-2.5-pro the input rate doubles and output rises 50%. The Flash and Flash-Lite models have no over-200k tier: an earlier version of this page listed one for gemini-3-flash, but its second listed input rate ($1.00/1M) is the audio rate, not a long-context tier.

Context Caching Pricing

Context caching lets you store a fixed block of input (a long system prompt, a reference document, a codebase snapshot) once and reuse it across many requests at a reduced per-token rate. Gemini bills caching two ways: a per-token rate to read from the cache, plus a per-hour storage fee for keeping the cache warm.

Caching pays off when the same large context is reused enough times that the discount on input tokens outweighs the storage fee. For gemini-2.5-flash, cached input at $0.03/1M is 10x cheaper than the $0.30/1M standard input rate. If you send the same 100k-token document to 50 requests in an hour, caching it once and reading it cheaply beats paying full input price 50 times, even after the $1.00/hr storage fee.

Caching does not help one-off requests or workloads where the context changes every call. It is a fit for chat sessions over a fixed document, agents grounded in a static knowledge base, or batch jobs that share a long system prompt.

Audio Input Costs More

For the models that accept audio, audio input tokens are billed at a higher rate than text, image, or video input. This is a separate counter, so a multimodal request can mix rates within a single call.

On gemini-2.5-flash, audio input is $1.00/1M versus $0.30/1M for text, more than 3x. If your workload streams audio (transcription, voice agents), the audio rate, not the text rate, drives the bill. Budget against the audio column for those use cases.

Worked Cost Example

Take a workload of 1,000,000 input tokens and 200,000 output tokens, all under 200k context per request. The table compares the cost on three tiers using standard base rates from the price list.

The same workload costs $0.18 on Flash-Lite, $0.80 on Flash, $1.50 on the new gemini-3.8-flash, and $4.40 on Pro. Flash-Lite is 24x cheaper than Pro here. If a meaningful fraction of these prompts are simple enough for Flash-Lite, sending all of them to Pro overpays by an order of magnitude. The whole pricing structure rewards matching the model to the task.

Monthly Cost Scenarios

The table prices three usage profiles over a 30-day month at standard-tier text rates, no caching, all requests under 200k context. Light and medium assume 1,000 input + 300 output tokens per request (a short chat or extraction call). Heavy assumes an agent workload at 3,000 input + 1,000 output tokens per request.

The arithmetic behind the first gemini-3.8-flash cell: 3M input tokens at $0.75/1M is $2.25, plus 0.9M output tokens at $3.75/1M is $3.375, totaling $5.63. Every other cell is the same two multiplications. Batch halves any of these numbers for traffic that can wait, and the January 2027 doubling turns the gemini-3.8-flash column into $11.25, $112.50, and $3,600 unless Google revises the schedule.

OpenAI-Compatible vs Native Calls

Gemini exposes two API surfaces at the same per-token prices. The native Gemini SDK (google-genai) uses Google's request and response shapes. The OpenAI-compatible endpoint accepts the standard OpenAI Chat Completions format, so existing OpenAI-SDK code can target Gemini by swapping the base URL and model name. Pricing is identical between the two; only the request schema differs.

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.GEMINI_API_KEY,
  baseURL: "https://generativelanguage.googleapis.com/v1beta/openai/",
});

const response = await client.chat.completions.create({
  model: "gemini-2.5-flash",
  messages: [{ role: "user", content: "Summarize this changelog." }],
});

// Billed at gemini-2.5-flash rates: $0.30/1M input, $2.50/1M output.
// Same prices as the native google-genai SDK.

The OpenAI-compatible surface matters for cost control because it lets one codebase address Gemini, OpenAI, and other providers through the same client. That is the foundation routing builds on: if every model speaks the same request format, a router can pick the cheapest one that clears the quality bar without rewriting call sites.

Cutting the Bill with Routing

The 24x spread between Flash-Lite and Pro is the cost-savings opportunity. Most prompts in a real workload (classification, extraction, simple edits, short answers) clear the bar on Flash-Lite or Flash. A minority (multi-step reasoning, long-context synthesis) need Pro. Sending everything to Pro pays Pro prices for work a $0.10/1M model handles identically.

A model router classifies each prompt by difficulty, then sends it to the cheapest model that can handle it. Easy turns drop to Flash-Lite; hard turns reach Pro. Across a mixed workload this cuts 40-70% off the bill. Morph Router classifies a prompt in ~430ms at $0.005 per request and returns the model to use. It does not proxy the call: your code sends the request to Gemini (or any other provider) with the model it picked.

The router also keeps requests on the right side of the 200k context line where it can, since that boundary doubles the input rate on Pro models. Combined with caching for reused context, routing plus a model-tier policy is the practical way to keep a Gemini bill proportional to the actual difficulty of the work. See how the LLM router works, the broader cost-optimization guide, and the cost calculator.

Frequently Asked Questions

Sources

• Google AI for Developers: Gemini API pricing (ai.google.dev), fetched October 25, 2026

Related Resources

• What Is an LLM Router? Automatic Model Routing for Cost and Quality
• LLM Cost Optimization
• LLM Cost Calculator
• Anthropic API Pricing
• OpenAI API Pricing
• LLM API Comparison
• LLM Context Windows Compared
• Best AI Model for Coding
• LLM Gateway