Google Gemini API Pricing Guide 2026: Flash, Pro, and Vertex AI

Google's Gemini line still has one thing the others don't quite match: a 1 million token context window on every model, right down to the cheapest. Rivals have reached 1M at the top of their lineups, but not at Flash-Lite prices. If your work involves long documents, whole codebases, or hours of transcript, that changes the cost math more than any per-token headline does.

This guide covers every Gemini model you can actually call as of July 2026: the Gemini 3 generation that now leads, the 2.5 models on their way out, image generation with Nano Banana, and what Vertex AI adds for production. Compare any of them against OpenAI, Anthropic, and others in our LLM Pricing Calculator.

All prices are per million tokens on the standard tier, verified against Google's official pricing pages on 12 July 2026.

The full Gemini pricing table (July 2026)

Above 200K tokens, the Pro models charge a higher rate: Gemini 3.1 Pro moves to $4/$18 and Gemini 2.5 Pro to $2.50/$15. The Flash and Flash-Lite tiers stay flat no matter how long the prompt.

What's already gone: Gemini 2.0 Flash and 2.0 Flash-Lite were shut down on 1 June 2026, and the Gemini 1.5 family now returns a 404. If any of your code still points at them, migrate now. 2.0 Flash maps to 3.5 Flash, and 2.0 Flash-Lite to 3.1 Flash-Lite. The 2.5 models still work, but Google has set their retirement for 16 October 2026, so treat them as a migration target, not a foundation.

The 1 million token context window

Every Gemini model, including Flash-Lite, supports a 1 million token context window. A year ago that window was Gemini's alone. It isn't now: OpenAI's frontier GPT-5.x models and Anthropic's Claude Sonnet 5, Opus 4.8, and Fable 5 all reach 1M too. What stays distinctive is that Gemini offers it on every tier, down to the cheapest. Where the frontier labs land today:

• Gemini: 1M tokens on every tier, down to Flash-Lite
• OpenAI: ~1M on the frontier GPT-5.x models; the cheaper nano and mini tiers cap at 400K
• Anthropic Claude: 1M on Sonnet 5, Opus 4.8, and Fable 5; Haiku 4.5 stays at 200K
• Mistral: 256K maximum (Large 3, Small 4); no 1M option

What does 1M tokens look like in practice? Roughly 750,000 words. A 3,000-page book. An entire medium-sized codebase, or eight-plus hours of transcribed audio.

The practical impact is that you can often skip building chunking pipelines, RAG systems, or summarization chains. Load the whole document into context and ask your question. That cuts engineering complexity, and it can cut cost too, by removing several LLM calls from a retrieval pipeline you no longer need.

The long-context pricing catch

Google prices long prompts in two bands. On the Pro models, once input crosses 200K tokens, both input and output jump to a higher rate. Gemini 3.1 Pro goes from $2/$12 to $4/$18, and Gemini 2.5 Pro from $1.25/$10 to $2.50/$15. The Flash and Flash-Lite tiers stay flat no matter how long the prompt. So the 1M window is cheap to fill on Flash and priced as a premium on Pro. For the common case, prompts under 200K (still around 150,000 words), standard rates apply.

Understanding the model families

Gemini 3: the current generation

Gemini 3 launched in November 2025, and the line has moved fast since. As of July 2026 there are four models you'll actually reach for:

• Gemini 3.1 Pro (preview, $2/$12) is the flagship for the hardest reasoning, world knowledge, and multimodal work. It defaults to high thinking, so reasoning tokens bill as output, and it has no minimal-thinking mode.
• Gemini 3.5 Flash (GA, $1.50/$9) is now Google's default model across the Gemini app, AI Mode in Search, and Enterprise. It's tuned for agentic and coding work, and it beats 3.1 Pro on several agentic benchmarks while running several times faster.
• Gemini 3 Flash (preview, $0.50/$3) gives you Pro-grade reasoning at Flash cost and latency. It's the value pick when 3.5 Flash is more than you need.
• Gemini 3.1 Flash-Lite (GA, $0.25/$1.50) is the high-volume workhorse: cheapest in the 3.x line, latency-optimized, with full thinking-level control.

All four carry the 1M-token context window and native multimodal input (text, image, audio, video, PDF, code). You set reasoning depth per request with the thinking_level parameter, from minimal up to high.

Two things to know. Google has announced Gemini 3.5 Pro but hasn't shipped it yet, so the current callable flagship is 3.1 Pro. And Deep Think, the high-compute reasoning mode, ships through the Google AI Ultra subscription in the app, not as a billed API model.

Gemini 2.5: the outgoing line

Gemini 2.5 Pro ($1.25/$10), Flash ($0.30/$2.50), and Flash-Lite ($0.10/$0.40) are still live and still cheap. In fact 2.5 Flash-Lite is the single cheapest Gemini you can call right now. But Google has scheduled all three for shutdown on 16 October 2026. They're fine for a short-lived project, or to squeeze out the lowest per-token cost today. For anything you'll still be running past Q4, start on the 3.x models.

Image generation: the Nano Banana family

Gemini also does native image generation, and here the pricing shifts from tokens to images. The Nano Banana line has been generally available since May 2026:

• Nano Banana Pro (gemini-3-pro-image) is the high-end model: roughly $0.13 per image at 1K to 2K resolution, about $0.24 at 4K. It renders legible text, supports multi-turn editing, and holds character consistency across up to 14 reference images.
• Nano Banana 2 (gemini-3.1-flash-image) runs about $0.045 for a 512px image, up to $0.15 at 4K.
• Nano Banana 2 Lite is the cheapest, around $0.034 for a 1K image.
• The original Nano Banana (gemini-2.5-flash-image, about $0.039 per image) still works but retires 2 October 2026.

Under the hood these bill per output token ($120, $60, and $30 per 1M image-output tokens respectively). The per-image figures are Google's own rounding of that math.

Cutting costs: batch, caching, and tiers

Three levers materially change what you actually pay.

• Batch API submits async jobs for roughly 50% off standard rates, on a 24-hour SLA. It's the obvious choice for document pipelines and evals.
• Context caching lets you cache a repeated prefix (a system prompt, a big reference doc) and pay a fraction for those tokens on every later call. Cached input runs about $0.20/1M on Pro, versus $2 fresh.
• Priority and Flex tiers let you pay up for lower latency and guaranteed capacity, or take cheaper Flex and Batch pricing when latency doesn't matter.

For a document pipeline that reuses the same instructions across thousands of files, batch and caching together can cut the bill by more than half.

Real cost calculations

Use case 1: codebase analysis

The setup: analyze a medium-sized codebase (200K tokens) once a week, producing about 5,000 tokens of analysis, four runs a month. That's 0.8M input and 0.02M output tokens. With the 1M window you load the whole codebase in one call, so no chunking and no RAG.

The cost is almost irrelevant here: under $2 a month even on the flagship. The context window is the real differentiator. A model capped at 128K or 200K would be right at its limit on a 200K-token codebase, with no room left for the answer.

Try this calculation

Use case 2: document processing pipeline

The setup: 2,000 documents a month, about 8,000 tokens each, with 1,500 tokens of structured output per document. That's 16M input and 3M output tokens.

At under half a cent per document on Flash-Lite, you can process all 2,000 for the price of a coffee, and Batch pricing halves even that. Add context caching for the shared extraction prompt and the effective cost drops further.

Try this calculation

Use case 3: real-time chat application

The setup: 15,000 conversations a month, about 1,200 tokens of input each (message plus context) and 600 tokens of output. That's 18M input and 9M output tokens.

Gemini 3.1 Flash-Lite at $18 a month for 15,000 conversations is hard to argue with. Step up to 3 Flash for stronger reasoning when conversations get complex, and save 3.5 Flash or 3.1 Pro for the flows that genuinely need them.

Try this calculation

Google AI Studio vs Vertex AI

Google gives you two ways to reach Gemini.

Google AI Studio is the simpler path: API keys, straightforward pricing, and a genuinely useful free tier. Best for startups, prototypes, and small-scale production.

Vertex AI is Google Cloud's enterprise platform. It adds SLAs and enterprise support, VPC Service Controls for data residency, Model Garden for third-party models, batch prediction, and provisioned throughput for guaranteed capacity.

Base model pricing is generally the same across both, with two caveats worth budgeting for. From 1 July 2026, Vertex charges roughly 10% more on non-global (regional) endpoints, so data-residency requirements carry a small surcharge. And enterprise features like provisioned throughput cost extra. The real reason to pick Vertex is organizational: enterprise compliance, data that can't leave your GCP project, or guaranteed throughput.

One data-privacy line is worth knowing. On the free AI Studio tier, your prompts can be used to improve Google's products. On any paid tier, and on all of Vertex, they are not. For anything touching customer data, that alone is reason enough to stay off the free tier in production.

The free tier

Google still offers the most generous free tier among the major labs. Real models, rate-limited, no credit card:

• Rate-limited, with limits that vary by model and are visible in AI Studio
• Covers the Flash and Flash-Lite line (3.5 Flash, 3 Flash, 3.1 Flash-Lite, the 2.5 Flash models, and Gemma). Pro models are paid-only
• No credit card required

OpenAI and Anthropic both require paid accounts for API access, so this stays a real advantage for prototyping. Same catch as above: free-tier prompts may be used to improve Google's products, so keep prototyping and production data separate.

Grounding with Google Search

Grounding adds live Google Search results to a response, billed on top of tokens, and the rate now depends on generation. Gemini 3 models get 5,000 grounded queries a month free, then $14 per 1,000. The 2.5 models get a smaller daily free allowance, then $35 per 1,000. Grounding input tokens themselves aren't charged. For static tasks like document analysis, code review, or content generation, you don't need it.

How Gemini compares

Where Gemini sits against the competition:

vs OpenAI. OpenAI's frontier GPT-5.x models now match Gemini's 1M context, so the window itself isn't the differentiator anymore. Price is. OpenAI's cheap tiers (nano at $0.20/$1.25, mini at $0.75/$4.50) cap at 400K context, and its 1M-context models start around $2 input (legacy GPT-4.1) or $2.50 (GPT-5.4). Gemini 3.1 Flash-Lite gives you the full 1M at $0.25/$1.50. On frontier reasoning the two trade blows, so check the calculator for current GPT pricing.

vs Anthropic. Claude caught up on context: Sonnet 5, Opus 4.8, and Fable 5 all offer a 1M window now, GA, with no long-context surcharge. Where Gemini wins is price. Claude's cheapest 1M-context model is Sonnet 5 at $2/$10, its sub-dollar tier (Haiku 4.5 at $1/$5) is capped at 200K, and there's no free tier. Claude still tends to lead on nuanced writing and complex instruction-following, so its premium can be worth paying for those tasks.

vs Mistral and open models. Mistral is cheap and has grown its context to 256K (Large 3 at $0.50/$1.50, Small 4 at $0.15/$0.60), well past its old 128K cap. But 256K is the ceiling. For genuinely long context, whole codebases or hours of transcript, Gemini's 1M window is still the one to beat.

The decision framework

• Need a big context window cheaply: Gemini is still the default. OpenAI and Claude reach ~1M at the frontier now, and Mistral tops out at 256K, but none of them match a full 1M window at sub-dollar input prices. Use Flash-Lite or 3 Flash if cost matters, 3.1 Pro if you need deep reasoning at length. If your prompts stay under 400K, it's worth pricing OpenAI's nano tier and Mistral's cheap models too.
• Cost-sensitive and high-volume: Gemini 3.1 Flash-Lite ($0.25/$1.50) is hard to beat going forward, or 2.5 Flash-Lite ($0.10/$0.40) if you're comfortable migrating before it retires in October.
• Need reasoning: Gemini 3.1 Pro for the hardest problems, or 3.5 Flash for agentic and coding work, where it beats Pro on several benchmarks at lower cost.
• Already on GCP: Vertex's IAM, VPC, logging, and monitoring integration cuts operational overhead versus juggling separate API keys.
• Creative and writing quality matters most: Claude still often edges Gemini here, so the Anthropic premium may be worth paying for those tasks.

Prices verified against Google's Gemini Developer API pricing and Vertex AI pricing on 12 July 2026. Preview-model pricing (3.1 Pro, 3 Flash) can change at GA. Compare Gemini against OpenAI and Anthropic in our LLM Pricing Calculator.