Gemini models explained: the 2026 family, in plain English

Gemini moves fast, and a lot of guides (including the earlier version of this one) still describe the 2.5 family as the state of the art. That is now two generations out of date. Since November 2025 Google has shipped the 3, 3.1 and 3.5 lines, plus a whole image-generation family called Nano Banana that is easy to confuse with the on-device Nano model. This guide is the current, accurate map, with interactive tools to help you choose.

Why does Gemini have so many models?

Because intelligence, speed and cost pull against each other. A model powerful enough to reason through a tangled coding problem is heavier and slower than one tuned to sort thousands of short messages a second. Rather than force every task through one expensive model, Google offers a tiered family so you spend capability only where it earns its keep. Read it as a toolkit, where each tool is sharpened for a specific kind of work.

Every model name is generation plus role. The number marks the generation: 3.5 is the newest line, 3 and 3.1 are the current generation that replaced 2.5, and 2.5 and 2.0 are now legacy, kept only for teams that have not migrated. The word marks the role: Pro for deep thinking, Flash for high velocity, Flash-Lite for sheer volume, and Nano for running on the device itself. Once you read a name as generation plus role, the lineup stops looking crowded and starts looking deliberate.

3.5 · newest line

3.5 Flash (GA, fast + cheap) · 3.5 Pro (preview: 2M context, Deep Think)

3 / 3.1 · current generation

3.1 Pro (GA flagship) · 3 Flash · 3.1 Flash-Lite · Nano (on-device)

2.5 / 2.0 · legacy

2.5 Pro, 2.5 Flash, 2.5 Flash-Lite and 2.0, now superseded; migrate new work to 3.x or 3.5.

Which Gemini model should you use? (interactive picker)

Answer four quick questions and we will recommend a model and tell you why. It updates as you click.

Gemini 3.5 Flash: the new default

Released at Google I/O on 19 May 2026, Gemini 3.5 Flash is the newest generally available model and, for most teams, the right starting point. It delivers frontier performance for agents and coding, and notably outperforms the previous flagship, 3.1 Pro, on many challenging coding and agentic benchmarks while being far cheaper and faster. It is the model behind the gemini-flash-latest alias, and it is live in the Gemini app, Search's AI Mode and the API.

On cost, Google priced it at roughly US$1.50 per million input tokens and US$9 per million output tokens at launch, about 40% cheaper per token than 3.1 Pro, with output speed several times faster than comparable frontier models (Google). For anything user-facing, that combination of speed and price is exactly why Flash earns its keep.

Gemini 3.1 Pro: today's flagship for deep work

Released on 19 February 2026, Gemini 3.1 Pro is Google's generally available flagship for the hardest work: complex reasoning, software development, scientific analysis and rich multimodal understanding across text, images, audio and video. It carries a 1-million-token context window, so it can hold huge documents, long transcripts or entire codebases in mind and reason across all of it at once. Reach for Pro when the answer has to be right and the problem has many moving parts, then step down to a lighter model once you have confirmed the quality you need.

Gemini 3.5 Pro: announced, not yet shipped

Announced at I/O on 19 May 2026, Gemini 3.5 Pro is the coming flagship. Its headline targets are a 2-million-token context window, the largest of any production frontier model announced so far, and a built-in Deep Think reasoning mode for the hardest problems. Treat these as targets, not confirmed facts: as of late June 2026 it remains in limited preview for some Vertex AI customers and has not reached the public Gemini app, with broad availability expected around mid-2026 (Gemini API docs). Until it ships, 3.1 Pro is the flagship you can actually deploy.

Gemini 3.1 Flash-Lite: cheap work at volume

Released on 3 March 2026, Flash-Lite is the most economical cloud model, tuned for enormous volumes of simple, repetitive work: classifying support tickets, tagging products, detecting sentiment, moderating short messages or extracting fields from forms. A useful pairing is to run a live chatbot on 3.5 Flash, then process the transcripts in bulk on Flash-Lite to surface trends and intents. One model holds the conversation; the cheaper one reads it back at volume.

What is Deep Think (and is it a model)?

Deep Think is a reasoning mode, not a separate model. It lets a Gemini model spend more compute deliberating before it answers, which lifts accuracy on the hardest maths, science and coding problems at the cost of speed. It debuted as Gemini 3 Deep Think in November 2025 and is part of the Gemini 3.5 Pro plan, generally reserved for the top subscription tier. Think of it as a higher gear you engage when correctness matters more than latency, rather than a new product to choose.

Gemini Nano: the on-device model

Gemini Nano runs directly on supported phones and edge hardware rather than in the cloud. Because the data never leaves the device, Nano is the privacy-first option: it can summarise a private message, suggest a smart reply or transcribe audio without sending anything anywhere, and it works offline with no per-request cost. The trade-off is capability: Nano is deliberately compact, so it is not the model for heavy coding or long-context reasoning. Privacy is more than a nicety here, as most consumers say they will not deal with a company they do not trust with their data (Cisco).

Nano vs Nano Banana: the image family explained

This is the single most confusing naming in the whole family, so let us be blunt: Nano and Nano Banana are completely different things. Nano is the on-device text model above. Nano Banana is Google's image generation and editing line, and it runs in the cloud.

• Nano Banana (officially Gemini 2.5 Flash Image) is the original, fast image model.
• Nano Banana Pro (Gemini 3 Pro Image) is built on the Gemini 3 Pro reasoning backbone and behaves like an art director, reasoning about layout and physics before rendering. It is the best in the family at rendering correct, legible text inside an image (DeepMind).
• Nano Banana 2 (Gemini 3.1 Flash Image) delivers near-Pro quality at Flash speed, with resolutions up to 4K, character consistency across multiple subjects, and image-search grounding for real-world accuracy.

So if you want pictures, you want Nano Banana. If you want a private, offline text feature on a phone, you want Nano. The shared word is a coincidence of branding, not a shared purpose.

The wider family: Veo, Gemma and Antigravity

Gemini sits inside a larger stack worth knowing by name. Veo is Google's video generation model. Gemma (now Gemma 4) is the open-weights family you can run yourself, useful when you need control, on-premise hosting or fine-tuning. And Google Antigravity is the agentic development environment that puts these models to work carrying out multi-step tasks. You will not choose between these and Gemini so much as combine them: Gemini to reason, Nano Banana for images, Veo for video, Gemma when you need to self-host.

What does a 1M vs 2M context window actually mean?

Token counts are abstract. Here is what they translate to in plain reading. Tap a model to see it.

Estimate your monthly Gemini cost

Enter your expected monthly usage to compare indicative costs across models. Move the inputs and the table updates live.

Gemini models compared at a glance

How the current family stacks up across the factors that decide which model fits a job. Tap any column heading to sort, or filter by where the model runs.

How Gemini compares to other AI models

In 2026 the frontier is crowded and close, so the honest answer is task-led rather than tribal. Gemini's clearest edges are context length (up to a target 2M tokens) and native multimodal and agentic breadth, which make it strong for long documents, mixed media and tool use. Competing frontier models from the likes of Anthropic and OpenAI trade blows on pure reasoning, code quality and price, and often lead on specific benchmarks at any given moment. The practical approach for a business is to choose per use case, keep your prompts and integrations portable, and re-test every few months as new versions ship, because the lead changes hands often. If you want help running that evaluation, that is exactly the kind of work we do.

Quick glossary

How to choose a Gemini model and control cost

The principle is simple: buy quality first, then optimise for cost. Prototype on the best model you might need, lock in a quality bar, then move traffic down to cheaper tiers only while the result still clears that bar. Done deliberately, model selection becomes a lever for both performance and margin rather than an afterthought, which is how we approach it when we build AI automation and AI chatbots for clients.

1. 1

   Prototype on the strongest model

   Build your first version on the most capable model you might need (Gemini 3.1 Pro today, or 3.5 Pro when it ships). This shows you the best the model can do and lets you lock in a quality bar to measure against.

2. 2

   Define your quality bar

   Write down what a good answer looks like for your use case, with a handful of representative test cases. This becomes the objective check every cheaper model has to pass.

3. 3

   Move steady traffic to 3.5 Flash

   Shift your routine, high-volume requests onto Gemini 3.5 Flash, which is fast and far cheaper per token, and confirm the output still clears your quality bar.

4. 4

   Use Flash-Lite or Nano for bulk and on-device work

   Route simple, high-volume tasks like classification or tagging to Gemini 3.1 Flash-Lite, and push private or offline features to Gemini Nano on-device where there is no per-request cost.

5. 5

   Re-test as new versions ship

   AI pricing and capability move quickly, so re-run your test cases every few months. Model selection then becomes a lever for both performance and margin rather than a one-off decision.

Sources

• Google: Gemini 3.5 announcement
• Google AI for Developers: Models
• Google DeepMind: Nano Banana Pro (Gemini 3 Pro Image)
• Wikipedia: Gemini (language model), release history

Picking the model is half the job. Structuring the request is the other half, and the PARTS framework is a good place to start if your course teaches that framework.