Google’s annual I/O wrapped on May 20 with a bunch of new announcements. The product naming is a mess. Here’s a map of Google’s AI products, split into sensible groups.
The chat models
These are the text/reasoning models you’d actually pick between. Prices are Gemini Developer API, paid tier, standard rate, per million tokens.
| Name | What it is | Use it for | Input | Output |
|---|---|---|---|---|
| Gemini 3.5 Flash | The I/O headline model | Default chat/search, agents, coding | $1.50 | $9.00 |
| Gemini 3.1 Pro (preview) | The older big model | Hard reasoning, long context | $2.00 | $12.00 |
| Gemini 3 Flash (preview) | Cheaper fast model | Bulk tasks below 3.5 Flash | $0.50 | $3.00 |
| Gemini 3.1 Flash-Lite | Cheapest Gemini 3 model | Extraction, tagging, translation | $0.25 | $1.50 |
Gemini 3.1 Pro is tiered: above 200k tokens of input it jumps to $4.00 input and $18.00 output. Gemini 3.5 Flash is the model now running in the Gemini app and Search by default. Google’s pitch is that it beats the older 3.1 Pro on coding tests while running about 4x faster.
If you’re starting a project today: Flash-Lite for high-volume work where mistakes are cheap, 3.5 Flash as the everyday default, 3.1 Pro only when being wrong costs more than the model bill. The “3.5 Pro” you might be waiting for isn’t out yet.
How that compares to Anthropic and OpenAI
Same units, per million tokens, standard rate, checked May 22.
| Model | Input | Output |
|---|---|---|
| GPT-5.4 Nano | $0.20 | $1.25 |
| Gemini 3.1 Flash-Lite | $0.25 | $1.50 |
| Claude Haiku 4.5 | $1.00 | $5.00 |
| Gemini 3.5 Flash | $1.50 | $9.00 |
| Gemini 3.1 Pro | $2.00 | $12.00 |
| GPT-5.4 | $2.50 | $15.00 |
| Claude Sonnet 4.6 | $3.00 | $15.00 |
| Claude Opus 4.7 | $5.00 | $25.00 |
| GPT-5.5 | $5.00 | $30.00 |
Google’s “half to a third the price of frontier models” line mostly holds, but only against the flagships. Gemini 3.5 Flash at $1.50/$9 undercuts Sonnet 4.6 and sits below GPT-5.4, and it’s well under the two $5-input flagships (Opus 4.7 and GPT-5.5). The comparison is less flattering against the budget tiers: GPT-5.4 Nano is $0.20/$1.25 and Claude Haiku 4.5 is $1.00/$5.00, both cheaper than 3.5 Flash. Pick the cheapest model that passes your own testing, not the one with the best launch slide.
The agents
| Name | What it is | Where to get it |
|---|---|---|
| Gemini Spark | Personal agent that acts across your Google apps | Beta, Ultra subscribers (US), rolling out |
| Antigravity | Coding app you work inside (now 2.0) | Free app, higher limits on Pro/Ultra |
| Jules | Coding agent you hand GitHub tasks to | Higher limits on Pro/Ultra |
| Managed Agents | Spin up an agent with one API call | Gemini API |
The naming gets annoying here. Gemini is the assistant and the model family. Spark is the agent inside Gemini, built on the Antigravity platform and running on Gemini 3.5. Ultra is the plan that gets Spark first. None of those names tell you what the thing does.
Spark runs in the background “even while [your devices are] turned off,” checks with you before major actions, and is rolling out to trusted testers first, then Ultra subscribers in the US.
Antigravity and Jules are both coding agents but the split is where they run. Antigravity is a desktop app you open and code inside, like VS Code but built around agents instead of around you typing. Your code stays on your machine and you work alongside the agent in real time: it suggests, you accept or reject, you iterate. Jules is the opposite. It’s a cloud service you hand a GitHub task to and then close the tab. It clones your repo into a Google VM, does the work there, and sends back a pull request for you to review. You never touch its environment.
So: Antigravity for hands-on work where you want to watch and steer, Jules for fire-and-forget tasks you’d rather not do yourself (“fix the flaky test, update that dependency”). They bill through the same subscription, so one Pro plan raises your limits on both. Antigravity 2.0 also got a CLI that absorbs the old Gemini CLI, plus an SDK and multi-agent orchestration. If you’ve been using Gemini CLI, Google wants you on the Antigravity CLI now.
The creative models
| Name | What it is | Price |
|---|---|---|
| Gemini Omni | Video gen and editing model | In app/Flow/Shorts; Omni Flash for subscribers |
| Veo 3.1 | Video model | $0.05 to $0.60/sec by tier and resolution |
| Imagen 4 | Image model | $0.02 fast, $0.04 standard, $0.06 ultra per image |
| Nano Banana | Native image gen/editing | In Gemini/Flow limits |
| Lyria 3 | Music model | $0.04 per 30s clip, $0.08 per full song |
| Flow | Creative studio | Included in paid plans, with credits |
| Project Genie | Type a place, walk through it in 3D | Ultra $200 benefit |
Gemini Omni is the new flagship here. Google describes it as a model that creates “anything from any input,” starting with video, with a better grasp of physics like gravity and how fluids move. In practice it’s video generation and editing you drive with plain language: upload a clip, tell it to change the background or add yourself to the scene. The consumer version is Gemini Omni Flash, rolling out to all paid subscribers in the Gemini app and Flow, and free in YouTube Shorts Remix.
Project Genie is the odd one out and worth explaining since the label “interactive world model” tells you nothing. You type a description of a place (“a hilly landscape of Hobbit homes with little gardens”), and it generates a 3D world you can actually move through in real time, walking or flying, rather than a fixed video you watch. It’s an experimental prototype, US Ultra-only, capped at about a minute per generation. Think tech demo, not product.
Veo pricing is per second. Veo 3.1 standard is $0.40/sec at 720p or 1080p and $0.60/sec at 4K. Fast is $0.10/sec at 720p, and Lite is about $0.05/sec for rough drafts. An 8-second clip runs $3.20 on standard or $0.80 on fast 720p. For images, Imagen 4 Fast is $0.02 and Google claims better text rendering, which is the polite way of saying it might finally spell.
The image model naming deserves its own note. The API lists it as Nano Banana, the consumer pages call it Nano Banana 2 and Nano Banana Pro in different spots, and there’s a separate Google Pics editing tool built on it. I can’t tell you which name wins. They couldn’t either, apparently.
The subscriptions
Google restructured the plans at I/O. Two changes matter: there are now two Ultra tiers, and the old $250 Ultra dropped to $200.
| Free | Plus | Pro | Ultra $100 | Ultra $200 | |
|---|---|---|---|---|---|
| Price/mo | $0 | $7.99 | $19.99 | $100 | $200 |
| Usage limits | Base | 2x | 4x | 5x Pro | 20x Pro |
| Storage | 15 GB | 200 GB | 5 TB | 20 TB | 30 TB |
| Flow credits | Limited | 200 | 1,000 | 10,000 | 25,000 |
| Cloud credits | None | None | $10 | $40 | $100 |
| YouTube Premium | No | No | Lite | Full | Full |
| Gemini Spark | No | No | No | Yes (US) | Yes (US) |
| Project Genie | No | No | No | No | Yes |
| Deep Think | No | No | No | No | Yes |
| Good for | Casual use | More headroom | Most builders | Daily builders | Power users |
The two Ultra tiers are the same plan at different volumes: the $100 one is aimed at developers who keep hitting Pro’s ceiling, and the $200 one (formerly $250) adds the experimental stuff like Project Genie and Deep Think, a mode that lets the model think longer on hard problems before it answers. Most people who pay will land on Pro.
Gotchas
A few things worth knowing before you subscribe or build.
Personal accounts only. You cannot subscribe to a Google AI plan with a Workspace account. Google’s FAQ is explicit: Workspace customers get a Gemini add-on to their existing subscription instead. If your main account is a Workspace email, you’re either signing up with a personal Gmail or going the add-on route.
Limits are now compute-based, not prompt-based. Google dropped daily prompt counts in favor of a “compute-used” model. A simple text prompt burns far less than a video or coding prompt. Your allowance refreshes every five hours up to a weekly cap. When you hit the ceiling on the big models, you get bumped to the smaller fast ones unless you buy pay-as-you-go top-up credits.
Spark is US-only and barely out. Gemini Spark is the headline agent but it’s beta, trusted-testers-first, US-only, and Ultra-gated. Don’t subscribe to Ultra for Spark expecting it to be there on day one.
Region and age limits are everywhere. A lot of features are tagged “US only” (Daily Brief, AI Inbox, Chrome auto browse, Personal Intelligence) or “18+” (most generative video). The headline list and the list you can actually use in your country are different lists.
The model on the box may not be the model you get. The consumer “Pro model” in the app is Gemini 3.1 Pro, not the 3.5 family. The 3.5 line so far is Flash only. If you’re paying for Pro-tier reasoning, check which model is actually serving your requests.
What to use, by job
| If you want to… | Use | Why |
|---|---|---|
| Chat or search day to day | Whatever the Gemini app gives you (3.5 Flash) | It’s the default and it’s fast |
| Write, summarize, or answer at high volume | Flash-Lite | Cheapest, fine when mistakes are cheap |
| Do hard reasoning or feed in a huge document | 3.1 Pro | Test it against Claude and GPT on your real task first |
| Code interactively, watching the agent | Antigravity | Desktop app, code stays local, you steer |
| Offload routine GitHub tasks | Jules | Fire-and-forget, sends back a pull request |
| Generate or edit video | Veo (lock the prompt on fast/lite, then spend on standard) | Per-second billing adds up quickly |
| Generate images | Imagen 4 Fast, or Nano Banana for editing | Fast is $0.02/image |
| Run an agent across your Google apps | Spark | Worth it if your life already lives in Gmail/Calendar/Drive |
Two rules cover most of it: pick the cheapest model that passes your own testing, not the one with the best launch slide, and don’t pay for a tier to get one feature that’s still in beta.
Sources: Google’s 100 announcements, Gemini API pricing, the AI subscriptions post, and the Google One plans page. Anthropic and OpenAI prices from their published rate cards, checked May 22.