New: AI API Cost Calculator — compare GPT, Claude, Gemini, Grok, Llama, DeepSeek & Kimi pricing live
Developers · News

Google's Gemini 4 Neon Is Here: Frontier Brains at $3 per Million Tokens — and Genie 4 Lets You Edit Worlds Mid-Walk

By ToolsDig Research · Updated November 19, 2026 · How we research

Nothing here is sponsored — no affiliate links on this page. Every fact below comes from Google's announcements of November 18, 2026 (the Gemini 4 Neon launch post and the Project Genie update), plus Google's November 17 Tasks announcement. Benchmarks and pricing are Google's own claims, not independently verified. We have not tested Neon or Genie 4 hands-on — this is researched analysis, not a review.

A glowing neon glass orb sending threads of light into a miniature vivid 3D landscape of mountains, trees and a river

Google did not wait long to answer. A day after OpenAI's GPT-6.2 Astra and Sol and Microsoft's MAI models landed on November 17, Google shipped two launches of its own on November 18: Gemini 4 Neon, a new mid-tier model aimed squarely at developers doing complex, high-frequency work — and Genie 4, now open to AI Ultra subscribers in Project Genie, the first world model that lets you rewrite the world while you are walking through it. The Neon launch is the one that matters for your API bill. The Genie launch is the one that matters for your imagination.

🔍 How we put this together

We used Google's official November 18, 2026 announcements — the Gemini 4 Neon launch post on the Google blog and the Project Genie update featuring Genie 4 — plus Google's November 17 announcement that Tasks in Gemini is now free in six countries. Where a figure is Google's own claim — the benchmark scores, the pricing, the capability list — we say so. Nothing here is independently verified yet, and we have not run Neon or Genie 4 ourselves.
🏆 Verdict: Neon is the most interesting Gemini 4 model for working developers — not because it tops the charts, but because it sits in the sweet spot: near-Carbon intelligence at a price and speed meant for daily, high-volume use. The honest catch is the calendar: Carbon's introductory $2/$10 pricing runs through November 30, so Neon is briefly the more expensive model per token. From December 1, the maths flips and Neon is 25% cheaper than Carbon. If your workload is long agent runs and multi-file refactors, Neon is worth a pilot this week — but run it against Carbon while Carbon is still on sale, and let your own cost-per-task numbers decide.

Where Neon sits in the Gemini 4 family

Google's Gemini 4 lineup now has a clear shape, and Neon fills the gap that was missing:

Google's pitch is that Neon delivers near-Carbon performance at roughly three times the speed and about 60% of the cost — and at its default high thinking level, a 32% lower cost per task than Carbon. It also gets four thinking levels, so you can dial intelligence and token spend up or down per workload instead of switching models.

The numbers — and what they actually mean

Benchmarks are vendor-selected, so read these as Google's best case, not gospel. Still, the pattern is consistent across independent evaluators:

There is also a security angle Google is pushing hard: Neon for cyber produces more secure code than Carbon across easy, medium and hard challenges on AutoBaxBench, hitting state-of-the-art on the hard tier (76%), and reaches 90.5% on SEC-bench Pro for finding vulnerabilities. The cyber variant rolls out to trusted defenders through the Fairwind Program over the coming days.

The price — read this part carefully

Neon costs $3 per million input tokens and $15 per million output tokens, with cached input at $0.15 per million — a 95% discount that matters a lot if your agent re-reads the same long prompts. Context window: 1 million tokens.

Now the catch nobody's headline will carry: Carbon is still on introductory pricing of $2 per million input and $10 per million output through November 30. That means Neon, the mid-tier model, is currently more expensive per token than the flagship above it. From December 1, Carbon rises to $4/$20 and Neon becomes 25% cheaper — which is the comparison Google actually wants you to make. Until then, the honest move is to benchmark both while Carbon is on sale and lock in whichever wins on cost per completed task, not cost per token. Google's own framing points that way: at its default high thinking level, Neon pairs a high Intelligence Index score with roughly 50% to 80% lower cost per task than top competing frontier models.

Where you can use it today: the paid Gemini API tier in Google AI Studio, Gemini for business (via API and model dropdown), Google Antigravity for paid Pro and Ultra subscribers — where Neon is also the fast, high-judgment Orchestrator directing parallel Hydrogen subagents — and the Gemini app for Pro and Ultra subscribers.

Genie 4: the world you can rewrite while you walk through it

The second launch is stranger and, in the long run, possibly bigger. Genie 4 is DeepMind's most realistic interactive world model — you type a prompt or drop in a photo and step inside a place you can explore with a keyboard. It was announced on October 27 as a research preview for trusted testers. From November 18, it is open to eligible Google AI Ultra subscribers in Project Genie; Google AI Pro members can join the waitlist.

Two firsts make it different from Genie 3:

Project Genie also gains a creation agent that helps you build your vision, and a community gallery to discover, explore and remix other people's worlds. Genie 3 is not going away — it stays selectable for longer runs and for worlds seeded from 360-degree panorama images — and it got upgrades of its own: sessions grew from 60 seconds to 3 minutes, with 360-degree panorama grounding. All video and audio generated in Project Genie carries SynthID, Google's invisible watermark for AI-generated content.

Google is unusually candid about the limits: camera latency, intricate physical interactions, real-time audio precision and long-term scene permanence are all still unresolved, and details can drift in longer sessions. Both Genie models remain frontier research systems. That honesty is worth noting — a world model that forgets its own geography after a few minutes is a toy, not a tool, and the 3-minute Genie 3 sessions tell you exactly where the technology stands today.

And quietly, the same week: Gemini Tasks is free

Easy to miss between the two big launches, but possibly the most widely felt change: on November 17, Google made Tasks in Gemini free in the US, Japan, South Korea, Canada, Australia and India — running on Gemini 4 Hydrogen. It is the agent formerly known as Spark: flip Gemini from Chat to Task and it works through multi-step jobs for you. Free users get standard usage limits, 10 active schedules and up to 15 tasks at a time; AI Pro adds 4x the usage and 50 schedules. Outside those six countries it still needs Pro or Ultra, and it is blocked entirely in the European Economic Area, the UK, Switzerland and Nigeria. Tasks in Gemini can now also run on Neon, which is the quiet connection between the week's three announcements: Hydrogen does the free chores, Neon does the hard ones.

Who this fits — and who should wait

Neon fits developers and teams running long agent workflows — multi-file debugging and refactoring, tool-heavy research, enterprise document work — where Carbon is overkill per task and Hydrogen is underpowered. If your agent sessions run long and your prompt cache hit rate is high, the $0.15 cached-input rate is the number to watch. Antigravity users get the most immediate benefit, with Neon as the orchestrator over Hydrogen subagents.

Wait, or pilot carefully, if: you are price-sensitive and can still grab Carbon at its $2/$10 introductory rate before November 30; your workloads are short and simple enough that Hydrogen already handles them; or you need the cyber-hardened variant, which is still rolling out to Fairwind members only.

Genie 4 fits anyone curious about where interactive media is going — game designers prototyping worlds, educators, storytellers — and anyone with an AI Ultra subscription who wants to be early. It does not yet fit production work: three-minute sessions and drifting details are research limits, not product features.

The honest limits

What to do this week

If you build with Gemini, the move is a small, measured pilot — not a wholesale model swap:

1. Run your hardest recurring agent task on Neon and Carbon side by side while Carbon is still $2/$10, and record cost per completed task — not per token. That is the number that decides which model you keep.
2. Turn on prompt caching for any long, repeated context. At $0.15 per million cached tokens, a 95% discount on input is where Neon's pricing actually wins.
3. In Antigravity, try Neon as the orchestrator over Hydrogen subagents for a multi-file refactor — that is the workflow Google built it for.
4. If you have AI Ultra, spend an evening in Project Genie with Genie 4 — change the weather mid-walk, add a character, see where it drifts. It is the fastest way to understand what world models can and cannot do yet.
5. If you are in one of the six free-Tasks countries, try Gemini Tasks on a real recurring chore. Free is the right price to find out whether an agent actually saves you time.

Sources: Google, "Introducing Gemini 4 Neon," November 18, 2026; Google, "Genie 4 is now available" (Project Genie update), November 18, 2026; Google, Tasks in Gemini announcement, November 17, 2026. Benchmark scores, pricing and capability details are Google's own claims; they have not been independently verified. Full version with FAQ on this page.