Gemini 4 Split into Three Models — Carbon, Neon, Hydrogen, and What Changed Across the Generations
The result first
Gemini 4 is no longer one model. It is three, divided by how heavy the job is.
- Carbon — for the hardest, longest work. The one that thinks slowly and thoroughly, and keeps a long task moving.
- Neon — the middle: depth and efficiency together. Announced on November 19, 2026 (Japan time), it is the center of this generation.
- Hydrogen — the fast everyday side. The free tier of Gemini runs on this one — including the free “Tasks” feature I wrote about in another article today.
The prices differ too. Per million tokens (a token is roughly the unit AI uses to count words), input/output: Carbon starts at $2 / $10 as an introductory price, Neon is $3 / $15, and Hydrogen is $1.50 / $7.50. One honest note before anything else: the benchmark rankings quoted for these models are numbers Google published itself. They have not been verified by third parties yet, and this article keeps that line visible.
Below: the three models one by one, then the whole family history — Gemini 1.0 through 4 — with what changed at each generation: how much it can hold, what it can do, and how the way we hand it work has changed.
I am Kurashiki Bouzu (kurashiki-bouzu), born in 1971, a former funeral director now living in an old house in Kurashiki, Okayama, writing articles every day by talking with Moka (Muse), my AI butler.
Carbon — the hardest, longest work
Carbon is the top of Gemini 4. It was announced on October 28, 2026, and became available on November 11 (Japan time). Google describes it as the model for cybersecurity, complex knowledge work, and long-running tasks.
There is also a defense specialist version, Carbon for Cyber, built to find weaknesses in software and help fix them. It was first handed to security defenders through a framework called the Fairwind Program. When today’s security article mentions a defensive AI, this is the one it means.
The API price is $2 per million input tokens (about ¥320) and $10 per million output tokens (about ¥1,600) — an introductory price. From December 1, it rises to $4 (about ¥640) and $20 (about ¥3,200). A million tokens is a lot: several paperback books’ worth of text.
One more note: Neon also has a defensive version, Neon for Cyber, which is likewise being rolled out gradually via Fairwind, as announced.
Neon — depth and efficiency, in the middle
Neon, announced on November 19, 2026 (Japan time), is the core model of Gemini 4, adapted from Carbon. It is aimed at complex coding, analysis, and business systems.
The price is $3 per million input tokens (about ¥480) and $15 per million output tokens (about ¥2,400). Cached input — reusing content it has already read — costs $0.15 (about ¥24), 95% cheaper than regular input. Its context window is one million tokens, and it can take spreadsheets, PDFs, charts, audio, and video together.
Google’s announcement cites rankings: 61 points on the Artificial Analysis Intelligence Index, third overall, and first place on a coding-agent index, among others. Here is the honest line again: these are figures Google reported. They are not third-party verified. The source article for this piece (Business+IT, November 21) states that explicitly, and so do I.
One more honest detail. You may see the claim “32% cheaper per task than Carbon.” That is a per-task claim, not a per-token one. Per token, Carbon’s introductory price is actually lower during November; Neon becomes cheaper per token only after Carbon moves to its regular price on December 1. Small print, but it changes how the headline should be read.
Hydrogen — the fast everyday side
Hydrogen, announced on November 12, 2026, is the speed-and-efficiency model: $1.50 per million input tokens (about ¥240) and $7.50 per million output tokens (about ¥1,200), with the same one-million-token context window.
What matters most is that Hydrogen has a free tier. The free version of Gemini’s “Tasks” feature — the one that became free in Japan and five other countries — runs on Hydrogen. The generational change that looks distant in a headline has already arrived on the free side.
The family history — how Gemini changed
Gemini 1.0 (December 6, 2023) — everything included, from the start
The first Gemini came in three sizes: Ultra, Pro, and Nano. Its core idea: designed from the beginning to handle text, images, audio, video, and code together — not bolted on later, but built in from day one.
Gemini 1.5 (February 15, 2024) — the amount it could hold jumped
With 1.5 Pro, the context window reached one million tokens for the first time: a whole book, or an hour of video, in one go. This became Gemini’s signature strength. (Shortly before, on February 8, 2024, the chat AI formerly called Bard was renamed Gemini.)
Gemini 2.0 (December 11, 2024) — from answering to acting
Gemini 2.0 arrived under the banner of “the agentic era.” An agent is an AI that does not just answer, but uses tools and carries work forward — searching, running code. The way work was handed over changed: from an answering partner to a working partner.
Gemini 2.5 (March 25, 2025) — think first, then answer
With 2.5, “thinking” moved to the center: the model reasons before it answers, and you can choose how deep it thinks. The thinking levels in today’s models were built here.
Gemini 3 (November 18, 2025) — everything in one, into Search on day one
Gemini 3 combined reasoning, tool use, and multimodal understanding into one generation, and entered Google Search on the very day it was announced. Its selling point was reading the room — grasping context and intent.
Gemini 4 (from September 30, 2026) — split by role
And now: element names, and models split by role. Argon came first (September 30), then Carbon (October 28), Hydrogen (November 12), and Neon (November 19, Japan time) — the lineup completed within about two months. Google’s own model-choosing guide uses these three — Carbon, Neon, Hydrogen — as its axis.
What changed, in three lines
- How much it can hold: one million tokens arrived with 1.5, and is now standard across all three Gemini 4 models. Handing over long documents whole is normal.
- What it can do: from reading text, to handling images, audio, and video; to using tools; to thinking before answering; to carrying long jobs through.
- How we hand it work: from a partner that answers questions, to one you can hand whole documents, to one that does the work — and now, one you choose by the weight of the job.
In my own project “Using Gemini to the Fullest,” the free-tier limit appeared on the eighth task of the second stage. The model I was touching there was the everyday, fast side of the family. I have not yet tried splitting real work between Carbon and Neon at home — that, too, belongs on the honest side of the ledger.
Cost and sources
Writing this article cost 0 yen. I read the announcements and official explanations, and matched them against my own records.
- Neon’s announcement date, prices, context size, and rankings: Business+IT, “Google announces Gemini 4 Neon” (2026-11-21, https://www.sbbit.jp/article/cont1/188650). The article itself notes the rankings are Google-reported and not third-party verified.
- Carbon’s announcement, Carbon for Cyber, and introductory pricing: Google’s official blog, “Gemini 4 Carbon is here” (2026-11-10; November 11 Japan time), and PC Watch’s report of the October 28, 2026 announcement.
- Hydrogen’s announcement, prices, and free tier: Google’s official blog, “Introducing Gemini 4 Hydrogen” (2026-11-12), and ITmedia NEWS (2026-11-13).
- Generational dates and changes: the model list on Wikipedia’s “Gemini (language model)” article, cross-checked with Naoky AI’s update-history series (checked November 5, 2026).
- Yen figures are converted at roughly ¥160 per dollar, matching the source article’s “3 dollars (about ¥480).”
- November 21, 2026 ― English draft (result-first rewrite of the Japanese article, not yet published)(6,445 characters)
- November 21, 2026 ― Published (all 7 languages at once, on the user's instruction). Japanese original: ledger No. 1245. New URL form with the ledger number at the end.(6,445 characters)
