A generated title card reading 'Gemini 4 Neon Is Adapted From Carbon' with the subtitle 'One sentence, repeated seven times, and never defined', beside a family-tree diagram: Gemini 4 Argon above Gemini 4 Carbon, labelled 'based on', branching down to Gemini 4 Hydrogen and Gemini 4 Neon, both labelled 'adapted from'.
Guides & Insights

Gemini 4 Neon Is Adapted From Carbon: What the Card Won't Define

Author

Elias Hawthorne

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

The published model card for Gemini 4​ Neon says exactly one thing about where the model came from, and it says it seven times: “Gemini 4​ Neon is adapted from Gemini 4 Carbon.” That is the entire published account of the relationship between Gemini 4​ Neon, Gemini 4 Carbon, Gemini 4 Argon and Gemini 4​ Hydrogen. No parameter count, no statement of what was adapted, no definition of the word. Six weeks of Gemini 4 releases hang off that single undefined clause, and it is now being read on X and in launch summaries as though “adapted” were a term of art with a settled meaning. It is not one Google has settled.

What follows separates the three things that are easy to blur together here: what the card literally prints, what that implies, and what is currently guesswork. Anything quoted is from a Google DeepMind page fetched on 19 and 20 November 2026; anything inferred is labelled as inference and kept in its own paragraph, so you can drop it without losing the facts.

Seven identical sentences

The card runs under the headings Model Information, Model Data, Implementation and Sustainability, Distribution, Evaluation, Intended Usage and Limitations, and Safety & Security. In seven of those slots the body text is the same clause, followed by a pointer to the Carbon card:

• Model dependencies — “Gemini 4​ Neon is adapted from Gemini 4 Carbon.”

• Architecture — the same sentence, then “see the Gemini 4 Carbon model card.”

• Training Dataset — the same sentence, then the same pointer to Carbon.

• Training Data Processing — the same sentence, then the same pointer.

• Hardware — the same sentence, then the same pointer.

• Software — the same sentence, then the same pointer.

• Acceptable Usage — the same sentence, then the same pointer.

The card is never silent about lineage. It is also never specific. “Adapted from” carries a pointer instead of content: everything a reader would want — parameter count, which layers changed, whether a smaller model was trained against a larger one, what data was new — is delegated to a document about a different model, and that document answers none of it either. The Carbon card says Carbon is “based on Gemini 4 Argon” and then points at Argon, which says Argon is “a large scale sparse mixture-of-experts (MoE) transformer-based model” and “is not a modification or a fine-tune of a prior model.” Following the chain end to end tells you that Neon descends from something that descends from something that is new. It does not tell you what “descends” did.

Two phrases, no stated difference

The more interesting comparison is not Neon against its parent. It is the vocabulary Google uses across the family, because the same publisher uses two different phrases for two different relationships and never explains why:

• Gemini 4 Argon, card published 13 October 2026 — “not a modification or a fine-tune of a prior model.” Used once, in Model dependencies.

• Gemini 4 Carbon, card published 10 November 2026 — “based on Gemini 4 Argon.” Used six times, in Model dependencies, Architecture, Training Dataset, Hardware, Software and Acceptable Usage.

• Gemini 4​ Hydrogen, card published 12 November 2026 — “adapted from Gemini 4 Carbon.” Used seven times, in the same seven places Neon uses it.

• Gemini 4​ Neon, card published 18 November 2026 — “adapted from Gemini 4 Carbon.” Seven times.

A generated four-column comparison card headed 'The Gemini 4 family — four cards, four rows', counting one lineage phrase across the published model cards: Argon published 13 October 2026, 'not a modification', stated once; Carbon 10 November 2026, 'based on Argon', six times; Hydrogen 12 November 2026, 'adapted from Carbon', seven times; Neon 18 November 2026, same phrase, seven times; every card listing a March 2026 knowledge cutoff.

So “adapted from” is not a phrase coined for Neon. Hydrogen carried it six days earlier, and the two cards use it identically — same wording, same slots, same parent. Whatever adaptation means here, it is the standard relationship for a model built on Carbon, and it is visibly distinguished in the text from “based on,” which is Carbon's relationship to Argon. Neither phrase is defined anywhere in the family's documentation. A reader who assumes “adapted from” means something weaker or narrower than “based on” is assuming; a reader who assumes the two mean the same thing is assuming just as hard in the other direction.

What the architecture section says

This is the part the circulating summaries skip. The Neon card contains no occurrence of “distill,” “distillation,” “prune,” “quantize” or “mixture-of-experts,” no parameter count, and the word “smaller” does not appear. The architecture section is one sentence and a pointer to Carbon. Nor does the card ever use “post-trained” — the exact word the X thread that prompted this piece noticed was missing. That absence is real and checkable.

Screenshot of the Google DeepMind model card page for Gemini 4 Neon, showing the published date of 18 November 2026, the Model Information tabs, and the Model dependencies line reading 'Gemini 4 Neon is adapted from Gemini 4 Carbon.'

The one place the card names training techniques is the Safety and Security section, in a sentence beginning “Safety and security were built into Gemini 4​ Neon throughout the training and deployment lifecycle, including pre-training data filtering, post-training alignment (supervised fine-tuning and reinforcement learning)…” Read quickly, that looks like the disclosure the rest of the card withholds. It is not. The identical sentence, word for word, appears on the Carbon card and on the Hydrogen card. The Argon card does not carry it at all. It is a safety-process template applied to every model built on Carbon, and using it as evidence that Neon was reinforcement-learning post-trained is reading boilerplate as a training disclosure. The card does not say Neon was post-trained. It does not say it wasn't.

The cutoff is the one dated fact

The card does give one hard, checkable fact about provenance, and it is easy to miss because it sits under Known Limitations rather than anywhere near the architecture paragraph: “The knowledge cutoff for Gemini 4​ Neon is March 2026, though in some domains the model's knowledge may be less up to date.”

• Gemini 4 Argon — March 2026 (the card states it for the family: “The knowledge cutoff for Gemini 4 is March 2026”).

• Gemini 4 Carbon — March 2026.

• Gemini 4​ Hydrogen — March 2026.

• Gemini 4​ Neon — March 2026.

Four models across five weeks, four identical cutoffs. Whatever was added between Argon in October and Neon in November, it was not a new slice of the world; any capability gain comes from method, data selection or training compute on material the family already had. That is a genuine constraint on the “they trained a fresh model on top” reading, and it is the only quantitative statement about the model's construction the card makes.

The MCC finding nobody quoted

Buried in the Frontier Safety Assessment section is a comparison Google ran and did not put in the launch materials. The company assessed whether Neon showed “meaningful new capabilities or material increases in performance relative to the most recent model version subject to a critical capability assessment, which was Gemini 4 Argon.” Its conclusion: “Gemini 4​ Neon showed an increase in general capability under our MCC assessment compared with Gemini 4 Argon, however this falls below our MCC threshold.”

Read that as a capability statement rather than a safety one. Google's own internal assessment places Neon above Argon, and the card then carries Argon's capability evaluations and threshold determinations forward instead of re-running them, on the grounds that nothing material changed. The 61 on the Intelligence Index that the card reports is therefore not just a vendor number in a table — it is consistent with an internal finding that the derivative outperforms the model it descends from.

The same section is where the two “adapted from” siblings stop being interchangeable. Hydrogen's card says the opposite of Neon's: “Gemini 4​ Hydrogen did not show a material increase in general capability under our MCC assessment compared with Gemini 4 Argon.” Same parent, same phrase, opposite finding. So “adapted from Gemini 4 Carbon” is not a strength label. It describes a family of derivatives, at least one of which added measurable capability and at least one of which did not.

Which produces an awkward price table. Through 30 November 2026, Gemini 4​ Neon costs $3.00 per million input tokens and $15.00 per million output tokens, while Gemini 4 Argon — the model Neon is measured against and, internally, beats — costs $2.00 and $10.00. The derivative is 50% dearer per token than the ancestor whose evaluations it inherits. From 1 December, when Carbon and Argon step to $4.00/$20.00, that inverts and Neon becomes 25% cheaper than both. Neither figure says anything about model size: per-token price is a commercial decision, and Google has never priced this family by parameter count.

Why the price cuts against Matryoshka

The most specific hypothesis in circulation is that the family is nested — that Carbon contains Neon the way a Matryoshka doll contains a smaller one, and that Neon was obtained by distilling the larger model into a cheaper one, possibly with extra reinforcement learning on top. It is a reasonable guess. It explains why a derivative would share a knowledge cutoff, why the card would defer architecture to the parent, and why Google would avoid “post-trained” for a model that, on that reading, was not post-trained so much as extracted.

Three things in the public record argue against the simple version. The first is price: a distilled, smaller model is normally cheaper per token than its teacher, and Neon launched at 50% above Carbon's introductory rate. A cheaper-to-serve model sold at a premium is possible — that is what a faster tier is — but it removes the strongest piece of circumstantial evidence for the distillation reading. The second is Hydrogen: the cheapest model in the family, $1.50/$7.50 with the only free tier, carries the identical “adapted from Gemini 4 Carbon” wording in the same seven slots. If the phrase described one specific technique, two models at opposite ends of the price ladder would not both carry it. The third is that the card itself, when it wants to talk about efficiency, does it in the Description — “Neon balances depth and efficiency” — and not in the architecture section, which is empty of content.

The honest position is that the card's language distinguishes three relationships — new model, based on, adapted from — without defining any of them, and that the Matryoshka hypothesis is a plausible reading of an underspecified document, not a finding. A nest-of-dolls model usually shows a parameter count and a serving-size story. This card shows neither, and the only size-shaped claim on it is a positioning sentence.

Four cards, three comparison traps

Anyone assembling a “Neon versus Carbon” claim out of these documents hits three separate mismatches, and each one is easy to walk into.

• Index version. The Neon card reports “AA Index v5.0.1: 61,” against Claude Opus 5.5 at 59 and GPT-6.2 Sol at 57, results dated November 2026. The Carbon card, published three weeks earlier, reports “AA Index v5.0: 59,” against Opus 5.5 at 58 and Gemini 4 Argon at 50, results dated October 2026. The suffix changed between the two cards, so a two-point gap measured on different index versions is not a two-point gap.

• Thinking level. The 61 is a ceiling, not a default. Artificial Analysis, whose index Google is citing, publishes four Neon scores on v5.0.1: 44.1 at low, 54.9 at medium, 59.6 at high — the level Neon actually defaults to — and 61.4 at xhigh. Carbon sits at 60.0. So the model's default configuration scores below its parent while its maximum configuration scores above it, and the vendor card prints only the maximum. Both sets of figures are the tracker's, not Google's, and both were current on 20 November 2026.

• Benchmark version and harness. Neon's card lists Terminal-bench 4.1 at 65.2% cheating-adjusted, against Opus 5.5 at 65.8% — a loss. Carbon's October card lists “Terminal-Bench 4.0 (AA)” at 65.2%. Two cards, the same number, different benchmark versions, which is a coincidence and not a comparison. The SWE-sweep rows are worse: Neon lists 39.0% on a “provider harness” and 12.8% on a “Mini-SWE harness,” while Carbon lists a single 9.6% with no harness qualifier. Those are three configurations and cannot be lined up as a 29-point improvement.

Screenshot of the Artificial Analysis page for Gemini 4 Neon, headed 'Intelligence, Performance & Price Analysis', showing ranks 46 of 96 and 73 of 95, 110 output tokens per second, $3.00 input and $15.00 output per million tokens, an intelligence index of 61 at xhigh effort, and the note that the model is slow per task and very verbose.

Everything else on the Neon card is vendor-reported and unreproduced: CursorBench 4.1 at 65.4% (Opus 5.5 58.1%), FrontierCode 2.0 at 46.2% (47.3%, a loss), GraphWalks 256k-to-1M at 95.2% (66.8%), AutomationBench 56.3% (42.5%), SEC-bench Pro 90.5%, RiemannBench 84.0%. Two of the coding numbers are losses to a competitor, which is worth noting because launch summaries quote the wins and skip those.

What would settle it

Four things would move “adapted” from inference to fact. None exists yet.

• A parameter count or serving-hardware disclosure for Neon. The Hardware section is a pointer to Carbon; a real figure would immediately separate a pruned or distilled derivative from a re-tuned sibling.

• The Gemini 4​ Neon Safety & Security Report, which the card says “will be available in the coming days.” Safety reports have historically carried capability detail the public card omits.

• Token-efficiency data at matched quality. Google's launch documentation claims Neon delivers “leaner token usage” and the card quantifies nothing. If Neon really emits fewer tokens per task than Carbon, the lineage question matters less than the invoice question.

• Any statement from Google defining “adapted from” as against “based on.” Until then the two phrases are a distinction the publisher has drawn and declined to explain.

How to use an underspecified model

None of this is a reason to avoid the model. It is a reason to evaluate it rather than infer it from the card. Neon is paid-tier only — no free quota — with four thinking levels (low, medium, high as default, xhigh), a unified 1,048,576-token window with 64K synchronous output and long decode to 1M, implicit caching only, batch and flex at half rate ($1.50/$7.50), and priority at 1.8× ($5.40/$27.00; cached input $0.27 there against $0.15 standard). Audio generation, image generation and the Live API are marked unsupported; code execution, computer use, function calling, structured outputs, file search, URL context and both grounding options are supported.

The practical consequence of an underspecified lineage is that you should measure the thing the card withholds — output tokens per completed task — at the setting you would actually deploy. Run the same 20 or 30 tasks through Neon at high and through Carbon at its default medium, count tokens rather than tokens-per-dollar, and compare against your own bill. A model that is 50% dearer per token and 30% leaner per task is cheaper; one that is 50% dearer and equally verbose is not, and no card in this family will tell you which you have.

Our own catalogue is a partial answer to the second half of that problem. It carries “google/gemini-4-carbon” and “google/gemini-4-argon”, and it does not carry Gemini 4​ Neon — the model page returns nothing today, so nothing here should be read as an offer to route it. What the router does give you while you evaluate is a single key and a single bill across the models you are comparing, provider list price passed through with no markup, so the 1 December step on Carbon and Argon ($2.00 to $4.00 input, $10.00 to $20.00 output) lands in your invoice the same day it lands on Google's pricing page instead of being absorbed into a reseller margin. And because a fallback chain can hold Carbon behind Neon, an unproven derivative can carry production traffic without being the only thing standing between your users and an outage.

The open question is not whether “adapted from” is honest — it plainly is, as far as it goes. It is whether a card that repeats one undefined clause seven times, ships without a parameter count, and defers architecture, data, hardware and software to a sibling document is enough disclosure for a model that costs more per token than the one it descends from. On the current evidence Google has told us the shape of the family tree and nothing about the branches. The Safety & Security Report, whenever it lands, is the next document worth reading.