Title card reading Gemini 4 Neon: Google's 'Small' Model Has No Published Size, with the subtitle one X post, three checkable numbers, and a size nobody outside Google has seen, above three numbered rows: the post calls Neon small with no parameter count published for any Gemini 4 model; Neon scores 61.4 at xhigh and 59.6 at default high against Carbon's 60.0, a 1.4 point lead or a 0.4 point deficit; Carbon shipped 10 November and Neon 18 November, eight days, with the RSI clause citing no training data. The footer reads scores per Artificial Analysis index v5.0.1, read 20 November 2026; size claims unaudited.
Guides & Insights

Gemini 4 Neon: Google's 'Small' Model Has No Published Size

Author

Elias Hawthorne

Date Published

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

At 01:57 UTC on 20 November, a post went up on X calling G​emini 4 Neon G​oogle's small model — “far ahead of Argon and Carbon”, competitive with Claude Opus 5.5 and Claude Fable 5.2, and, in its closing four words, evidence of RSI. It was edited three times before it settled and carried 316 likes when we read it. Every other clause can be checked against a document. The word “small” cannot. Go looking for a parameter count or a size ordering across G​oogle’s documentation for G​emini 4 Neon, G​emini 4​ Carbon, G​emini 4​ Argon and G​emini 4 Hydrogen and you find nothing: no parameter count, no active-parameter figure, no sentence that ranks the four by size, and — on the DeepMind model page and the model cards alike — not one use of the word “small” in reference to any of them.

That does not make the post wrong. It makes it unfalsifiable on its own central word, which is a different and more interesting problem: the claim bundles a measurement anyone can check (the score) with a description of the model’s internals that nobody outside G​oogle can check, and then draws a conclusion about how the model was trained from the unverifiable half. Below, the claim is taken apart into its three parts and each is tested against what is published. All scores are Artificial Analysis on index v5.0.1 as read on 20 November 2026; everything Google-stated is labelled as such; anything inferred is marked as inference.

What the post actually says

Stripped of framing, there are three assertions. One: Neon is G​oogle’s small model. Two: as a small model it is “far ahead of Argon and Carbon”, i.e. a smaller model has overtaken larger siblings. Three: that outcome is what “RSI and such” looks like. The second resolves into arithmetic, the first resolves into a definition G​oogle has not published, and the third does not resolve at all — it is a claim about a training process with no disclosed input.

Worth noting before the arithmetic: this framing is not unique to one account. A day earlier a different poster attributed a Neon demo to “swarms” and to “not doing distillation and taking time with RSI”. The vocabulary is circulating faster than any document supports it, which is exactly why the individual words are worth pricing out.

G​oogle never calls Neon a small model

G​oogle does describe each of the four models, and the descriptions are specific enough to be quoted. From the G​emini API model documentation, as published this week:

• G​emini 4​ Carbon — “Our strongest reasoning model, built for complex software engineering, multimodal knowledge work, scientific reasoning, and cybersecurity defense.”

• G​emini 4 Neon — “Frontier-level intelligence with fast end-to-end task completion across complex software engineering, knowledge work, and science workflows.”

• G​emini 4​ Argon — “Our efficient G​emini 4 frontier reasoning model, delivering faster responses and lower token spend per turn compared to G​emini 4​ Carbon…”

• G​emini 4 Hydrogen — “Brings cutting-edge intelligence to everyone, with speed and efficiency for everyday coding and agentic workflows.”

Three of those four sentences are about speed and cost. None of them is about size, and no G​oogle page we could read uses the word “small” for any of the four. According to the same documentation, the model G​oogle points at for everyday, low-cost work is G​emini 3.5 Flash-Lite — which is not part of this family and is not what the post is about. So “Neon is the small one” is not a summary of G​oogle’s positioning. It is a reader’s inference from behaviour, stated as though it were a spec.

The price card argues against the plain reading too. Neon is the most expensive model in the family per token: $3.00 in and $15.00 out per million, flat, with no free tier and no expiry, against Carbon and Argon at $2.00 and $10.00 through 30 November and Hydrogen at $1.50 and $7.50 with a free tier. A small model that costs 50% more per token than the flagship it is being compared to is not small in the way buyers usually mean the word. Whatever is small about Neon, it is not the bill.

Where the numbers do back the claim

The score half of the post is checkable and mostly holds — with one substitution that changes the verdict. On the Intelligence Index at each model’s deepest published setting:

• G​emini 4 Neon — 61.4 at xhigh (59.6 at its default, high)

• G​emini 4​ Carbon — 60.0 at high, which is its ceiling; the model has no xhigh rung

• G​emini 4 Hydrogen — 54 at xhigh

• G​emini 4​ Argon — 50 at high, also its ceiling

So “far ahead of Argon” is true and not close: 11 points. “Far ahead of Carbon” is not true at all. Neon leads Carbon by 1.4 points, and only when you push it to a thinking rung Carbon does not have. At the settings both vendors recommend — Neon’s default high against Carbon’s high — Neon is 0.4 points behind. The post’s strongest form would have been the honest one: Neon is the first model in this family to score above the flagship, by about one point, if you are willing to pay for the deepest rung.

“Within weeks” is also literally true, and the arithmetic is worth stating because it is the only dated fact in the post. Carbon went generally available on 10 November. Neon went generally available on 18 November. Eight days separate the flagship and the model being described as having overtaken it.

A single-column scoreboard titled Gemini 4 Neon - the scoreboard, with six rows: Intelligence Index 61.4 at xhigh and 59.6 at default high; Carbon at its ceiling 60.0 at high; cost per index task $6.44 against Carbon's $7.76; list price $3.00 in and $15.00 out per million; output tokens on the suite 123M against Carbon's 205M; published parameter count none, for any Gemini 4 model. The footer reads scores per Artificial Analysis index v5.0.1, prices per Google's rate card, size claim unaudited.

The one comparison the post skipped

The “competes with Opus / Fable” line is the part that undersells itself, because the flattering version of it requires a number the post did not use. On the same index, Claude Opus 5.5 scores 59 at its max effort and Claude Fable 5.2 scores 65. Neon’s 61.4 therefore sits between them: ahead of Opus 5.5 by 2.4 points at Neon’s deepest setting, behind Fable 5.2 by 3.6. “Competes with” is fair for the first and generous for the second.

What makes it a real story is what those points cost, measured per completed task rather than per token:

• Cost per Intelligence Index task — G​emini 4 Neon $6.44, Claude Opus 5.5 $10.04, Claude Fable 5.2 $26.54

• List price per million, input / output — Neon $3.00 / $15.00, Opus 5.5 $4.00 / $20.00, Fable 5.2 $8.00 / $40.00

• Time per index task — Neon 21.5 minutes at xhigh and 14.8 at high; Opus 5.5 25.7 at max; Fable 5.2 41.6 at max

• Output speed — Neon 110 tokens per second, Opus 5.5 112, Fable 5.2 about 110

Four times cheaper per task than Fable 5.2 for 3.6 index points is a better argument than anything in the original post, and it does not depend on a single word about size. It is also the argument that survives the nearest objection: at Fable 5.2’s price, a 3.6-point gap is the cheapest thing on the board to close if the workload needs it, and irrelevant if it does not.

The Artificial Analysis page for Gemini 4 Neon, headed Gemini 4 Neon: Intelligence, Performance & Price Analysis, showing intelligence 61 with 110 tokens per second, $3.00 input and $15.00 output per million tokens, a 1M token context window, and the notes that it scores 61 on the Intelligence Index against a median of 44, costs $6.44 per task, and takes 21.5 minutes per task against a median of 12.4.

Small as speed, not as size

If there is a defensible reading of “small” here, it is the one the tracker measures rather than the one the model card withholds, and it is about work per task rather than parameters. At the same high effort, Neon finished the Intelligence Index suite in 123M output tokens against Carbon’s 205M, in 14.8 minutes against Carbon’s 41.8. That is a 40% smaller token footprint and a task that completes nearly three times sooner, and it is the closest thing to evidence for the post’s intuition that exists in public: whatever Neon is, it does the same job with less emitted work than the model G​oogle calls its strongest.

But that is inference from behaviour, and behaviour has other explanations — a more efficient decoder, a different thinking budget, a shorter reasoning trace — none of which require a smaller parameter count. The same numbers are also consistent with Neon simply being a newer, better-trained model of comparable size, which is the least exotic explanation available.

The speed framing also breaks somewhere the post does not look. Neon is not the fastest model in the family. G​emini 4 Hydrogen runs at 167 tokens per second to Neon’s 110, and finishes an index task in 15.4 minutes at xhigh — within a whisker of Neon’s 14.8 at high, at half the price. If “small and fast” is the claim, Hydrogen is the model it describes, and Hydrogen scores 54.

An infographic titled Small as speed, not as size, with six rows: output tokens for one index suite 123M on Neon's default and 205M on Carbon; time per index task 14.8 minutes on Neon's default and 41.8 on Carbon; output speed 110 tokens per second on Neon and 167 on Hydrogen; score at the deepest setting Neon 61.4 and Hydrogen 54; fastest model in the family Hydrogen at half Neon's price; parameter count behind any of it unpublished. The footer reads tokens, time, speed and scores per Artificial Analysis index v5.0.1, read 20 November 2026; prices per Google's rate card; size claims unaudited.

Why 'RSI' is doing the work here

Recursive self-improvement, in the sense the phrase has in AI research, means a lab using a model to measurably improve its successor — a loop with a model inside it, not merely a fast release cadence. Nothing published supports that reading of this release, and the post offers nothing to support it: no compute figure, no disclosure of what Neon was trained on, no evidence that any G​emini model contributed to Neon’s training.

What G​oogle does publish about lineage is one clause, repeated. The model cards for both G​emini 4 Neon and G​emini 4 Hydrogen state that the model is “adapted from G​emini 4​ Carbon” — the same sentence in the same seven slots on both cards, with no definition of “adapted”, no parameter count, and no account of what changed. Following Carbon’s card back leads to Argon, whose card says it “is not a modification or a fine-tune of a prior model”. Every G​emini 4 card gives the same knowledge cutoff, March 2026, across five weeks of releases. Four models, one data pipeline, one undefined word: that is a description of a shared training base, which is ordinary, not of a loop.

The eight-day gap between Carbon and Neon is a schedule. Schedules compress for reasons that have nothing to do with self-improvement — a shared base model, staged rollout, a launch calendar — and the observation that this one compressed is not evidence for any of them. If a lab did build a model with a model, the way you would know is that it said so, or that a benchmark measured it. Neither has happened.

What would settle it

Three disclosures would convert the post from an assertion into a finding, and all three are things G​oogle could publish this week:

• A parameter count or active-parameter figure for any G​emini 4 model. One number would make “small” checkable and would immediately confirm or kill the premise.

• A definition of “adapted from” on the model cards. The phrase appears seven times on the Neon card and is never explained; “based on” is used for Carbon-to-Argon and is equally undefined.

• Any statement about training contribution — what data was new in Neon relative to Carbon, or which model, if any, generated training signal. Absent all three, RSI remains a word in a post.

What to do with this on Monday

The practical consequence is small, and it is the opposite of what the post implies. Nothing in the record says Neon is a small model that beat a big one; it says Neon is a $3/$15 model that scores one point above Carbon’s ceiling when pushed and half a point below Carbon’s default when not. That is a tuning-ladder decision, not a size decision, and the ladder spans 17.3 index points from low (44.1) to xhigh (61.4) — far more than the gap between any two models in the family. Which rung you call matters more than which model you call.

Two dated things are worth putting in a calendar regardless of how the size question resolves. Carbon and Argon double on 1 December, from $2.00 and $10.00 to $4.00 and $20.00 per million tokens, while Neon’s rate card carries no expiry. On today’s prices Neon is 17% cheaper per completed index task than Carbon ($6.44 against $7.76); after 1 December that gap widens without Neon changing anything. Anyone who reads the “small model” post and concludes Neon is the budget option has the direction right for the wrong reason — it is cheaper per task because it emits less, not because it is smaller, and it becomes much cheaper on a date G​oogle already published.

On our side, google/gemini-4-carbon and google/gemini-4-argon are in the catalogue at provider list price with no markup, which means the 1 December step arrives on the same day it arrives at G​oogle. G​emini 4 Neon is not in the catalogue yet, so the useful thing to say about it is not a price but a shape: when it lands, the request worth writing is one that names Neon and falls back to Carbon if Neon is rate-limited or errors. A model whose main published advantage is finishing tasks quickly is one you want a failover behind, not a single point of dependency. One key covers the rest of the family today, so a comparison run against Carbon or Argon costs nothing but the tokens.

The honest close is the narrow one. A post that says a lab’s small model overtook its large ones has made a claim about sizes. Nobody has published a size. The scores say a fast model landed within a point of a slower one in eight days, which is interesting enough on its own, and the training story attached to it is, for now, a caption.

Compared in this article1

Detected from this article · Benchmarks: Artificial Analysis · updated daily