Title card reading Gemini 4 Neon's First Independent Numbers, with the line The 61 is the xhigh setting. At its default it scores 59.6, and Carbon scores 60.0, above three cards reading Gemini 4 Neon at xhigh 61.4, Gemini 4 Neon at high (default) 59.6 and Gemini 4 Carbon at high 60.0, footed with Artificial Analysis Intelligence Index v5.0.1, 19 November 2026.
Guides & Insights

Gemini 4 Neon's First Independent Benchmarks: The 61 Is an xhigh Number, and at Its Default It Lands Behind Carbon

Author

jinhao.song

Date Published

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

Google's launch post for Gemini 4 Neon leads with an Artificial Analysis Intelligence Index of 61, third overall. Artificial Analysis now publishes its own page for Gemini 4 Neon, and that page says 61.4 — for the xhigh variant, the deepest of four thinking settings, ranked sixth of 96 models on index v5.0.1. At the setting Neon actually defaults to, high, the same tracker scores it 59.6. Gemini 4 Carbon, the model Google's announcement measures Neon against in nearly every paragraph, sits at 60.0 at its own maximum. Read the two pages the way they are published and the newest Gemini 4 is not the smartest one: it wins by 1.4 points when you push it to a rung Carbon does not have, and loses by 0.4 points when you compare the two models at the settings their vendors recommend.

None of that makes Neon a bad model, and the speed numbers in the same panel are the strongest argument for it that anyone has published. But it does mean the one figure most likely to end up in a slide deck this week is the least representative configuration of the model, and the correction is now checkable by anyone with a browser rather than an inference from a vendor's own table.

What the tracker actually published

Artificial Analysis lists Gemini 4 Neon as released 18 November 2026, proprietary, Google, with a 1M-token context window and text, image, speech and video input. It carries four separately scored variants — low, medium, high and xhigh — and the page opens on xhigh by default, which is why the headline figure is the 61.4 rather than the number you would get by calling the model with no thinking parameters at all. The index in force is v5.0.1, which matters for anyone holding an older score: this is a rebased index, and a 61 from a summer leaderboard is not the same measurement as a 61 today.

The same page scores the rest of the family, and the ordering is worth reading before the individual numbers. All four use the same evaluation set — AA-Briefcase v1.1, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode-AA 2.0, AA-Omniscience v1.1, GDP.pdf, HLE-Diamond, CritPt v2 and Terminal-Bench-Science 0.1 — so the comparison is at least like for like, which is more than can be said for the benchmark table in the launch post.

The Artificial Analysis model page for Gemini 4 Neon, headed Gemini 4 Neon: Intelligence, Performance and Price Analysis, showing Intelligence 61 at rank #6 of 96, Performance #73 of 95 with 56.51 seconds per response, 110 output tokens per second and 21.5 minutes per task, Cost #82 of 95 at $3.00 input and $15.00 output per million tokens with a 95% cache discount and $6.44 per Intelligence Index task, and Verbosity #93 of 95, above a low, medium, high and xhigh variant selector.

One model, four scores, and a much steeper ladder

• Neon at low — 44.1, against Carbon at low 54.2 and Gemini 4 Hydrogen at low 26.9.

• Neon at medium — 54.9, against Carbon 58.3 and Hydrogen 41.2.

• Neon at high — 59.6, against Carbon 60.0 and Hydrogen 50.8. This is Neon's default and Carbon's ceiling.

• Neon at xhigh — 61.4, against no Carbon entry at all. Gemini 4 Hydrogen has a fourth rung and it scores 53.9.

The shape of those four numbers is the part the launch post does not convey. Neon spans 17.3 index points from its cheapest to its deepest setting; Carbon spans 5.8 across its three rungs. Neon is not a fixed quantity that happens to have a tuning knob — it is four materially different models depending on which rung you call, and the differences are largest at the bottom, where Neon's low is 10.1 points under Carbon's low. If you have been sizing Neon against Carbon using a single number per model, the honest version is a curve against a flat line.

That has a practical consequence that runs opposite to the launch framing. Google's documentation tells developers to evaluate at all four levels and pick the lowest one that clears their quality bar, and warns that xhigh "can cause overthinking or exhaust max_output_tokens on simpler prompts such as text or JSON extraction." The independent numbers show why that advice is load-bearing rather than cautious: stepping down is not a rounding error on this model. The 1.8 points between Neon's default and its ceiling is more than four times the distance between Neon and Carbon; the 4.7 points between high and medium is nearly twelve times it.

A two-column scoreboard titled Gemini 4 Neon — the scoreboard. Gemini 4 Neon at xhigh reads AA Index 61.4, cost per task $6.44, output tokens 283M, time per task 21.5 min, output speed 110 tok/s; Gemini 4 Carbon at high reads AA Index 60.0, cost per task $4.90, output tokens 205M, time per task 36.1 min, output speed 57 tok/s. The footer says both figures are per Artificial Analysis on index v5.0.1, read 19 November 2026, and that Google's 32% cheaper-per-task claim is not reproduced at these settings.

What a task costs, measured rather than claimed

The claim worth testing is the per-task one. Google says Gemini 4 Neon at default high costs 32% less per task than Gemini 4 Carbon, and 36% or more less than competing frontier models at their highest thinking setting. Per-task cost is a function of how many tokens a model emits, and the launch post discloses no token counts, so until this week the number was untestable from outside.

The tracker's panel puts Neon at $6.44 per Intelligence Index task and Carbon at $4.90. That is not a 32% saving; it is a 31% premium, and the direction is the opposite of the claim. The honest caveat is a configuration mismatch, and it cuts both ways: $6.44 is Neon at xhigh, because xhigh is the variant the page opens on, while $4.90 is Carbon at high, its own maximum. Google's 32% figure is about Neon at high, which the tracker scores separately at 59.6 but does not headline with a cost in the same panel. So the correct statement is narrower than "Google is wrong": the published configuration of Neon costs more per task than the published configuration of Carbon, and the specific comparison Google is making has not been independently reproduced.

What is reproducible is the token profile, and it undercuts the intuitive reading of a faster model. Neon generated 283M output tokens across the Intelligence Index against Carbon's 205M — 38% more — on 8.12B input tokens, with output running at 3.5% of input. Median across the models the tracker scores is 65.1M output tokens, so Neon is not merely chattier than Carbon; it is among the most verbose models on the board, ranked 93rd of 95 on that measure. Carbon, which had a reputation for verbosity, is 90th. Whatever Neon is saving you, it is not saving it in emitted tokens.

Where it genuinely is faster — and this is the strongest case for it — is wall clock:

• Time per Intelligence Index task — Neon 21.5 minutes, Carbon 36.1, Hydrogen 15.3, Argon 29.6. The median for comparable models is 12.4 minutes.

• Output speed — Neon 110 tokens per second, Carbon 57, Hydrogen 168, Argon 65.

• Time to first answer token — Neon 56.5 seconds, Carbon 75.0, Hydrogen 47.2, Argon 47.1.

Neon finishes the same evaluation set in 40% less wall clock than Carbon at roughly twice the generation rate, while emitting more tokens and scoring 0.4 points lower at matched default settings. That is a coherent product, and it is a different product from the one the benchmark table implies: it is a model you pick to finish sooner, not one you pick to be right more often.

Three numbers in the launch post that the tracker does not corroborate

The rank. Google says third overall. The tracker's live page ranks the xhigh variant sixth of 96 on index v5.0.1. Both figures can be traceable to Artificial Analysis and still be different, because the model set moves; what cannot be true is both being current. Cite the live page, and say which variant it belongs to, or the number will not survive the next release.

The setting behind the score. A 61 with no qualifier reads as the model's score. It is one of four, it is the most expensive one to run, and it is the one Google's own docs warn about for extraction-shaped tasks. The default is 59.6.

The per-task advantage. Discussed above: at the configurations the tracker publishes in its summary panel, Neon is 31% dearer per task, not 32% cheaper. The vendor claim may still hold at high on Google's own harness. It has not been shown to hold anywhere a reader can check.

Everything else in the announcement remains what it was — third-party leaderboards Google is citing, and Google's own unreproduced evaluations, from Chartography at 73.5% to FrontierCode 2.0 at 46.2% and the cyber figures that ship only through the Fairwind defenders programme. None of that is contradicted by the independent numbers. It simply is not the part anyone can verify, and the part that can be verified now reads differently.

What this changes if you were about to standardise on Neon

If your reason for picking Neon was the 61, the number you should carry into a capacity plan is 59.6, and the model it beats at that setting is nothing in the current family — Carbon's 60.0 is ahead of it and Carbon's price is $2.00 and $10.00 per million tokens against Neon's $3.00 and $15.00 until the introductory rate on Carbon and Argon steps to $4.00 and $20.00 on 1 December 2026.

If your reason was the speed, the case is real and the numbers above support it. Twenty-one minutes against thirty-six on the same evaluation set is the kind of difference that changes how much you can run in a working day, and the 1.9x generation rate shows up in interactive work well before it shows up in a batch job.

If your reason was Google's per-task economics, run the test yourself before you commit a production path to it. The token profile is the one thing that decides whether the per-task claim holds on your traffic, and the independent measurement says Neon emits more tokens than the model it is supposed to be undercutting. On input-heavy work the picture inverts, because Neon's input price is 50% above Carbon's today and Carbon's step in December closes even that gap in Neon's favour — which is a fourteen-day window, not a strategy.

Running the comparison on your own traffic

The reason to care about a four-rung ladder rather than a single score is that the rung is a cost decision, and cost decisions are cheapest to make when both models are one call away. Our catalogue carries google/gemini-4-carbon and google/gemini-4-argon at provider list price with no markup, which means the 1 December step on those two models appears on your invoice the day it appears on Google's, and Neon's December position relative to them changes without anyone renegotiating anything. Gemini 4 Neon is not in the catalogue yet, so this is not a recommendation to run it through us today — it is the shape of the test to run when it lands.

The OrcaRouter model page for Gemini 4 Carbon, showing the model identifier google/gemini-4-carbon, a maximum output of 65K tokens, multimodal input across text, image, file, audio and video, capability chips for vision, audio, tools, JSON, reasoning and output text, the line by Google · 2026-11-10, a p50 time to first token of 10.005 seconds, and the description noting a 1,048,576-token context window.

The evaluation the numbers above argue for is not a benchmark run. It is the same prompt at high on both models, measured in tokens to finish rather than seconds, on a sample of your own tasks. That test needs both models reachable from one key with a fallback behind it, because the point of running Neon is to find out whether the per-task claim survives contact with your traffic, and the point of the fallback is that a model you are still evaluating never gets to be the only path to production. Carbon behind it, Argon behind that, and the ladder comparison becomes a routing change rather than a migration.

The one thing not to do is size the decision on the launch table. The tracker has now published four scores for one model, the gap between the top and the default is larger than the gap between the default and the model it replaces, and the cost claim that makes the whole proposition work is the one number on the slide that nobody outside Google has reproduced.

What to watch next

Three things, all checkable. Whether the tracker's per-task cost for Neon at high lands near Google's 32% figure — that single row would settle the argument, and it is a variant that already has a published index score. Whether the Fairwind gate on the cyber variant opens, since the security numbers are still claims about a model almost nobody outside Google can call. And whether the 1 December step on Carbon and Argon survives contact with demand, because that is the date on which Neon's flat $3.00 and $15.00 stops looking expensive and starts looking like the cheapest Gemini 4 reasoning tier per token.

Until then the useful summary is one sentence long: Gemini 4 Neon is the fastest way to get a Gemini 4 answer, it is not the most accurate one at the setting it ships with, and the efficiency claim that justifies its price is still a vendor's.