ModelBrief · Gemini 4 Neon · Topic Page
Last updated Nov 20, 2026

Gemini 4 Neon vs Carbon: Same Lineage, 61 vs 60, and When Switching to Neon Pays

Published Nov 19, 2026 · Last updated Nov 19, 2026

Neon and Carbon are not two rivals but one lineage: DeepMind's model card states plainly that Neon is "adapted from Gemini 4 Carbon", with architecture, training data and hardware details all pointing to the Carbon model card. Google's positioning is equally direct: Carbon takes the hardest long-horizon projects, Neon takes high-frequency complex work. So the question becomes: when is the mid-tier model adapted from the same base worth using? The answer: it leads by 1 point only at the top tier, it is cheaper only from December 1, and the real difference is the subscription bar — Neon is open to Pro users, while Carbon is Ultra-only.

Same lineage: Neon is adapted from Carbon

The model card's wording is restrained: "Gemini 4 Neon is adapted from Gemini 4 Carbon", and the sections on architecture, training dataset, training data processing, hardware, software and acceptable usage all say the same thing — see the Carbon model card for details. Google has not published what exactly changed between Neon and Carbon — neither parameter counts nor architecture details are disclosed. Three known differences: Neon adds an xhigh thinking tier; Neon's safety capability determinations carry forward from Argon (the MCC assessment found Neon's capability increase over Argon below the material-change threshold); and Neon's behavior metrics — fabrication rate, shortcut rate — beat Carbon's (Google's internal measurements). Think of Neon as "Carbon specialized for high-frequency work" rather than "a new generation".

Scores: the truth about 61 vs 60

The AA index is the most direct comparison (v5.0.1, measured independently by Artificial Analysis):

Thinking tierNeonCarbon
Top tier (Neon xhigh / Carbon high)6160
Default tier (both high)6060
medium5558.3
low4454

The headline "Neon beats Carbon 61 to 60" holds only at the top tier; at the default tier they tie; at medium and low, Carbon leads by 3.3 and 10 points. Neon's score advantage is top-tier-only, while Carbon's middle and low tiers are stronger. Neon has one genuine headline on single indexes: the AA Coding Agent Index at 58.6, No. 1. But in the model card table, Neon loses Terminal-Bench 4.1 and FrontierCode 2.0 to Opus 5.5 by a hair — and the only comparable Carbon figure is Terminal-Bench 4.0 at 65.2% (Google's official-table figure), tying Neon's 65.2%; there is no comparable Carbon score for FrontierCode 2.0 (the 61.2% in Carbon's official table is FrontierSWE 2, a different benchmark — do not mix them).

The pricing calendar: which is cheaper depends on today's date

Per token, Neon at $3/$15 is a fixed price, while Carbon at $2/$10 is a promotional price ending November 30. Until November 30: Neon costs 50% more than Carbon. From December 1: Carbon rises to $4/$20 and Neon is 25% cheaper. Per task, Google claims Neon costs 32% less than Carbon at the default tier (Google's own measurement, AA index tasks); but LMArena's Agent cost snapshot points the other way (median $2.66 per task for Neon vs $1.34 for Carbon) and cannot be reconciled yet. See our pricing guide. In short: choose by the calendar on per-token price; wait for independent measurements on per-task cost.

Speed: Neon is a little faster, but neither is fast

AA measured Neon at 21.5 minutes per index task and 56.51 seconds to first answer token at xhigh; Carbon at High takes about 26–27 minutes per index task with a first token at about 64.7 seconds. Neon is indeed a bit faster, but both sit in the "slow" band — AA notes Neon at xhigh is among the slowest in its price tier. Google's "Flash-level speed" is a qualitative claim relative to Carbon, not absolute speed; neither model suits interactive use as a default.

Subscriptions: the biggest practical difference

In the Gemini app, Carbon is available only to Google AI Ultra subscribers; Neon is open to Pro and Ultra. For users already paying for Pro, this is Neon's most concrete advantage: near-flagship capability without upgrading to Ultra. On the API side both are paid-tier only with no free tier — no difference. In Antigravity both are available to paid Pro / Ultra subscribers, and Neon adds an Orchestrator role (directing Hydrogen subagents).

Decision table

Task typePickWhy
Already paying for Pro, want near-flagship capabilityNeonCarbon is Ultra-only; Neon is the strongest Gemini 4 model a Pro subscriber can get
High-frequency complex coding, enterprise knowledge workNeonCoding Agent Index No. 1; Google claims 32% less per task at default tier (own measurement)
Hardest long-horizon projects, tough debugging, scienceCarbonGoogle's official positioning; stronger middle/low-tier AA scores (medium 58.3 vs 55, low 54 vs 44)
Interactive coding, quick daily loopsNeitherBoth are slow; use Hydrogen (see our Neon vs Hydrogen comparison)
Batch work before Nov 30CarbonPromotional $2/$10 is 50% cheaper than Neon
Batch work from Dec 1NeonCarbon rises to $4/$20; Neon is 25% cheaper
Cybersecurity workNeon for cyberSEC-bench Pro 90.5%, AutoBaxBench hard tier 76% (Google-reported); Fairwind application only

Migration notes

Moving from Carbon to Neon needs no request changes — change the model ID from gemini-4-carbon to gemini-4-neon. Three defaults deserve attention: the thinking tiers go from three to four, with high still the default; only switch xhigh on for the hardest jobs — 2.3x the tokens for 1 AA index point is poor value; and set max_output_tokens to at least 32,768 (thinking and answer share the cap, and the empty-answer trap already confirmed on Carbon applies here too). Also note: Neon supports only implicit caching, not explicit caching; like Carbon, it reportedly ignores temperature, topP and topK (tracker-reported; not verified against official documentation in this round).

Our take: Neon is not Carbon's replacement but its "Pro edition": the same base, the same score at the default tier, a 1-point lead at the top tier, 25% cheaper from December 1 — and the biggest practical difference is that Pro subscribers can use it. Ultra subscribers doing the hardest work should stay on Carbon; Pro subscribers doing high-frequency complex work should pick Neon. Before making it your default, ask one question: is today before or after November 30?

Reference sources: this article is an original rewrite produced by this site; facts and figures were checked against DeepMind's model card (Nov 18, 2026), Artificial Analysis' model page (read Nov 19, 2026) and Google's official blog (Nov 18, 2026)

← Back to the Gemini 4 Neon topic page