Google launches Gemini 4 Neon: completes tasks in 14.8 minutes – versus Carbon's 37.1

Google launched Gemini 4 Neon on November 18–19, 2026, positioned between Carbon and Hydrogen in the Gemini 4 family at $3/$15 per million tokens.

Illustration: two parallel tracks of metal beads in glass channels, one surging ahead while the other lags behind – a visual metaphor for the speed gap between two AI models.
Illustration

Google launches Gemini 4 Neon: completes tasks in 14.8 minutes – versus Carbon's 37.1

Google launched Gemini 4 Neon on November 18–19, 2026, positioned between Carbon and Hydrogen in the Gemini 4 family at $3/$15 per million tokens. The first independent figures from Artificial Analysis largely support Google's story of speed and price — but they also reveal how dependent the comparisons against Claude Opus 5.5 and GPT-6.2 Sol are on settings and tooling.

What was launched

Gemini 4 Neon is the newest model in Google's Gemini 4 family, described by the company as a model with "frontier-level intelligence" and fast end-to-end task completion in software development, knowledge work and scientific workflows. It became generally available on November 18 through the paid Gemini API, which developers use from their own software, with rollout to the Gemini app and the Google Antigravity developer platform for Google AI Pro and Ultra subscribers, according to VentureBeat and Winbuzzer. Neon is also available in Google AI Studio and Android Studio, but not in the free tier of the Gemini app — the same applied to Hydrogen, which launched on November 12 (nemui).

The launch timing varies slightly between sources: Winbuzzer and the Japanese analysis from EasyAI研究所 date the announcement and availability to November 18 (EasyAI研究所), while GSMArena wrote on November 19 that Google had "just" announced the model. The most consistent reading is that the model became available on November 18, with coverage articles and Artificial Analysis' comparison published on November 18–19.

Pricing and position in the family

Neon is priced at $3 per million input tokens and $15 per million output tokens. For comparison, Carbon costs $4/$20 and Hydrogen $1.50/$7.50 per million input/output tokens, according to GSMArena. That places Neon exactly in the middle of the two: roughly 25% below Carbon and twice as expensive as Hydrogen.

The model has a one-million-token context window and four thinking levels, according to GSMArena's coverage of Google's information. VentureBeat frames the pricing against competitors: Neon's rates are around 50% higher than Anthropic's mid-tier Claude Sonnet 5.5 and OpenAI's mid-tier GPT-6.2 Sol, but around 25% below Claude Opus 5.5's standard prices — while Google claims Neon beats both top models on the Artificial Analysis Intelligence Index at the highest reasoning setting.

Google's claims — and what is independently confirmed

It is important to distinguish between two types of numbers here. Google's own benchmark comparison, reproduced by GSMArena, shows Neon ahead of Claude Opus 5.5 and GPT-6.2 Sol on most listed tests, including the Artificial Analysis Intelligence Index (61 versus 59 and 57) and Vals Index (70.9% versus 70.1% and 66.1%). Opus 5.5 is nonetheless ahead on Terminal-Bench 4.1 and FrontierCode 2.0. Google also claims Neon has 32% lower cost per task than Carbon at the default high setting, and that the model "maintains performance even in long agent tasks, reduces the need to re-read files, and handles context compression reliably in long sessions," according to the EasyAI rendering of Google's official blog. Google's figures place Neon in first place on Artificial Analysis' code-agent metric (58.6) and on Zapier AutomationBench (56.3%). All of these numbers are Google-reported — none of them are independently confirmed in themselves.

The independent figures come from Artificial Analysis, published via Winbuzzer as of November 19. There, Neon scores 60 on the Intelligence Index at the default high setting — exactly level with Carbon at the same setting — and 61 at xhigh. (GSMArena gives 61; Winbuzzer gives 60 at high and 61 at xhigh, which most likely reflects different reasoning settings.) The speed figures are more unambiguous: Artificial Analysis estimates 14.8 minutes per task for Neon versus 37.1 for Carbon, and Neon generates 101 output tokens per second against Carbon's 55. Estimated cost per task on the overall index was $6.44 for Neon at the highest setting versus $7.76 for Carbon at high with standard prices — and Neon uses around 15% more output tokens per task to get there. At high, Neon came in, according to Winbuzzer, at less than half of Carbon's standard price estimate.

In short: Google's claim of better performance than Carbon is not what the independent figures show at like settings — they show equal performance, but significantly faster and cheaper. That is a different, and more credible, advantage.

A warning before you run your own numbers

There is a trap for anyone running their own cost comparisons right now: Carbon's offer of $2 per million input tokens and $10 per million output tokens only applies through November 30. From December 1, prices rise to $4/$20, according to Winbuzzer. A comparison between Neon and Carbon based on November prices will look different from December onward — Google's "32% lower cost per task" claim is precisely based on the standard prices, not the discount.

The customer's voice

In Google's announcement, VentureBeat quotes Thai Tran, head of AI product at Glean: "In our early evaluations, it matched or exceeded Gemini 4 Carbon in overall quality at 3x the speed and 60% of the cost." This is a customer evaluation published in Google's own announcement — not an independent verification — but it points in the same direction as the Artificial Analysis figures: equal or better quality, significantly faster and cheaper than Carbon.

What remains

Several questions are open. Google's head-to-head comparison against Opus 5.5 and GPT-6.2 Sol is not documented in methodological detail — which settings and tools were used has not been disclosed in the available sources, and Winbuzzer's coverage of the coding-agent comparison suggests the results may be strongly tool-dependent: Neon leads Opus by almost 29 percentage points with separate tools, but is only about one point behind Sol in the same mini-SWE tooling. That means "Neon beats Opus" may in practice depend on which toolchain you run.

The composition of the Gemini 4 family is also not fully clarified: GSMArena describes Carbon, Hydrogen and Neon, while the Japanese sources mention four models including Argon. And Google has promised a "Gemini 4 Neon for cyber" variant "in the coming days," according to EasyAI's rendering of the official announcement — which may say something about where Google sees the most willing-to-pay customers for mid-tier models with agent capabilities.

The most concrete takeaway from the launch, however, is this: Neon does not deliver better results than Carbon according to the independent figures — it delivers the same results faster and cheaper, and that is probably exactly what the mid-tier market is about.

AIMag.no
AIMag.no
The AIMag.no editorial team covers artificial intelligence, tools, research, and regulation.

Get the best of AI MAG in your inbox

News, analysis, and ideas at the intersection of AI and society.