# Google launches Gemini 4 Neon, a faster and cheaper AI model

> The new Gemini 4 Neon model offers double the speed of its predecessor, lower per‑task costs and a 1 million‑token context window.

Oossa · 2026-11-19 · https://oossa.com/en/google-launches-gemini-4-neon-a-faster-and-cheaper-ai-model

Google DeepMind announced Gemini 4 Neon on 18 November 2026. The model is the fourth release in the Gemini 4 series in seven weeks. Neon runs about twice as fast as the previous Gemini 4 Carbon, scores higher on most benchmarks and is priced lower per token. It can handle up to one million input tokens across text, images, audio and video in a single request.

## Performance and cost improvements

At its top "xhigh" reasoning level Neon scores 61 on the Artificial Analysis Intelligence Index, edging out Carbon’s 60. It generates roughly 110 tokens per second, compared with Carbon’s 57, cutting benchmark run time from 36 minutes to about 21 minutes. Per‑task cost at xhigh is $6.44, about 17 % cheaper than Carbon’s $7.76, even though Neon uses 15 % more output tokens – the savings come from a lower list price. Input tokens cost $3 per million (or $0.15 when cached) and output tokens $15 per million, 25 % below Carbon’s regular rates.

Neon also leads on coding benchmarks. On the Coding Agent Index it scores 59 points, beating Carbon’s 55 and charging an average coding task $6.86, versus $12.61 for Claude Opus 5.5 and $15.47 for Claude Fable 5.2. In Google’s internal SWE‑sweep test Neon’s hit rate reaches 39 %.

Visual and scientific tests show similar gains. Neon tops the document test GDP.pdf with a 33.4 % score at "high" level, beats Carbon on the AA‑Briefcase presentation test, and leads on RiemannBench (84 %).

## Availability and safety

Neon is now available through the Gemini API, Google AI Studio, Gemini for Business and to paid Google AI Pro and Ultra subscribers. Google also offers a cyber‑focused variant via the Fairwind Program for trusted defenders.

The model keeps the safety protections of Gemini 4 Carbon, blocking 100 % of indirect prompt‑injection attacks on the Gray‑Swan benchmark and refusing only 1.98 % of tasks on a terminal safety test. Its knowledge cutoff is March 2026.

## The facts

- Google DeepMind released Gemini 4 Neon on 18 November 2026.
- Neon processes about 110 tokens per second, nearly double Carbon’s 57 tokens per second.
- Input tokens cost $3 per million (or $0.15 cached); output tokens cost $15 per million.
- Neon scores 61 on the Artificial Analysis Intelligence Index at its xhigh level, one point higher than Carbon.
- On the Coding Agent Index Neon averages $6.86 per coding task, versus $12.61 for Claude Opus 5.5.

## Why it matters

For developers and businesses, Neon means faster AI‑assisted coding and analysis at lower price points, making large‑scale automation more affordable. The expanded 1 million‑token context lets users feed whole documents, spreadsheets or video transcripts in one go, reducing the need to split work into many calls. However, the model still has a March 2026 knowledge cutoff and may produce occasional hallucinations, so critical decisions still need human review.

## Sources & references

1. [Google's latest AI model is faster, cheaper and smarter than its flagship predecessor from just weeks ago](https://the-decoder.com/googles-latest-ai-model-is-faster-cheaper-and-smarter-than-its-flagship-predecessor-from-just-weeks-ago/) – The Decoder, 2026-11-19
2. [Google](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/?utm_source=openai) – Google
3. [Google DeepMind](https://deepmind.google/models/model-cards/gemini-4-neon/?utm_source=openai) – Google DeepMind

Last updated: 2026-11-19
