Google's Gemini 4 Neon runs twice as fast as Carbon and costs 17% less per task
Google DeepMind has released Gemini 4 Neon, its fourth model in seven weeks. It generates about 110 tokens (chunks of text) per second, against 57 for the previous Gemini 4 Carbon, so an assistant answers almost twice as fast. One task in the Artificial Analysis test costs $6.44 instead of $7.76, which is 17% cheaper. If you pay for the API, run your own tasks on Neon, but note that the savings come from the lower price per token, not from shorter answers.
AI-processed from The Decoder; edited by Hamidun News
Google DeepMind has released Gemini 4 Neon, its fourth model in seven weeks. It targets a problem familiar to anyone building AI into a product: a smart model is either slow or expensive. Imagine a coding assistant that you ask to fix a bug in a repository.
On Neon, it generates a response at about 110 tokens (chunks of text) per second. On the previous flagship, Gemini 4 Carbon, the speed is 57, so Neon is almost twice as fast. One such task in the Artificial Analysis test costs $6.
44, versus $7.76 for Carbon at its regular price. That is where the 17% savings come from.
If you pay for the API or are choosing a model for an agent, these numbers translate directly into your bill and your users' wait time.
Where Neon sits in the lineup
Neon sits between Carbon and Hydrogen. It has a context window of one million tokens and accepts text, images, audio and video. On the Artificial Analysis general Intelligence Index at the maximum reasoning level ("xhigh"), Neon scores 61 points to Carbon's 60. That puts it in third place, behind only OpenAI's GPT-6.2 Astra and Anthropic's Claude Fable 5.2.
Interestingly, Google's own comparison table leaves out those two models. The company compares Neon only with Claude Opus 5.5 and GPT-6.2 Sol. When choosing a model, look at independent measurements, not just the developer's slides.
Where the savings and speed come from
Through Google AI Studio and Gemini for Business, one million input tokens cost $3 ($0.15 with caching) and one million output tokens cost $15. That is 25% below Carbon's regular price ($4 and $20). But Carbon has a promotion running through the end of November, and at that rate Neon is 50% more expensive. So until December, Carbon is formally the better deal at list price.
An important caveat: Neon spends roughly 15% more output tokens per task than Carbon, meaning its answers are more verbose. The $6.44 versus $7.76 saving comes entirely from the lower price per token, not from the model becoming more economical. If Google raises its prices, the advantage will shrink.
At the second-highest reasoning level ("high"), Neon posts the same intelligence score as Carbon, while the cost per task is half as much. A benchmark run drops from about 36 minutes to 21, precisely because of the 110 versus 57 tokens-per-second speed. For interactive products where a user waits for the answer, the difference is noticeable.
Coding, documents and security
In the coding agent index within Google's Antigravity environment, Neon at "xhigh" scores 59 points. Carbon gets 55, while Claude Fable 5.2 and Claude Opus 5.5 each get 58. The average coding task costs $6.86, compared with $12.61 for Opus and $15.47 for Fable. In SWE-sweep, Google says the Antigravity environment triples Neon's share of solved tasks, to 39.0%.
There is also progress in working with visual content. On the GDP.pdf document test, Neon at the "high" level scores a record 33.4%, five points above Carbon. But at "xhigh" the result unexpectedly drops to 28.0%, so the maximum reasoning mode is not always the best choice. In AA-Briefcase, the model leads in presentation quality (1771 Elo, ahead of Opus 5.5). By Google's own measurements, Neon is also ahead in individual Vals Index categories, in Zapier's AutomationBench (56.3%) and in the scientific RiemannBench (84.0%).
For security professionals, Google is launching a "cyber" variant through its Fairwind program. It scores 90.5% on SEC-bench Pro and 51.26% on WebExploitBench v1.2, at half the price of competitors. According to the very short model card, Neon blocks 100% of indirect prompt injection attacks in the Gray Swan benchmark. That matters for agents that read emails and web pages, but the figure is self-reported by Google, and the methodology is hard to verify.
What to do in practice
If you are on Carbon, run your real tasks on Neon in "high" mode: according to Artificial Analysis, the intelligence is the same while the cost per task is about half. Do not rely on a single index, especially if you work with documents: test your own scenario. And keep in mind that Carbon's promotional price through the end of November changes the math, and that Neon produces more output tokens than its predecessor.
- Nov 19, 2026A GitHub Script Catches App Store and Google Play Rejection Reasons Before Review: 542 Stars in a Day
- Nov 19, 2026Nvidia posts record $115.8 billion revenue, yet trades cheaper than at any point since 2012
- Nov 19, 2026After an Anthropic ban, Codex moved 43 Claude Code sessions in 20 minutes
- Nov 19, 2026Google releases Gemini 4 Neon: 61 on the Artificial Analysis index at $3/$15 per million tokens
Key Questions Answered
Where Neon sits in the lineup?
Neon sits between Carbon and Hydrogen. It has a context window of one million tokens and accepts text, images, audio and video. On the Artificial Analysis general Intelligence Index at the maximum reasoning level ("xhigh"), Neon scores 61 points to Carbon's 60. That puts it in third place, behind only OpenAI's GPT-6.2 Astra and Anthropic's Claude Fable 5.2. Interestingly, Google's own comparison table leaves out those two models. The company compares Neon only with Claude Opus 5.5 and GPT-6.2 Sol.
Where the savings and speed come from?
Through Google AI Studio and Gemini for Business, one million input tokens cost $3 ($0.15 with caching) and one million output tokens cost $15. That is 25% below Carbon's regular price ($4 and $20). But Carbon has a promotion running through the end of November, and at that rate Neon is 50% more expensive. So until December, Carbon is formally the better deal at list price. An important caveat: Neon spends roughly 15% more output tokens per task than Carbon, meaning its answers are more verbose.
What are the key details on Coding, documents and security?
In the coding agent index within Google's Antigravity environment, Neon at "xhigh" scores 59 points. Carbon gets 55, while Claude Fable 5.2 and Claude Opus 5.5 each get 58. The average coding task costs $6.86, compared with $12.61 for Opus and $15.47 for Fable. In SWE-sweep, Google says the Antigravity environment triples Neon's share of solved tasks, to 39.0%. There is also progress in working with visual content.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.