Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableHacker News AI→ original

Google Releases Gemini 4 Neon: Tasks Cost 32% Less Than on the Flagship Carbon

Google has released Gemini 4 Neon, the mid-sized model in the Gemini 4 lineup. It comes close to the flagship Carbon, but runs faster and costs less. At the high reasoning setting, a completed task costs 32% less than with Carbon, and on the independent Artificial Analysis Intelligence Index, Neon ranks third with 61 points. Tokens cost $3 per million on input and $15 on output. If an agent is running up your bill, compare models by the cost of a solved task.

AI-processed from Hacker News AI; edited by Hamidun News
Google Releases Gemini 4 Neon: Tasks Cost 32% Less Than on the Flagship Carbon
Source: Hacker News AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHacker News AI ↗
Editor-in-Chief
Cached input$0.15 per million tokens, a 95% discount
SWE-sweep39.0% in Google Antigravity versus 12.8% in mini-SWE
1 million token contexttables, PDFs, audio, and video in a single request

Strong models are usually slow and expensive, and on long agentic tasks the bill grows faster than quality does. Google is trying to close that gap with Gemini 4 Neon. It is a mid-sized model in the Gemini 4 family.

The top model in the family is Carbon, which handles the heaviest projects but costs more. The headline number of the release is 32%: according to Google, that is how much less Neon costs than Carbon per completed task. The metric is not the price per token but the total cost of a solved task, meaning how many tokens the model spends on reasoning and on its answer.

Alongside it is a score on the Artificial Analysis Intelligence Index, an independent composite ranking that runs models through a battery of intelligence tests. At the highest reasoning level (xhigh), Neon scored 61 points there and took third place, behind only larger and slower models.

Where Neon Sits in the Gemini 4 Family

The lineup is named after chemical elements. Carbon is the heaviest model, built for long, complex projects. Hydrogen is a fast workhorse for everyday tasks. Neon sits in the middle, offering deep reasoning at a lower price and better speed.

Pricing: $3 per million input tokens and $15 per million output tokens. Cached input, meaning the repeated part of a request, costs $0.15 per million, which is 95% cheaper. The context window is 1 million tokens. You can load tables, PDFs, charts, audio, and video at once and get structured output.

The model has four reasoning levels, and the higher the level, the more tokens go to reasoning. At the high level by default, the savings against Carbon are the 32% cited above. On the independent Vals and Agent Arena leaderboards, Neon sits on the Pareto frontier: among the models tested, no one delivered better results for the same money. Against top competitors, a task comes out roughly 50–80% cheaper. At xhigh, where intelligence is at its maximum, the savings are 36% or more.

Code, Agents, and "Working"

Benchmarks

For developers, the most important thing is long agentic sessions. Google says Neon re-reads the same files less often, reports on its progress, and compresses context more reliably when a session drags on. It can work on its own or act as an orchestrator that hands tasks to parallel Hydrogen subagents in the Google Antigravity environment.

Numbers: first place on the Artificial Analysis Coding Agent Index (58.6) and 46.2% on FrontierCode 2.0 at the high setting. On SWE-sweep (a test of editing real repositories) in the simplified mini-SWE harness, the model solves 12.8% of tasks, while in Antigravity the result more than triples, to 39.0%. The agent harness affects the outcome as much as the model does, so these numbers cannot be compared with others' runs in a different harness.

Among office tasks, Google highlights 73.5% on Chartography (reading dense professional charts), first places on several financial and spreadsheet tests, and $16,843 in Vending-Bench 2, a long business-management simulation. On the Agent Arena Index, the model debuted in second place with 1336 points, based on more than 4,200 real agent sessions. In demos of an agent "swarm," Neon built a WebGL ski-slope simulator and turned a single image into a Google Sheets table of 100,000 cells with more than 500,000 formulas.

Cybersecurity and the Model's Own Safety

A separate focus is code security. Neon writes safer code than Carbon at every difficulty level and reaches 76% on the hardest one. On a penetration-testing benchmark (a simulated attack to find weaknesses), it scores 51.26% at half the price of large competitors. On vulnerability discovery it scores 90.5%, and on root-cause fixing 70.7%, which is 3.4% better than Carbon. A cybersecurity version of Neon will become available to verified defenders through the Fairwind program in the coming days. Access is requested through a Google Cloud account manager.

In terms of behavior: on Terminal-Bench 4.0 the model refuses 1.98% of tasks, and on Terminal-Bench-Science 0%, meaning it cuts off legitimate work less often. The rate of corner-cutting in the cheating check is 2.45%, the share of dangerous actions in the agentic suite is 0.09%, and honesty on its own test is 99.2%. Against hidden commands planted in web pages, emails, and documents, Neon with real-time protection blocked all attacks in the Gray Swan benchmark.

What to Do About It

Almost all of these figures come from Google itself or were commissioned from partners, and there has not yet been independent re-verification. The practical takeaway is clear, though: if an agent running on Carbon or a rival flagship is running up your bill, run Neon on your own tasks at the high and xhigh levels and compare the cost of a completed task, not the price per token. The model is already available in the Gemini API (paid tier), in Google Antigravity, and in the Gemini app for Pro and Ultra subscribers, while all subscribers to those plans can try it in AI Studio.

Story

Key Questions Answered

Where Neon Sits in the Gemini 4 Family?

The lineup is named after chemical elements. Carbon is the heaviest model, built for long, complex projects. Hydrogen is a fast workhorse for everyday tasks. Neon sits in the middle, offering deep reasoning at a lower price and better speed. Pricing: $3 per million input tokens and $15 per million output tokens. Cached input, meaning the repeated part of a request, costs $0.15 per million, which is 95% cheaper. The context window is 1 million tokens.

What are the key details on Code, Agents, and "Working" Benchmarks?

For developers, the most important thing is long agentic sessions. Google says Neon re-reads the same files less often, reports on its progress, and compresses context more reliably when a session drags on. It can work on its own or act as an orchestrator that hands tasks to parallel Hydrogen subagents in the Google Antigravity environment. Numbers: first place on the Artificial Analysis Coding Agent Index (58.6) and 46.2% on FrontierCode 2.0 at the high setting.

What are the key details on Cybersecurity and the Model's Own Safety?

A separate focus is code security. Neon writes safer code than Carbon at every difficulty level and reaches 76% on the hardest one. On a penetration-testing benchmark (a simulated attack to find weaknesses), it scores 51.26% at half the price of large competitors. On vulnerability discovery it scores 90.5%, and on root-cause fixing 70.7%, which is 3.4% better than Carbon. A cybersecurity version of Neon will become available to verified defenders through the Fairwind program in the coming days.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

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.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…