# Gemini 4 Neon cuts Google's coding to $3 per input

> Google launches Gemini 4 Neon at $3 per million input tokens and $15 for output, with a one-million context window and its own figures against Opus 5.5 and GPT-6.2 Sol.
> By José Maldonado · Artificial Intelligence · 2026-11-20

Google presented Gemini 4 Neon, its new AI model with a rate designed for heavy use: $3 per million input tokens and $15 per million output tokens, with a context window of one million tokens. A token is the unit used to bill the text that goes into and comes out of the model, and input caching lowers the price to $0.15, 95% less. The [announcement of Gemini 4 Neon](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/) is signed by Koray Kavukcuoglu, of Google DeepMind, and talks more about cost per task than about scores.

Neon is the fourth Gemini 4 in less than two months. Argon opened the series on September 30 with restricted access, Carbon arrived on October 28, Hydrogen on November 12 and Neon on the 18th. Google places it between Carbon, the model for long projects, and Hydrogen, the one for fast responses.

The figures behind the promise come from Google, which sells Neon as the value member of the family. The company compares it with Claude Opus 5.5 and GPT-6.2 Sol in 19 rows of its technical sheet, wins in 13 of the 16 where all three appear and admits in the same document that the model "may present general limitations such as hallucinations." That detail matters more than it seems, because a low price per token only pays off if the work comes out right the first time.

## A rate below the bigger sibling

The natural comparison is with Carbon, the large model of the same family. Its standard rate is $4 per million input and $20 per million output and takes effect on December 1. Until November 30 an introductory price of $2 and $10 applies. Neon comes in 25% below the standard rate and 50% above the promotion, so the advantage depends on the calendar.

Model and rate

Input

Output

Input cache

Gemini 4 Neon

3 $

15 $

0,15 $

Gemini 4 Carbon, standard rate from December 1

4 $

20 $

0,20 $

Gemini 4 Carbon, introductory rate until November 30

2 $

10 $

0,10 $

Anyone comparing the two rates today will see that Carbon is cheaper per token, and anyone comparing from December 1 will see the opposite. The [Gemini API price list](https://ai.google.dev/gemini-api/docs/pricing) adds the rest: Neon charges $0.15 per million cached tokens versus Carbon's $0.20, and the batch rate drops to $1.50 for input and $7.50 for output. The cache discount matters for agents that reread the same repository for hours.

## The tests Google chooses

On the Artificial Analysis index, an external measurement, Neon scores 61 points at the xhigh reasoning level and 60 at the high level. Google presents it as third globally, behind larger and slower models, and there the first crack appears: the best number requires the most expensive setting, with 21.5 minutes per task versus 14.8 at the high setting, according to the measurements published by the index itself.

That trade-off between score and time decides real-world use. An agent working overnight can think for twenty minutes per task; an assistant inside the editor cannot, because the developer is waiting in front of the screen.

Against the rivals the company chooses, the split is uneven. Neon wins on the Vals index with 70.9% and on the Coding Agent Index with 58.6. It loses on Terminal-Bench 4.1 with 65.2% versus Opus 5.5's 65.8%. It also falls on FrontierCode 2.0, where Opus reaches 47.3% and Neon stays at 46.2%.

There are rows where the gap is enormous and Google shows them for good reason. The GraphWalks long-context test leaves Neon at 95.2, versus 66.8 for Opus and 80.1 for Sol. On SWE-sweep, with Antigravity's own harness, it rises to 39.0% while Opus stays at 10.1%. The [analysis of the 19 rows](https://www.therundown.ai/news/google-gemini-4-neon-release) notes that three rows omit a rival and that, in the 16 comparable ones, Neon beats both in 13.

## Cost per task, which is another calculation

The central promise is about cost, not capability. With the default high setting, Google puts cost per task 32% below Carbon and between 50% and 80% below frontier models on the Vals and Agent Arena indexes. With the xhigh setting the gap narrows on intelligence and holds on money, 36% less per task than those rivals at their highest level.

All those figures are Google's and measure tasks from its own tests. The team that buys has to redo the math with its repository and its documents, because cost depends on the code the model receives.

The [analysts consulted](https://www.infoworld.com/article/4250618/google-launches-gemini-4-neon-touting-lower-costs-for-complex-ai-coding-tasks.html) point in the same direction. Prabhu Ram, of CyberMedia Research, expects companies to adopt a tiered scheme rather than replace the expensive model with the cheap one, and recommends evaluating the total cost of completing a task rather than a test score. Anand Joshi, of JP Data, warns that the generated code can be a black box for whoever maintains it.

## From Argon to Neon in seven weeks

The speed of the series is the other story. Argon appeared on September 30 in restricted access for advocates and opened to everyone on October 13. Carbon arrived on October 28 and Hydrogen on November 12, so Neon closes four launches in 49 days, a pace Google had not sustained in 2026. The [presentation of Gemini 4 Hydrogen](https://observatorioblockchain.com/en/ai/gemini-4-hydrogen/) marks the family's previous scale.

The rollout also changes target. Carbon required the Ultra plan for consumers and Neon enters Pro, an access cut that is worth as much as the price cut because the model reaches more people without going through the highest tier.

Neon appears in the paid API and in AI Studio. Also in Gemini for business and in Antigravity for Pro and Ultra subscribers, and in the Gemini app. Advocates in the Fairwind program will receive access in the coming days, and the company uses cybersecurity tests as a selling point. Inside Antigravity, Neon can distribute work among Hydrogen subagents in parallel, so the cheap model directs the expensive ones.

## The fine print of the launch

All the performance figures Google publishes are its own, including the cybersecurity ones: 76% on the hard level of AutoBaxBench and 90.5% on SEC-bench Pro. The technical sheet acknowledges hallucination as a general limitation of the model and sets the knowledge cutoff at March 2026, eight months earlier, so any recent data needs verification.

The fine print is joined by test selection. Google chooses 19 rows and wins 13 of the 16 comparable ones, but loses precisely on the two that matter most to a programmer. Neither defeat is serious and both matter when the selling point is coding. Among open models, rivals such as [Qwen4 Max from Alibaba](https://observatorioblockchain.com/en/ai/qwen4-max-alibaba-modelo-salvaguardas/) compete with the same promise of contained cost, so Neon's margin does not depend on Carbon alone.

The playing field has shifted to the bill. The question is no longer which model gets the best score, but how much each task that ends well costs and how many attempts it takes to reach that end.

## The bill as the battleground

The fourth Gemini 4 in less than two months is not fighting for first place on any table, and that is the change. Google competes on the cost of the completed task, a metric companies already use to decide which model they put behind each workflow, and it does so with a family of four models instead of a single product. The pressure comes from both sides: Anthropic is preparing its stock market listing with [Claude Fable 5.2 and Mythos 5.2](https://observatorioblockchain.com/en/ai/claude-fable-5-2-mythos-5-2-anthopic-ipo/) in reserve and OpenAI already has GPT-6.2 on the market.

We will have to wait until December 1 to see how long Neon's advantage over Carbon lasts, and for the first independent cost-per-task evaluations to know whether the 32% Google announces holds up. In the meantime, the sensible decision is to test the model with your own repository and count the retries it needs.
