Google releases Gemini 4 Neon: 61 on the Artificial Analysis index at $3/$15 per million tokens
Google has released Gemini 4 Neon, a mid-tier model between the expensive Carbon and the cheap Hydrogen. It costs $3 per million input tokens and $15 per million output tokens, and scores 61 on the independent Artificial Analysis quality index (Carbon scores 60, Hydrogen 54). The confusing part: Carbon carries a discount of $2/$10 until November 30, so for now Neon costs more than both of its neighbors. From December 1, Carbon returns to $4/$20, so budget your project at that price, not the promotional one.
AI-processed from Simon Willison; edited by Hamidun News
On November 18, Google introduced Gemini 4 Neon, the third model in the Gemini 4 lineup. It costs $3 per million input tokens and $15 per million output tokens (a token is roughly a chunk of a word, and it is what the API bills by). It scored 61 on the Artificial Analysis Index, an independent composite index of model quality.
For comparison, the top-tier Carbon scores 60 and the entry-level Hydrogen 54. The problem is that prices across the lineup currently look illogical, and if you choose a model by today's price tag, you can underestimate your project's budget. The same day, Simon Willison released llm-gemini 0.
38, a version of his plugin for the llm utility that can work with Neon.
What the model is and where it fits in the lineup
Gemini 4 used to consist of Hydrogen, Argon and Carbon. In the Neon announcement, Google showed a new lineup without Argon: Carbon, Neon and Hydrogen. Willison also dropped Argon from his comparison grid.
Here are the numbers from Google's slide (price per million tokens, input / output):
- Carbon: $4 / $20, 60 points on the Artificial Analysis index, speed 2 out of 5. Aimed at the heaviest long-running projects and deep research.
- Neon: $3 / $15, 61 points, speed 3.5 out of 5. Aimed at deep reasoning in development, knowledge work and science.
- Hydrogen: $1.50 / $7.50, 54 points, speed 5 out of 5. Aimed at fast everyday coding and agentic scenarios.
Neon and Hydrogen have reasoning levels low, medium, high and xhigh; Carbon has only low, medium and high. Neon is available in the Gemini API through Google AI Studio (paid), in Gemini for Business, in the Gemini app with AI Pro and Ultra subscriptions, and paid in Google Antigravity. Hydrogen, by contrast, is open on all AI Studio tiers.
Why the prices are confusing
Carbon's $4/$20 price is not in effect right now: until November 30 it has a promotional price of $2/$10. So today the most expensive of the three is Neon ($3/$15), followed by Carbon ($2/$10), then Hydrogen ($1.50/$7.50). From December 1, the order returns to the logical one: Carbon, Neon, Hydrogen. Willison recently wrote that Gemini 4 pricing is convoluted, and in his words, Neon made the situation even worse.
A practical example: say your service gets a 10 000-token response from the model. On Neon that is about 15 cents, on Hydrogen about 7.5 cents, and on Carbon 10 cents right now, but 20 cents after December 1. If you build the promotional Carbon price into your calculations, your real bill will double in a week or two.
What the pelican test showed
Willison traditionally asked the models to draw an SVG of a pelican riding a bicycle at three reasoning levels. This is not a benchmark but a single attempt per mode, though the tokens and cost it reports are accurate.
At the high level:
- Carbon: 20 305 output tokens, 20.31 cents, 4 minutes 11 seconds;
- Neon: 9 691 tokens, 14.55 cents, 1 minute 11 seconds;
- Hydrogen: 5 311 tokens, 3.99 cents, 19 seconds.
An important detail: Carbon was billed at the promotional price of $10 per million output tokens. At the regular $20, the same run would have cost about 40 cents, nearly three times more than Neon. Most of the bill is tokens spent on reasoning, and Neon spends noticeably fewer of them than Carbon. So comparing models by price per token is not enough: what matters is the cost of a finished result.
On image quality, the order is as follows: Carbon has more detail (a sailor's cap, a basket of fish, a waterfront), Neon draws cleanly and neatly but more simply, and Hydrogen is the most minimalist. In every case it is a recognizable pelican on a bicycle.
Willison also ran Neon at the extra level in Google Antigravity on a three.js task: an animation in which Yonkers rises as a giant mech and strides across Manhattan, Queens and Brooklyn using OpenStreetMap data. He published the result as a demo page and a video.
What developers should do
If you already use llm, update the plugin (llm install -U llm-gemini) and look for the model under the ID gemini-4-neon, as in Willison's grid. After that, it makes sense to run your own tasks on Neon and Hydrogen rather than pelicans: for simple code and agentic loops, Hydrogen may deliver the quality you need at a fraction of the price, while Neon suits cases that need stronger reasoning without Carbon's price.
The limitations are obvious: a one-point gap between Neon and Carbon on the Artificial Analysis index is nearly a tie, and the pelican test shows only the ratio of price, time and token volume in a single episode. It is best to make the budget decision after December 1, when Carbon's price tag stops being promotional.
- Nov 19, 2026Anthropic goes public and raises $81.3 billion, beating SpaceX's $75 billion record
- Nov 19, 2026Amazon and Google's Anthropic Stakes Valued at $308 Billion in IPO at $83 a Share
- Nov 19, 2026Google's Gemini 4 Neon runs twice as fast as Carbon and costs 17% less per task
- Nov 19, 2026A GitHub Script Catches App Store and Google Play Rejection Reasons Before Review: 542 Stars in a Day
Key Questions Answered
What the model is and where it fits in the lineup?
Gemini 4 used to consist of Hydrogen, Argon and Carbon. In the Neon announcement, Google showed a new lineup without Argon: Carbon, Neon and Hydrogen. Willison also dropped Argon from his comparison grid. Here are the numbers from Google's slide (price per million tokens, input / output): Carbon: $4 / $20, 60 points on the Artificial Analysis index, speed 2 out of 5. Aimed at the heaviest long-running projects and deep research. Neon: $3 / $15, 61 points, speed 3.5 out of 5.
Why the prices are confusing?
Carbon's $4/$20 price is not in effect right now: until November 30 it has a promotional price of $2/$10. So today the most expensive of the three is Neon ($3/$15), followed by Carbon ($2/$10), then Hydrogen ($1.50/$7.50). From December 1, the order returns to the logical one: Carbon, Neon, Hydrogen. Willison recently wrote that Gemini 4 pricing is convoluted, and in his words, Neon made the situation even worse. A practical example: say your service gets a 10 000-token response from the model.
What the pelican test showed?
Willison traditionally asked the models to draw an SVG of a pelican riding a bicycle at three reasoning levels. This is not a benchmark but a single attempt per mode, though the tokens and cost it reports are accurate. At the high level: Carbon: 20 305 output tokens, 20.31 cents, 4 minutes 11 seconds; Neon: 9 691 tokens, 14.55 cents, 1 minute 11 seconds; Hydrogen: 5 311 tokens, 3.99 cents, 19 seconds. An important detail: Carbon was billed at the promotional price of $10 per million output tokens.
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.