Under the post of Gemini 4 Neon on Hacker News, someone gave Carbon a very damaging definition: it is Argon with a temper that is more willing to burn thinking tokens. He added casually that Neon probably wouldn't be able to escape this problem either.
Another reply downstairs was more specific. He used Carbon, Hydrogen and the newly released Neon as demonstration pages. The token usage of the three models is three to four times that of Argon.
I think this complaint is only half true.
Google released the Gemini 4 Neon on November 18, which is the fourth Gemini model in a month. The official selling point is straightforward: 3 times faster than Carbon and 40% cheaper to complete the same task. Looking at the price list, Neon costs US$3 per million input tokens and US$15 per million output, which is half more expensive than the US$2 and US$10 currently listed on Carbon. The unit price is more expensive, but each question saves money. The only difference is how much thinking tokens are burned.
Once November 30th passes, the price list will turn over on its own.
Carbon's $2 and $10 are not regular prices. It's an introductory price that says it's available until November 30th, reverting to $4 and $20 on December 1st, and it's the same arrangement with Argon.
So I went to OpenRouter to compare prices today, and I saw that Carbon was cheaper than Neon; when I looked at it ten days later, Neon was actually 25% lower than the regular price of Carbon. Many price comparison pages on the Internet are snapshots of a certain day without writing the date. This is the easiest pitfall.
When budgeting, use 4 and 20.
The context is in the millions on both sides. Carbon's window is 1,048,576 tokens. Input, thinking, and output share this pool. The longer the thinking is written, the smaller the space left for formal output. Neon 官方给的是约 100 万上下文,最多能生成 100 万 token 的输出。
Also in high gear, Neon only takes 40% of the time
The data from Artificial Analysis is the most interesting to watch this time. When driven to xhigh, Neon has a smart index of 61 points, ranking third overall, 1 point higher than Carbon's high gear; it ranks second on the Vals Index.
But this 1 point is obtained in exchange for a higher thinking gear. Put both sides in the high setting, and the scores are both 60. The difference is all elsewhere: Neon took 14.8 minutes to complete the entire set of questions and spent $3.74, while Carbon took 36.1 minutes and spent $4.90.
The time difference is 2.4 times, which is close to what Google says is “3 times faster”. The money difference is only 24%, which is still far from "40% cheaper", and the $4.90 is based on Carbon's promotional price.
Recalculating based on the regular price will widen the gap.
Another calculus calculated that Neon costs $6.44 per question and Carbon’s regular price is $7.76, a saving of only 17%; Google’s own figure in the default high setting is 32%. Each of the three numbers is different, because the thinking gears and price assumptions are different. My judgment is that "40% cheaper" only holds true for a specific gear and specific task combination. It would be more optimistic to use it directly to write a budget. An estimate of 20% to 30% is more stable.
| project | Gemini 4 Neon | Gemini 4 Carbon |
|---|---|---|
| Unit price (per million tokens, input/output) | 3 / $15 | Introductory price $2 / $10 (until November 30), $4 / $20 starting December 1 |
| context | About 1 million, output up to 1 million | 1,048,576, shared by input, thinking, and output |
| AA Intelligence Index | 61(xhigh)/ 60(high) | 60(high) |
| Time taken to finish running in AA high gear | 14.8 minutes | 36.1 minutes |
| AA high total cost | $3.74 | $4.90 (introductory price) |
In the round of writing code, Carbon did not lose.
The code review tool Bito did a round of programming tests by itself, and the results were a bit interesting: Carbon's total score was 80.7%, Neon's 79.8%, and Carbon's score was 0.9 points higher. But the cost per session is $7.37 for Neon and $16.50 for Carbon, which is more than twice as expensive.
This is just a test of a single blog. There is no third-party reproduction. I only use it as a reference. But its direction is the same as AA: the capabilities of the two are almost close to each other, and what really separates them is time and billing. On the other hand, if your task is very sensitive to less than 1 point, such as a large-scale refactoring that requires manual rework if it fails once, and the labor of reworking once is far more expensive than the extra $9, Carbon still has reason to stay in the toolbox.
Most daily tasks are not included here.
The summary of thoughts on the screen has limited reference value.
There is another article in the same post that no one will talk about at the press conference. One reply said that the summary of thinking displayed to users by Gemini is very different from the original reasoning process of the model, and sometimes even leads people astray.
The original reasoning is not open to the public, so this cannot be verified. But it reminds me of a practical thing: If you want to know how much the model has thought about, it is useless to look at how carefully you write the summary. You need to look at the actual number of billed thinking tokens in the interface usage. Consider that tokens are charged based on the output price, which is $15 per million for Neon and $20 for Carbon.
This is also the other side of the "burning token" complaint. According to AA statistics, the Carbon high level outputs about 123,000 tokens per question, which is 29% less than the 173,000 tokens of the Claude Opus 5.5 max level. It is not the most talkative among cutting-edge models. It seems heavy only when compared with Argon, which costs about 62,000 per question. The people who complained used Argon as the frame of reference, so they felt that Carbon and Neon were squandering; if Opus was used as the frame of reference, the conclusion was reversed.
The shortcomings of both sides
The problem with Neon is threefold. The unit price is indeed not cheap, 3 and 15 are twice as expensive as Hydrogen's 1.5 and 7.5. Its "fastness" depends on the gear. It only scored 61 points when it was driven to xhigh. At this time, part of the speed advantage will be eaten away. Furthermore, it has only been online for one day, and there is not enough third-party data on stability under long tasks and multiple rounds of tool calls.
Carbon's problem is more immediate. Slower, 36.1 minutes vs. 14.8 minutes in the same gear. The promotional price will expire immediately, and the regular price of 4 and 20 from December is the most expensive among this generation of Gemini. Considering that the token usage is large and the window is shared with the output, feeding a long document task with hundreds of thousands of tokens will squeeze the space available for formal output.
Google itself probably knows that Carbon’s situation is a bit awkward.
Four models were released in one month, Argon, Carbon, Hydrogen, and Neon. Neon’s position is right in the middle of Hydrogen and Carbon. Once the promotion price passes on November 30, Carbon’s remaining reason will be less than one point of programming points.
Around the end of the month, two different usages
For batch processing tasks that are already running on Carbon, there is no need to rush before the end of the month. The promotional price is still available for about ten days, so save as much as you can. Before December 1st, switch the default model to Neon's high setting and run a round with your own mission set to see if you can lose points.
For interactive scenarios that require low latency, such as customer service conversations and completion in IDEs, Neon is not the optimal solution. The unit price of Hydrogen is only half that, and the Terminal-Bench 4.1 score announced by Google is 57%, which is enough for daily coding.
As long as there are long-term tasks that require manual cleaning after one failure and the accuracy must be compared, I will continue to keep Carbon and pay $4 and $20 honestly from December.
Related comparisons: "Claude Fable 5.2 and Gemini 4 Carbon: Carbon is four times cheaper per unit, but each question in the low range is more expensive", "Gemini 4 Carbon, Claude Opus 5.5, GPT-6 Astra: Carbon is one point ahead, most people can only look at the list this month", "Gemini 4 Hydrogen and Claude Sonnet 5.5: 57% and 70.6% are not the same paper at all"