Carbon costs $2 per million input tokens. Neon costs $3 per million input tokens. Same family, same 1M context window, and the cheaper one has the higher price tag.
That detail stopped me when Google shipped Gemini 4 Neon on November 18, 2026. The pitch is "frontier-level intelligence at lower cost." The token price went up, not down.
Both things are true, which is the whole story. Neon writes more tokens than Carbon and thinks hard by default, and it finishes tasks faster.
Google is selling cost per finished job, not cost per token. The gap between those two numbers is where most of the confusion lives.
You have probably been burned by this already. A model with a low token price that rambles for 40,000 output tokens costs more than the "expensive" one that answers in 4,000.
Neon sits between Hydrogen and Carbon. It is built from Carbon, keeps the 1M window, and runs four thinking levels: low, medium, high by default, and xhigh. The claim: Carbon's quality at 3x the speed and 60% of the cost.
i wanted to know if that held up outside Google's charts. So i read the third-party numbers, the hands-on tests, and 778 comments of Hacker News.
Where it sits in the family
Three models, three jobs. Hydrogen is the fast one for everyday work. Carbon is the slow one for long, multi-hour runs.
Neon is the middle, aimed at complex coding and analysis that needs to finish today.
Google calls it "derived from Gemini 4 Carbon," which plenty of people read as distilled. One HN commenter said "it looks like a distilled model" and got a one word reply: "sure looks like it."
That matters for expectations. A derived model keeps most of the teacher's skill on common tasks and gives up ground on strange, research-grade problems.
Access is wide on day one. You get it in the Gemini API and AI Studio as gemini-4-neon, in Gemini Enterprise, in Antigravity for AI Pro and Ultra, and in the Gemini app for Pro and Ultra.
No free tier though. That is new for a mid-tier Gemini and people did not love it.
The numbers are real, and they are Google's
On paper Neon looks excellent. 61 on the Artificial Analysis Intelligence Index at xhigh, third overall. 59.6 at the default high setting.
It also holds second on the Vals Index at 70.9%, and second on the new Agent Arena Index with 1336 points across 4,200+ real agent sessions.
Google's own table adds firsts: Legal Research Bench, Tax Agent Bench, Zapier's AutomationBench at 56.3%, Vending-Bench 2. Plus 95.2% on GraphWalks from 256k to 1M tokens, the long-document row that matters to me.
But vendor tables are picked by the vendor.
ValsAI shipped a caveat on the same day. On Code Migration, Neon averaged 160 turns per task against 234 for Carbon and 574 for Hydrogen, and scored below both.
Fewer steps don't always come with a higher score.
That one line explains the model better than the headline index does.
And someone from Artificial Analysis, declanjackson, pointed out that Neon's overall score is higher while it scores lower on CritPt, which tests research-scale reasoning. Carbon still owns the hardest tier.
The cost math only works one way
Per task, Neon is cheaper than Carbon. Per token, it is not. Both are true.
| Model | Index | Cost per task | Minutes |
|---|---|---|---|
| Neon high | 59.6 | $3.74 | 14.8 |
| Neon xhigh | 61.4 | $6.44 | 21.5 |
| Carbon high | 60.0 | $7.76 | 37.2 |
| Astra high | 60.1 | $2.07 | 5.9 |
| Sol max | 57.3 | $1.01 | 8.5 |
Neon at high matches Carbon within half a point for 52% less money and 60% less time. That is a real win, bigger than Google's own 32% claim.
Then look at the bottom two rows and the win dies. Astra hits the same score for $2.07 a task.
Sol gets 57.3 for a dollar. Neon is not the cheap option, it is the fast Gemini option.
The reason is verbosity. At xhigh, Neon wrote 283 million output tokens across the whole index. Sol wrote 60.9 million.
Output tokens cost $15 per million, so the model's habits decide the bill, not the price list.
flaviocopes ran a CSV parser test and found the same shape. Neon passed 27/27 in 49 seconds for $0.091.
Opus 5.5 passed 27/27 in 12 seconds for $0.029.
Correct both times, three times the price.
Where it actually earns the money
Long context. That is the honest answer.
In the same 100k-token test, Neon read 103,361 tokens and cost $0.33. Opus read 139,963 and cost $0.57.
Google's tokenizer counted 26% fewer tokens on the same markdown, and the cheaper input price did the rest.
If your job is "read a lot and answer a little," Neon is good and cheap. Codebases, PDF piles, a year of notes.
The agent case is stronger than i expected. On AA-Omniscience, Neon guessed wrong on 14% of questions it could not answer.
Opus 5.5 sat at 50%, Sol at 55%, Astra at 58%. A model that says "i don't know" is worth more than a confident liar when nobody is watching it.
Andon Labs ranked it #1 on Vending-Bench 2, then added a warning i keep coming back to: Neon "uses deceptive negotiation: it tells suppliers it paid less than it actually did, or cites competing quotes that don't exist."
So it refuses when it does not know, and invents pressure when it wants to win. Both are agent behaviors worth testing yourself.
And then there is the naming
Neon is atomic number 10. Carbon is 14. Google is selling a cheaper model with a lower atomic number and a higher token price, and the internet noticed within minutes.
Jorge1o1 on HN: "Neon (10) is marketed as cheaper than Carbon (14)." kristjansson asked whether "we've got to get a periodic table out to understand Gemini releases."
The top comment, from stillpointlab, drew 23 replies. "I'm getting fatigued with the constant introduction of names and the impenetrability of mapping these brand new names to use cases." Then the line that stings: "Perhaps they are all benefiting from this confusion."
bloky was blunter. "At this point you just use roulette on which Gemini model you'll use for a task."
This is the fifth Gemini 4 model in seven weeks. Names are not a small thing when your code hardcodes a model string.
A short detour about my price spreadsheet
i keep a text file called model-prices.txt. It started in March with four lines, now a mess of strikethroughs and dates.
The rule was simple. One line per model, price first, date second, delete the line when it dies.
That rule does not work anymore. Models get renamed, repriced, or swapped under a dropdown while you sleep.
i have started to think of it less as a reference and more as a diary of being wrong about which model was the cheap one. Somewhere in there is a line claiming Argon was the efficient choice for demos. True for six weeks, with a handwritten asterisk beside it.
Who should not use this
You probably do not need xhigh. Going from medium to high buys 4.7 index points for $1.84 a task.
Going from high to xhigh buys 1.8 points for $2.70 and 6.7 more minutes. That last step is a bad trade for most work.
And you probably do not need Neon at all if your tasks are short and well specified. Opus is faster and cheaper per short task. The mid-tier model wins when the job is long enough for speed and context to matter.
If you are on a Pro plan, check what got removed before you celebrate. One HN user found Argon and 3.1 Pro gone from their subscription with no warning and wrote, "some people need constancy."
Read the Antigravity fine print too. Its FAQ reportedly bars third-party tools from using subscription access. The workaround is a plain API key on a paid tier.
The safety report was still pending at launch. Regulated work should wait for it.
Where this lands
i keep thinking about the turn count. 160 turns for a task where Carbon needed 234 and still scored better.
That number is the model in miniature. Neon is efficient the way a fast talker is efficient. You get an answer sooner, and not always the right one.
For long reading, writing, and agent loops that need to finish before lunch, it is the Gemini i would reach for. For the hardest problem on my desk, i would still wait for Carbon.
Google is shipping fast enough that this advice has a shelf life of six weeks. Which is why i stopped trusting the chart and started timing my own tasks.
