Gemini 4 review roundup: Hydrogen, Argon or Carbon, and what each costs
Gemini 4 now has three models: Hydrogen for fast execution, Argon in the middle and Carbon for the hardest problems. The Artificial Analysis index, Google's figures and Bilibili creators' tests, how prices change and which group to pick.
In short
- Gemini 4 names its three models after elements: Hydrogen for everyday high-volume work, Argon as the first choice for most software engineering and knowledge work, Carbon for the hardest problems.
- List prices: Hydrogen is $1.50 in / $7.50 out per 1M tokens. Argon and Carbon are $2 / $10 until November 30, 2026 and $4 / $20 from December 1.
- On the Artificial Analysis Intelligence Index Carbon scores 60 and Hydrogen 51 at its default, one point above Argon's 50. Carbon is strong, and it uses many tokens and runs slowly.
- On MoeRouter all three cost far less in the google-proxy group than in default or vertex: choose the group before the model.
On this page
What the three models are#
Gemini 4 drops the Pro and Flash names for three elements[1]:
- Hydrogen, the fast execution engine, for everyday coding, agent workflows and multi-step data processing. Released on November 12, 2026, it is the first of the line open to everyone and succeeds Gemini 3.8 Flash[1][2].
- Argon, the first of the line to ship. Google says it remains the first choice for most software engineering and knowledge work[1][3].
- Carbon, Google's most capable reasoning model, for complex multi-step programming, long-horizon agents, cybersecurity and research. It went to allow-listed security vendors on October 28 and opened generally on November 10[1].
| Hydrogen | Argon | Carbon | |
|---|---|---|---|
| List price per 1M tokens (input / output) | $1.50 / $7.50, with no change announced | $2 / $10 until November 30, $4 / $20 from December 1 | Same as Argon |
| Context | 1,048,576 tokens, shared by input and output | Same | Same |
| Model name on MoeRouter | gemini-4-hydrogen |
gemini-4-argon |
gemini-4-carbon |
| Endpoints on MoeRouter | Gemini, Chat Completions | Gemini, Chat Completions, Responses, Messages | Same as Argon |
Prices and context are as Ray Liu's blog and a DEV Community post report them, both from the Gemini API pricing page and model documentation[2][3]. Hydrogen reads text, images, audio and video and writes text[2]. Note that from December 1 a token on Argon or Carbon lists at 2.67 times a token on Hydrogen[3].
What they cost on MoeRouter#
| Model | Group | Input | Output | Cache read |
|---|---|---|---|---|
gemini-4-hydrogen | google-proxy | CN¥0.72≈ US$0.108 | CN¥3.6≈ US$0.54 | CN¥0.036≈ US$0.0054 |
default | CN¥3.15≈ US$0.473 | CN¥15.75≈ US$2.36 | CN¥0.1575≈ US$0.0236 | |
vertex | CN¥3.15≈ US$0.473 | CN¥15.75≈ US$2.36 | CN¥0.1575≈ US$0.0236 | |
gemini-4-argon | google-proxy | CN¥0.96≈ US$0.144 | CN¥4.8≈ US$0.72 | CN¥0.048≈ US$0.0072 |
default | CN¥4.2≈ US$0.63 | CN¥21≈ US$3.15 | CN¥0.21≈ US$0.0315 | |
vertex | CN¥4.2≈ US$0.63 | CN¥21≈ US$3.15 | CN¥0.21≈ US$0.0315 | |
gemini-4-carbon | google-proxy | CN¥0.96≈ US$0.144 | CN¥4.8≈ US$0.72 | CN¥0.048≈ US$0.0072 |
default | CN¥4.2≈ US$0.63 | CN¥21≈ US$3.15 | CN¥0.21≈ US$0.0315 | |
vertex | CN¥4.2≈ US$0.63 | CN¥21≈ US$3.15 | CN¥0.21≈ US$0.0315 |
Reference prices per 1M tokens in Chinese yuan (CNY), the currency accounts are charged in, from the model square (). Dynamically priced models settle at the route that served each request; the model square is authoritative. The same model takes other routes in another group and costs differently there; a key's group is chosen when the key is created. About groups Amounts in USD are estimates, converted at 1 CNY = 0.15001 USD ().
Choose the group first, then the model. All three are in the default, vertex and google-proxy groups at prices several times apart; google-proxy is the cheapest at present. If you only use Gemini, create a key in the google-proxy group; when one group's routes are unstable, move to one of the other two. See about groups.
Benchmarks#
Independent#
The Artificial Analysis Intelligence Index, on one footing:
| Model and effort | Index |
|---|---|
| Carbon (high) | 60 |
| Hydrogen (xhigh) | 54 |
| Hydrogen (high, the default) | 51 |
| Argon | 50 |
| Hydrogen (medium) | 41 |
| Hydrogen (low) | 27 |
Hydrogen's four settings are the Artificial Analysis figures Ray Liu quotes, and Carbon's 60 is from the DEV Community post[2][3]. How to read them:
- Hydrogen at its default beats Argon, at over 60% less per evaluation task. Which is why, when Hydrogen launched, people at once asked what Argon is for now[1].
- Carbon is strong and slow. It takes 32.1 minutes per index task against a median of 13.2 for all models, and wrote 205 million output tokens against a median of 78.6 million[3].
- Carbon is not first. Since GPT-6.2 the order on this index is Claude Fable 5.2 65, GPT-6.2 Astra 62, Gemini 4 Carbon 60, Claude Opus 5.5 59[4].
- Going from high to xhigh gains Hydrogen 3 points for nearly twice the output tokens. Google's documentation says to start at high and keep xhigh for hard multi-step tasks[1].
Reported by Google#
- At its default high setting Hydrogen finishes tasks about 2.5 times faster than Gemini 3.8 Flash, and scores 57.0% on Terminal-Bench 4.1 against 18.2%. It is still behind Claude Sonnet 5.5's 63.9%[1][2].
- One step back worth knowing: against 3.8 Flash, Hydrogen refuses harmless or borderline requests 5.57 points more often, most of all on multimodal prompts[1].
- On a benchmark of complex visual documents published by Google and Artificial Analysis, Claude Fable 5.2 scores 57% and Carbon 54%, at about $31.97 and about $6.29 per task: the gap in cost is far wider than the gap in score[1].
Hands-on: what Bilibili creators found#
程序员阿江-Relakkes tested two of them. Carbon: Argon had ranked 11th on his ModelDial board, and Gemini 4 Carbon went straight to 2nd[5]. Hydrogen: 80 points on his private question bank, 11th place, with 83 on back-end work, enough to write most back-end code; it can do front ends, with taste a little behind Sonnet 5.5 and about level with DeepSeek. On a ten-thousand-line legacy ERP codebase, four hours through the API ran up a bill of over ¥100, and through Antigravity it scored 77.9. His overall impression: much better than 3.8 Flash, and Google's coding model is finally usable (our translation)[6].
kate人不错 had Hydrogen do nine tasks in Antigravity at High effort: 3D modelling, a creative web page, an advertisement animation, a Godot game and more. The plainest comparison is speed: a mooncake gift-box task that took Carbon three hours took Hydrogen 23 minutes. Her verdict: pleasant surprises in speed, multimodal analysis and mechanism logic, and shortcomings in modelling detail and some structural design[7].
实验式 set Argon against Claude Opus 5.5 on 3D tasks: Argon still trails Opus 5.5 overall, though it is pleasant to use, steady and fast[8].
Carbon's bill#
Carbon lists at the same price as Argon, but a rate card is not a bill. It spends more tokens than Argon exploring and checking, thinking tokens are billed too, and the real cost of a task on Carbon is usually far above what the rate card suggests. The community's experience agrees: worth the wait, and hard on the quota[1]. Even Hydrogen ran up over ¥100 on one long task through the API[6]; token use on long tasks easily exceeds what one expects.
One way to spend less is to divide the work: Carbon or Argon plans and untangles, Hydrogen edits code and runs tests. Google's Orchestrator in Antigravity does this, and the internal figures Google gives are a success rate close to Carbon alone with 24% less quota and 19% more speed. Those figures are Google's own; nobody has reproduced them yet[1].
Which to pick#
The DEV Community post puts it directly: start on Hydrogen and move up only when it fails your own tests[3].
| If you | Then |
|---|---|
| Do everyday coding, back-end endpoints, tables and multi-step data processing | Hydrogen at high |
| Weigh designs across modules or have complex program problems | Argon |
| Trace root causes across services, plan large refactors, do security or research | Carbon |
| Ran fixed pipelines on Gemini 3.8 Flash | Run a few real inputs on Hydrogen, then see which can drop to low or medium |
Gemini 4 is good for
- Tasks that read images, audio, video and long documents
- Back-end code and agents that call many tools (Hydrogen)
- The hardest reasoning and long-horizon work (Carbon)
Less suited to
- Front ends where visual taste matters most
- Long tasks on a fixed budget: Carbon's token use is hard to predict
- Interactive use that needs steady latency: one Carbon task can run half an hour
For comparison with other makers, see the Claude Opus 5.5 review and the GPT-6.2 Sol and Astra review.
How to call them#
Gemini CLI and the Google GenAI SDK use the native Gemini endpoint at https://moerouter.com. The OpenAI SDK works too:
import os
from openai import OpenAI
client = OpenAI(
base_url="https://moerouter.com/v1",
api_key=os.environ["MOEROUTER_API_KEY"],
)
reply = client.chat.completions.create(
model="gemini-4-hydrogen",
messages=[{"role": "user", "content": "Find every call to deprecated_api in this repository and propose a replacement"}],
)
print(reply.choices[0].message.content)
A plain HTTP request to the native Gemini endpoint:
curl -sS "https://moerouter.com/v1beta/models/gemini-4-hydrogen:generateContent" \
-H "x-goog-api-key: $MOEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{"contents":[{"parts":[{"text":"Introduce yourself in one sentence"}]}]}'
gemini-4-argon and gemini-4-carbon also answer on the Messages and Responses endpoints, so they can be set in Claude Code and Codex; gemini-4-hydrogen is offered on the native Gemini endpoint and Chat Completions only. The docs are in Chinese for now.
Frequently asked questions#
What is the difference between Gemini 4 Hydrogen, Argon and Carbon?#
Hydrogen is the fast execution model that succeeds Gemini 3.8 Flash, for everyday high-volume work. Argon is the balanced model for most software engineering and knowledge work. Carbon is the most capable reasoning model, uses the most tokens and is the slowest: keep it for the hardest problems.
How much does the Gemini 4 API cost?#
List prices per 1M tokens: Hydrogen $1.50 input and $7.50 output; Argon and Carbon $2 / $10 until November 30, 2026 and $4 / $20 from December 1. MoeRouter's reference prices are in the table above and differ by group.
Which MoeRouter group should I use for Gemini 4?#
At present google-proxy is the cheapest; default and vertex cost several times as much. If you only use Gemini, create a key in the google-proxy group. Prices change: the table and the model square are authoritative.
Hydrogen scores above Argon. Is Argon still worth using?#
On the Artificial Analysis index Hydrogen scores 51 at its default and Argon 50, and Hydrogen is cheaper. Google still positions Argon for when Hydrogen is not quite enough and Carbon costs too much; comparing on your own tasks is the reliable way to decide.
Is Gemini 4 Carbon the strongest model now?#
No. It was first on the Artificial Analysis index when it launched, and Claude Fable 5.2 (65) and GPT-6.2 Astra (62) have since passed its 60. Its strength is value: on a benchmark of complex visual documents it is 3 points behind Fable 5.2 at about a fifth of the cost per task.
Can I use Gemini 4 in Claude Code or Codex?#
gemini-4-argon and gemini-4-carbon, yes: on MoeRouter they support the Messages and Responses endpoints. gemini-4-hydrogen is on the native Gemini endpoint and Chat Completions only, which suits Gemini CLI and clients that speak the OpenAI format.
Sources
The test results and figures quoted in this post come from the public sources below and belong to their authors; each quotation is marked where it appears.
- Gemini三个模型怎么选:Hydrogen、Argon、Carbon适用场景梳理 (in Chinese)
- Gemini 4 Hydrogen 介紹:接棒 3.8 Flash,和 Carbon、Argon 怎麼選? (in Chinese)
- Gemini 4 Carbon vs Hydrogen vs Argon: Which to Use (2026)
- OpenAI launches GPT-6.2 Astra and Sol with lower API prices and some benchmark gains
- 【全国首发】Gemini 4 Carbon跑分测试,谷歌这次真站起来了吗?! (video, in Chinese)
- 【实测】Gemini 4 Hydrogen 发布!谷歌编程终于能用了吗?! (video, in Chinese)
- Gemini 4 Hydrogen 实测 9 个任务,用掉 Pro 26% 周额度,速度惊喜,质量还不错 (video, in Chinese)
- 能否挑战天花板? Gemini 4 Argon 对比 Claude Opus 5.5 实测 (video, in Chinese)
Disclosure: MoeRouter sells API access to the models discussed here. The test results in this post come from the third-party sources cited; prices and endpoint details come from MoeRouter's own data.