On November 10, 2026, Google released Gemini 4 Carbon to everyone. It is the most capable model in the Gemini 4 family. On ChatX, Carbon has replaced Gemini 4 Argon since November 11, a model we had added only a few days earlier. Gemini 3.8 Flash stays on as the all-rounder and Gemini 3.5 Flash Lite as the budget option. Here’s what Carbon does better than Argon, where it doesn’t lead, and which of the three models is worth using for which tasks.
What sets Gemini 4 Carbon apart
Google announced Carbon on October 28, 2026, and at first gave it only to government agencies and selected security experts. According to the Gemini API changelog, it has been open to everyone since November 10. Google calls it its most capable model for long coding work, knowledge work with text, images, and spreadsheets, science questions, and defending against cyberattacks.
Three features matter in everyday use:
- Very long context: Input and output combined can run to as many as 1 million tokens. That’s enough for very long documents or extensive chat histories.
- Up-to-date knowledge: According to the model card, its knowledge extends to March 2026. For anything more recent, you can turn on web search on ChatX.
- Images as input: You can upload photos, screenshots, or charts and ask questions about them.
Carbon thinks before every reply. On ChatX, you control how thoroughly with the levels Low, Medium, and High. The default is Medium. This model has no “Off” level. ChatX doesn’t display the thinking steps themselves, but they do count toward billing.
Carbon instead of Argon: what changes
Argon and Carbon belong to the same family, and Google charges the same price per token for both. Google describes Argon as the faster variant for conversation and Carbon as the model that thinks more thoroughly. We compared the two on November 11 using the same tasks:
- Reasoning tasks: On a math problem with two trains, Argon got it wrong at the Low and Medium levels. Carbon arrived at the correct result at Medium.
- Code: Carbon wrote a PHP function faster than Argon and with fewer tokens (6.5 seconds instead of 10.4).
- Usage: On reasoning tasks, Carbon needs about three times as many tokens as Argon because it thinks longer. On simple tasks at the Low level, usage is about the same.
At the Low level, however, Carbon was slow and unsure on the math problem: it took more than a minute and corrected itself in the middle of its answer. That’s why Carbon runs at Medium by default on ChatX. If you had selected Argon before, you now get Carbon automatically. You don’t need to change anything.
Benchmarks: ahead almost everywhere
The figures come from Google’s model card dated November 10, 2026. For comparison, they’re shown next to Argon and three models you also know from ChatX.
| Test | Gemini 4 Carbon | Gemini 4 Argon | GPT-6 Astra | Claude Opus 5.5 | Claude Fable 5.1 |
|---|---|---|---|---|---|
| Overall score (Artificial Analysis Index v5.0) | 59 | 50 | 57 | 58 | 49 |
| Knowledge work (Vals Index v2.2) | 70.4% | 69.2% | 65.0% | 70.2% | not reported |
| Very long texts (GraphWalks, 256k to 1M) | 96.2% | 84.2% | 71.8% | 66.8% | 65.0% |
| Math (RiemannBench) | 80.0% | 76.0% | 72.0% | 69.6% | 65.6% |
| Working in the terminal (Terminal-Bench 4.0) | 65.2% | 57.1% | 59.1% | 59.6% | 52.0% |
| Coding (FrontierSWE 2) | 61.2% | 55.0% | 65.5% | 62.3% | 56.3% |
| Operating a computer (OSWorld-2.1) | 81.5% | 74.7% | 80.9% | 87.2% | 76.8% |
The lead is clearest with very long texts: for inputs between 256,000 and 1 million tokens, Carbon is twelve points ahead of Argon and more than 24 points ahead of the next model from another provider. That suggests Carbon copes well with a thick manual or many files at once.
Carbon isn’t ahead everywhere. For coding in FrontierSWE 2, GPT-6 Astra and Claude Opus 5.5 are in front, and Claude Opus 5.5 has a clear lead when it comes to operating a computer. There are two more things you should know: Google published these numbers itself. And Claude Fable 5.2, which has been available since November 9, isn’t in the table yet.
Carbon, 3.8 Flash, or Flash Lite?
There are still three Gemini models to choose from on ChatX. They differ mainly in thoroughness and price.
| Model | Designed for | Access |
|---|---|---|
| Gemini 4 Carbon | hard tasks, long documents, difficult code | free account |
| Gemini 3.8 Flash | everyday use: writing, analysis, translation, code | free account |
| Gemini 3.5 Flash Lite | quick, simple questions and short texts | no sign-up needed |
Gemini 3.8 Flash has been on the market since September 2, 2026. Google calls it “our most intelligent workhorse model”. For most questions, it’s the sensible choice. You can try Flash Lite right away without signing up. You’ll find all three on the ChatX Gemini page.
Our test: three tasks, three models
We gave the three Gemini models the same tasks: a math problem with two trains, a short email to a landlord, and a PHP function for validating an IBAN. Each model ran once at its default setting, which means Carbon ran at the Medium thinking level and 3.8 Flash at Low.
| Task | Gemini 4 Carbon | Gemini 3.8 Flash | Gemini 3.5 Flash Lite |
|---|---|---|---|
| Math problem | 20.8 seconds, about 4,000 tokens | 9.7 seconds, about 1,000 tokens | 2.1 seconds, about 370 tokens |
| Email to a landlord | 12.0 seconds, about 1,800 tokens | 3.4 seconds, about 170 tokens | 1.2 seconds, about 100 tokens |
| IBAN validation in PHP | 6.5 seconds, about 1,050 tokens | 4.7 seconds, about 440 tokens | 1.4 seconds, about 220 tokens |
On the math problem, Carbon and 3.8 Flash arrived at the correct result: the trains meet shortly after 10:10 a.m. Flash Lite miscalculated and said 10:29 a.m. Carbon also double-checked its result by verifying the answer. The email was usable from all three models. For tasks like that, Flash Lite is enough, and it was done in just over a second. At the Low level, Carbon wrote the email in 5 seconds using about 420 tokens.
The test is a spot check with one run each, and we didn’t execute the code. But it does show the pattern: Carbon takes longer and costs more, and in return it delivered the most thorough answer to the tricky math problem.
What a reply costs on ChatX
ChatX bills based on the actual costs charged by the provider. In our test, a reply from Carbon cost between about 1,050 and 4,000 tokens, compared with about 170 to 1,000 tokens for 3.8 Flash and about 100 to 370 tokens for Flash Lite. Depending on the task, that makes Carbon roughly two to ten times as expensive as 3.8 Flash. Most of that goes to thinking. At the Low level, usage drops noticeably on simple tasks, and at High it climbs further. In a longer chat, the conversation so far is added as input.
Important for planning: Google is offering Carbon at an introductory price until November 30, 2026. Starting December 1, Google will double the price. From then on, twice as much credit will be deducted on ChatX as well, which in our test would mean about 2,100 to 8,000 tokens per reply. For Gemini 3.8 Flash, the introductory price ends at the turn of the year.
Tips for getting started
- Flash first, then Carbon: Ask Gemini 3.8 Flash your question first. If the answer isn’t good enough, switch to Carbon in the same chat.
- Low for writing: Emails, summaries, and translations don’t need long thinking. In our test, this cut usage for the email from about 1,800 to about 420 tokens.
- Medium or High for hard cases: Tricky calculations, debugging code, or comparing several long documents are what Carbon was built for. On reasoning tasks, Low was the worse choice in our test.
- A little patience: Carbon thinks before it writes. In our test, it took 10 to 15 seconds for the first word to appear.
- Web search for current topics: Carbon’s knowledge ends in March 2026. For news, prices, or new products, turn on web search.
- Ask targeted questions about long documents: According to the benchmarks, this is where Carbon is strongest. Keep in mind, though, that the longer the input, the more tokens each message costs.