Gemini 4 Carbon: Almost Fable 5.2 for a Quarter of the Price, but Only Through November

November 10, 2026, evening

Less than a day has passed since Anthropic launched Claude Fable 5.2 and took back first place on the leaderboards, and Google is already posting: Gemini 4 Carbon, the model that until now was open only to a closed group, mostly cyber-defense organizations, is available to everyone starting today.

In my article on Fable 5.2, I wrote that in my opinion Anthropic's launch was a leaderboard move, meant to take first place back from Carbon. I expected Google wouldn't wait long. It took 21 hours.

After a day of reading every test that was published, my conclusion: Carbon is the best-value model at the top right now. Just not for the agent that answers your customers, and only until the end of the month.

Gemini 4 Carbon is Google's most powerful model, and on hard tasks it comes close to Claude Fable 5.2, for a small fraction of the cost. In a real-world test of fixing 105 bugs, it scored 57.3 to Fable's 58, and cost $31 against $140. But there are three catches: the low price is a launch price that ends on November 30, a model that thinks this much isn't suited to quick conversation, and Google itself recommends sticking with Gemini 4 Argon for most work.

So my opinion: Carbon shouldn't replace the model that handles your day-to-day work. But it's worth trying on one big, hard task, preferably before the end of the month.

In short

Gemini 4 Carbon has been available to everyone since November 10, 2026, through Google's API (the direct connection for developers, paid by usage) and the Google AI Ultra subscription. Until November 30 it costs $2 per million input tokens and $10 per million output tokens. On December 1 the price doubles to $4 and $20, the same as Claude Opus 5.5.
It's at the top: first on Cursor's coding benchmark and in Arena's user vote, and on the overall index a little behind Fable 5.2 and roughly level with Opus 5.5.
It's cheap per task: in the bug test, about a quarter of Fable's cost, at almost the same score. Even at the regular price, a task on it costs about a third of a task on Fable at its maximum level.
Where it's weak: documents and office work, interface design, and real-time conversation. At the high thinking level, more than a minute passes before it starts answering.
My opinion: Do you have one big, hard task with a clear finish? It's worth trying on Carbon now. An agent that answers customers and prepares reports? There's no reason to switch its model.

What Gemini 4 Carbon is, and what it costs

Carbon is the model Google places above Gemini 4 Argon, for the longest and hardest tasks: a tricky bug, scientific research, cybersecurity, and hours of uninterrupted work. Google announced it on October 28 and at first opened it only to cyber-defense organizations. On November 10 it opened to developers, businesses and Ultra subscribers. Argon took the same path: announced to a closed group on September 30, and opened to everyone 13 days later.

$2 / $10Price per million input and output tokens, until November 30. From December 1: $4 / $20, exactly like Opus 5.5. Fable 5.2 costs $8 / $40

One millionTokens in the context window, the amount of text the model holds in mind at once. The same price across the whole window, with no surcharge for long text

Ultra onlyIn the Gemini app and in AI Mode in Search. Google AI Pro subscribers get Argon. In the API it's open to anyone who pays

Sources: Google's two launch announcements, the Gemini API price list and the support page on models by plan (checked on November 11, 2026). If the term is new to you, we explained what a token is in a separate article.

To understand Carbon, you need to read three sentences Google wrote itself, in three different documents:

What Google writes about Carbon

The launch announcement, the model card and the developer docs

The model card: Carbon "is based on Gemini 4 Argon"
The developer docs: both share the same context window, the same tools and the same per-token price. "The main difference is how many thinking tokens each model spends per turn"
The launch announcement: Argon "remains the ideal choice for the majority of software engineering and knowledge work," while Carbon is built for tasks that test the limits of a model's intelligence

What it means in practice

How I read it

By Google's own explanation, Carbon is Argon given more time to think before each step
The price per token is the same, but a task on Carbon uses more tokens, so it costs more than the same task on Argon
The gap in capability is big: 60 against 50 on the Artificial Analysis index, and 57.3 against 40.7 in the bug test below

How smart Gemini 4 Carbon is, against Fable 5.2 and Opus 5.5

Every benchmark lists a "thinking level" (Low, Medium, High, Max): a setting that decides how long the model thinks before it answers. More thinking usually gives a better answer, but a slower and more expensive one.

Gemini 4 Carbon vs. Claude Fable 5.2 and Claude Opus 5.5 in independent measurements

Benchmark	Gemini 4 Carbon	Claude Fable 5.2	Claude Opus 5.5
Overall indexes
Overall indexArtificial Analysis Intelligence Index	60	65	59
Professional workVals Index	70.5%	71.3%	70.1%
User vote, textArena Text	1526	Not measured yet	1506
Code and agent work
Real coding tasksCursorBench 4.0	60.4%	60.0%	57.8%
Commands an agent runs on its own on a serverTerminal-Bench 4.0 (AA)	65.2%	59.6%	59.6%
Business automationAutomationBench-AA	77.8%	76.4%	69.5%
Long coding tasksFrontierCode 2.0	40.6%	48.8%	46.4%
Office work judged by expertsAA-Briefcase, Elo	1611	1785	1801
Science and knowledge
Expert-level questionsHLE-Diamond	69.1%	80.2%	67.9%
Science on the command lineTerminal-Bench-Science	64.8%	73.3%	59.0%

Top score in the row. Sources: Artificial Analysis (index version 5.0.1), Vals AI, Arena and Cursor, as of November 11, 2026. Carbon was measured at its high thinking level (High), while Fable and Opus were measured at their maximum level (except on Vals and Arena). Carbon's Arena rating is still preliminary.

What the table shows:

In real code and automation, it's very strong. On Cursor's coding tasks, on commands an agent runs on a server, and on business automation, it's first in these measurements, ahead of Fable 5.2. The picture isn't uniform: on Vals it's fifth on that same commands benchmark, and in Arena's agent ranking it's second, behind Opus 5.5.
In documents and office work, it's far behind Claude. When experts compared presentations, spreadsheets and reports, they clearly preferred the ones from Opus 5.5 and Fable 5.2. An agent's weekly report is exactly this kind of work. In writing long code from scratch, too, Claude leads by 6 to 8 points.
In science and the hardest questions, Fable 5.2 still leads by 8 to 11 points. Against Opus 5.5, Carbon is equal or better.

One more number, for anyone whose agent answers questions: on the Artificial Analysis knowledge test, when Carbon didn't know the answer, it made one up 11% of the time, against 50% for Opus 5.5. At the medium thinking level, only 4%. For an agent that answers customers from your own documents, the instructions and the documents matter more than this figure. But in research, where the model answers from its own knowledge, it's a real advantage.

What developers saw in the first hours

The most convincing test I found is Paweł Huryn's. He planted 105 bugs back into two real codebases and had six models find and fix as many as they could:

105 bugs in two codebases: score and cost per run

Model	Score	Cost
Claude Fable 5.2Max	58	$139.65
Gemini 4 CarbonHigh	57.3	$31.32
Claude Sonnet 5.5Max	51.3	$154.03
GPT-6 AstraMax	45	$33.03
GPT-6.1 SolMax	44.3	$5.88
Gemini 4 ArgonHigh	40.7	$7.16

Source: Paweł Huryn on X, November 10, 2026, launch day. His test and his calculation, not ours.

Paweł Huryn's post on X, November 10, 2026

99% of Fable 5.2's score, at about a quarter of the cost. In his words, "Nobody should be laughing at Google anymore."

Argon, the same base model with less thinking, scored 40.7 on the same test, which means Carbon added 16.6 points. And GPT-6.1 Sol scored 44.3 for under six dollars. If you're after the lowest price per point, Sol is still the pick.

Another test, by Atomic Agent, gave the models a code library with 13 bugs. They all fixed everything, so the only difference between them was time and money. At the high thinking level, Carbon finished faster than Fable and cost about a sixth as much: $1.29 against $7.70.

The team at Every, which tests every new model on real work, summed it up as "I'll use it sometimes." They gave Carbon a task backlog. It identified six tasks as ready, then did 15, including work someone else was already doing. After four and a half hours, they stopped it. One of the developers there wrote that Fable remains his first choice, because Carbon can produce huge files, while Fable keeps the code more modular and organized.

Post by Dan Shipper of Every on X, November 10, 2026

Their advice: "Give it one big job with a clear finish line."

In 3D and visual builds, several developers who had early access wrote, each independently, that Google has fixed the "laziness" of its earlier models, and that the model writes thousands of lines in one go. Simon Willison ran his 3D robot test on both Carbon and Fable 5.2, and you can see Carbon's result and Fable's result on his blog. On the other hand, almost everyone agrees that in interface design, meaning screens, buttons and landing pages, Claude still does nicer work.

kimmonismus's post and video on X, November 10, 2026

13 coding tasks in Antigravity, Google's development tool. Among them, a geared model of the solar system that you can run forward and backward with a single button. His impression of Carbon, not ours.

What Gemini 4 Carbon really costs

What decides the bill isn't the price per token, but how many tokens the model uses to finish a task. Artificial Analysis publishes an average cost per task on its index for every model:

Average cost per task on the Artificial Analysis index

Claude Fable 5.2 · Max · score 65

$26.54

Claude Opus 5.5 · Max · score 59

$10.04

Claude Fable 5.2 · High · score 62

$9.08

Gemini 4 Carbon · High · score 60

$7.76

Gemini 4 Carbon · High · until November 30

$3.88

GPT-6.1 Sol · Max · score 55

$1.27

Source: Artificial Analysis, as of November 11, 2026. Artificial Analysis calculates Carbon's cost at the regular price, $4 and $20. The "until November 30" row is our calculation: at the launch price both input and output are half off, so the cost per task is too.

On CursorBench, where Cursor publishes a score and a cost for each thinking level, the picture is similar. Carbon at high scores 60.4% at $9.22 per task, and Fable 5.2 at max scores 60.0% at $32.79. Roughly the same result, for less than a third of the price.

The three ratios in this article, a quarter, a third and a seventh, are all correct. The difference is which price and which test we're talking about:

At the launch price, $2 and $10

Almost Fable, at a seventh of the price.

On the Artificial Analysis index: about $3.88 per task, against $26.54 for Fable 5.2 and $10.04 for Opus 5.5 at their maximum level. In Huryn's bug test: $31 against $140, at almost the same score.

At the regular price, $4 and $20

Still cheaper than both, but the gap narrows.

The price per token is the same as Opus 5.5, but Carbon uses fewer tokens to finish a task, so it still comes out about 23% cheaper, and costs about a third of Fable, with both at their maximum level. Against Fable at high, which scores 62 at $9.08, the gap almost disappears.

Lower the thinking level to save money?

To medium, yes. To low, no.

Going from high to medium saves money: on Artificial Analysis, from $7.76 to $6.59 per task, for a 2-point drop in score. But at the low level the model takes more steps and rereads all the material at every step, so it costs more than at medium: $7.19. On CursorBench the gap is even bigger: $11.89 at low, against $8.49 at medium. The medium level, the default, is the best value.

A monthly subscription, not the API

Ultra only, and no published limit.

In the Gemini app, Carbon is available only on the Google AI Ultra plan. Google doesn't publish a separate limit for each model; it only writes that "more advanced models and higher thinking levels require more of your usage." On Hacker News, a user with an Ultra subscription wrote that the limit barely covered a single project.

Sources: Artificial Analysis, Cursor, Google's API price list and help page, Paweł Huryn and the launch thread on Hacker News. The cost calculation at the launch price is ours.

Where Carbon is weak

The problem that came up most is speed. At the high thinking level, 67 seconds pass before an answer starts coming out, according to Artificial Analysis's measurement. At medium, 23 seconds. One Hacker News user called it "completely unusable at the moment," and another replied that it's a model for long tasks, not for chat.

To be fair, this isn't a weakness unique to Carbon. In the same measurement, Opus 5.5 and Fable 5.2 at their maximum level start answering only after several minutes, and Carbon finishes a whole task faster than both. That's the price of any model that thinks a lot. On a half-hour task, a minute at the start doesn't matter. In a chat with a customer, a minute of silence feels like a glitch, and there you need a lighter model or a low thinking level.

What works

Benchmarks and developer tests, November 10 and 11

Long solo work: fixing existing code, commands on a server, and automation
3D and large builds in one go
Makes things up less than Opus 5.5 when it doesn't know the answer
Research in tax, law and finance: first on three Vals benchmarks

What works less well

Benchmarks, developer complaints and my impression from the tests

More than a minute before it starts answering at the high level: not suited to real-time chat with a customer
Doesn't always know when to stop: keeps building beyond what was asked
Documents, presentations and spreadsheets: far behind Opus 5.5 and Fable 5.2
Interface design: "boxes within boxes," as Every put it

One more thing that matters for businesses: Google writes in its price list that with paid use of the API, the data you send is not used to improve its products. On Google Cloud's enterprise platform, Carbon is for now open only to "verified enterprise customers in select regions."

Who Gemini 4 Carbon is for, and who it isn't

Worth trying now

One long, hard task with a clear finish: a bug no model has solved, moving an old site to a new system, research
Automation that runs in the background, with nobody waiting for an answer
Research where it matters that the model doesn't make things up
Anything you'd send to Fable 5.2 if the price weren't stopping you

Check first

The date: on December 1 the price doubles
The thinking level: medium is the best value, and low doesn't save money
Your plan: in the app it's Ultra only, and on Pro you'll get Argon
That you've set limits on the scope of work, so it doesn't keep building things you didn't ask for

Stay with what you have

An agent that answers customers in chat or on WhatsApp, where every second is felt
Routine work: website updates, replies to inquiries, short tasks
Documents, presentations and reports, where Opus 5.5 leads by a wide margin
Designing a website or a landing page

For a business owner whose agent already works, the practical takeaway: there's nothing to switch this week. An agent that handles inquiries, updates a website and prepares a weekly report does lots of short tasks, and needs to answer within seconds and write a well-organized document. For that, a lighter model or a low thinking level is better, and Opus 5.5 still does the reports better. And if price is your first consideration, you'll find that GPT-6.1 Sol is much cheaper.

Carbon is worth testing if you have one big project your current agent is stuck on, such as cleaning up an old website or analyzing a full year of sales data. In that case, ask whoever runs your agent to run the project on Carbon, with a spending cap, while it's still November.

Platin Digital is an OpenAI Select Partner, and we build our agents with the models and tools that best fit each business, from OpenAI, Anthropic, Google and others.

Who's first now

Twenty-one hours after the Fable 5.2 launch, Google opened Carbon to everyone. I don't know if the timing was a response or planned in advance: Argon also took 13 days from announcement to general availability, exactly like Carbon. Either way, today each company leads on different leaderboards:

Ranking	First today	Carbon
Artificial Analysis, overall index	Fable 5.2, 65	Fourth, 60
Vals Index	Fable 5.2, 71.3%	Third, 70.5%
CursorBench 4.0	Gemini 4 Carbon, 60.4%	First
Arena, text	Gemini 4 Carbon, 1526	First, preliminary score
Arena, web development	Fable 5.2, 1836	Fourth, 1768
Arena, agents	Opus 5.5	Second

Sources: Artificial Analysis, Vals AI, Cursor and Arena, as of November 11, 2026.

In my opinion, Anthropic leads in capability, and Google offers the best ratio of capability to price at the top.

Where Gemini 4 Carbon stands in Sigma AI Ranking

Until yesterday, Carbon appeared in Sigma AI Ranking (our ranking, which combines several open sources into one score) with a preliminary score of 97.8. That score was inflated, because the model had no public price, so cost didn't go into the calculation. Now it has a price, and enough measurements to get a full score:

Sigma AI Ranking: weighted score, as of November 11, 2026

Gemini 4 Carbon

91.0

Claude Fable 5.2 · preliminary score

91.0

Claude Opus 5.5

89.0

Gemini 4 Argon

86.8

GPT-6.1 Sol

85.3

Source: Platin Digital's Sigma AI Ranking, which combines open sources (Epoch AI, Arena, Terminal-Bench, and OpenRouter for prices).

Carbon is now first in the ranking among models with a full score. Fable 5.2 got the same score, 91.0, but it has been measured by only three sources so far, so its score is preliminary. Two things are worth knowing about Carbon's number: the ranking calculates cost at the current price, $2 and $10, so on December 1 its score will drop by about a point. And it's still missing Epoch AI's general capability index, so for now the ranking estimates that part from the other measurements.

Further reading: Claude Fable 5.2: smarter than Opus 5.5, Gemini 4 Argon and why we built Sigma AI Ranking.

Frequently asked questions

What is Gemini 4 Carbon?

Gemini 4 Carbon is Google's most powerful model, one tier above Gemini 4 Argon. It's based on Argon, and according to Google the main difference between them is how much thinking it does before each step. It's built for long, hard tasks, such as fixing tricky bugs, research and cybersecurity. Google announced it on October 28, 2026 and opened it to everyone on November 10. It has a context window of one million tokens.

How much does Gemini 4 Carbon cost?

In the API: $2 per million input tokens and $10 per million output tokens, until November 30, 2026. On December 1 the price rises to $4 and $20, the same as Claude Opus 5.5 and half the price of Claude Fable 5.2. In the Gemini app it's available only on the Google AI Ultra plan.

Is Gemini 4 Carbon better than Claude Fable 5.2?

Not at everything. On overall indexes and science benchmarks, Fable 5.2 leads: 65 against 60 on the Artificial Analysis index. On real coding tasks and automation, Carbon is equal or better, and in a test of 105 bugs it scored 57.3 against 58, at about a quarter of the cost. In documents and interface design, Claude leads.

What's the difference between Gemini 4 Carbon and Gemini 4 Argon?

According to Google, both have the same context window, the same tools and the same price per token. The difference is the amount of thinking: Carbon spends more thinking tokens, so it's smarter, slower and more expensive per task. Google recommends Argon for most work, and Carbon only for the hardest tasks.

Should I switch my agent's model to Gemini 4 Carbon?

Usually not. An agent that answers customers needs to reply within seconds, and Carbon at the high thinking level takes more than a minute just to start answering. It's suited to long tasks that run in the background, like a coding project, automation or research. If you have a task like that, it's worth trying Carbon on it before December 1, while the price is still half.