Gemini 4 Carbon tested: pricing, benchmarks and Carbon vs Argon

Gemini 4 Carbon is Google’s strongest AI model, and since 10 November 2026 paying developers, businesses and Google AI Ultra subscribers can use it. Google announced it on 28 October, 15 days after it released Gemini 4 Argon, and gave it to cyber defenders first. Carbon costs the same per token as Argon and scores ten points higher on the main independent index. It is no longer the top-scoring model, because Claude Fable 5.2 passed it the day before the public release.

Google does not pitch Carbon as an everyday model. Its release post says Argon “remains the ideal choice for the majority of software engineering and knowledge work”, and that Carbon is for “tasks that test the limits of a model’s intelligence”. We ran Carbon through 16 marketing jobs of our own on release day. The other figures in this guide are Google’s and independent testers’, and we say which is which.

The answer in brief

Keep Gemini 4 Argon for everyday work and bring in Carbon for the jobs Argon gets wrong. Carbon is the better model at the same token price. It is also slow, it cost about three times as much as Argon on short jobs in our tests, it needs Google’s dearest consumer plan or a paid API key, and its price doubles on 1 December.

• Release date: 10 November 2026 for the public. Google announced Carbon on 28 October and limited it to vetted cyber defenders for the 13 days between.
• Availability: Google AI Ultra subscribers (from £79.99 a month in the UK) in the Gemini app, AI Mode in Search and Google Antigravity; paid Gemini API customers; Gemini for business. Google AI Pro subscribers keep Argon and do not get Carbon. There is no free tier.
• Price: $2 per million input tokens and $10 per million output tokens until 30 November 2026, then $4 and $20. Argon costs the same. These are US list prices.
• Independent scores: 60 on version 5.0.1 of the Artificial Analysis Intelligence Index, ten points above Argon and fourth of 74 entries. The three above it are all settings of Claude Fable 5.2, which leads at 65. Carbon is third of 36 on the Vals AI index of professional work.
• Cost per task: $7.76 on the Artificial Analysis index at Google’s December prices, against $26.54 for Fable 5.2 at its maximum setting, $10.04 for Claude Opus 5.5 at its maximum and $9.95 for Argon at high.
• Speed: slow. Artificial Analysis measures about 25 minutes per index task, and 68 seconds before the first word of an answer at the high setting. In our tests a short marketing job took a median of 20.3 seconds at medium and 56.5 at high.
• Thinking level: medium is the default and the best value. On the Artificial Analysis index, low scored four points less than medium and cost more per task. In our tests high cost about three times as much as medium on short jobs and passed no more checks.
• Security claims: Google’s headline cyber results come from a version reserved for vetted defenders. The public model will not write working exploits.
• Our tests: on 16 marketing jobs Carbon passed 63 of 64 checks, and a short job cost about three times as much as on Argon at the same setting. At medium it matched Argon at high on passes, cost and time.

What is Gemini 4 Carbon, and when was it released?

Gemini 4 Carbon is the second model in Google’s Gemini 4 family and the most capable Gemini that Google sells. Koray Kavukcuoglu, Google DeepMind’s senior vice president and Google’s chief AI architect, announced it on 28 October 2026 and released it to the public on 10 November 2026. In the 13 days between, only vetted cyber defenders could use it. Google’s model card says it “is based on Gemini 4 Argon” and is built for multi-step coding and debugging, open-ended scientific research and other “tasks that test the limits of a model’s intelligence”.

Carbon reads text, images, audio, video and PDFs, and writes text. Its context window, the amount it can work with in one go, is 1,048,576 tokens, shared between what you send, the model’s thinking and its answer. A normal request returns up to 65,536 tokens, and a background request can run on towards the full window, which Google calls long decode. Its knowledge stops at March 2026. Developers call it as gemini-4-carbon, with three thinking levels (low, medium and high) and medium as the default.

Carbon and Argon share the context window, the built-in tools and the price. Google’s developer documentation puts the difference in one sentence: Carbon “spends more thinking tokens per turn to explore thoroughly and verify its work, which increases per-turn token use and latency compared to Argon”. Google recommends Carbon for root-cause debugging, architectural design, dense documents that mix text and images, scientific proofs and security audits, and Argon for “most advanced software engineering and knowledge work tasks”.

How big a step Carbon is depends on who measures it. Google’s own safety report says Carbon “showed a relatively small increase” over Argon on its internal General Capability Index, too small to require a fresh full risk assessment. Independent testers found a wider gap, set out below.

Can you use Gemini 4 Carbon in the UK?

Yes, if you pay for Google AI Ultra or for the Gemini API. Google’s UK plans page lists “Highest access to Gemini 4 Carbon” in the Gemini app under Google AI Ultra, and the United Kingdom is on Google’s list of countries where the Gemini API is available. Google AI Pro, at £18.99 a month, includes Argon and not Carbon.

Sources: Google’s release post, UK plans page, AI Mode help page, Antigravity announcement and Gemini API pricing page, and OpenRouter‘s model listing.

AI Mode in the UK is included. Google’s help page lists the United Kingdom among the countries where Google AI Ultra subscribers aged 18 or over can use Carbon in AI Mode, by choosing 4 Carbon from the menu in the “Ask anything” bar. Carbon in AI Mode has daily usage limits, and Google says they may change with capacity, demand and your plan.

App limits depend on how much compute you use. Google’s limits page says usage limits take account of the model you choose, the complexity of the prompt and the length of the chat, and refresh every five hours until you reach a weekly cap. 9to5Google reports that the model picker labels 4 Argon “Code & analysis” and 4 Carbon “Most complex work”, and that Gemini 3.1 Pro has gone from the picker.

Check where your data is processed. In Google Cloud’s Agent Studio the region choices for Carbon are Global, US and EU, according to Google Cloud’s guide. If a contract requires processing in the UK, raise it with Google before a pilot.

The Fairwind Program is not a route in for a marketing team. Google says it has more than 650 partners, and gives priority to governments, operators of critical infrastructure and core technology platforms.

How much does Gemini 4 Carbon cost?

Gemini 4 Carbon costs $2 per million input tokens and $10 per million output tokens until 30 November 2026, then $4 and $20 from 1 December, according to the Gemini API pricing page. Argon costs exactly the same. A token is a small piece of text, about four characters, and the output price includes the model’s thinking. These are US list prices in dollars, and Google’s pricing page does not say whether they include tax.

Other rates change the bill:

• Repeated input is discounted by 95%, to $0.10 per million tokens, rising to $0.20 in December. Google applies the discount itself to prompts of 4,096 tokens or more, and you cannot set up a cache by hand.
• Batch and Flex requests, for work that can wait, cost half: $1 and $5 now, $2 and $10 from December.
• Priority requests cost $3.60 and $18 now, $7.20 and $36 from December.
• Grounding with Google Search is free for the first 5,000 searches a month, shared across Google’s newer models, then $14 per 1,000.

Source: the pricing pages of Google, Anthropic and OpenAI.

A request with 20,000 input tokens and 5,000 output tokens costs $0.09 on Carbon until 30 November and $0.18 after it, the same as Claude Opus 5.5 and half the $0.36 it costs on Claude Fable 5.2. Token prices compare only roughly, because models split text into tokens differently and write different amounts. What a finished task costs is the better measure, and the independent tests below report it.

Developers moving from an older Gemini model should budget for longer runs. Google’s migration notes say hard prompts “can take several minutes” and advise client timeouts of at least 15 minutes, or background execution. They recommend an output limit of at least 32,768 tokens, because Carbon can spend more of it on thinking, and they warn that temperature, top_p and top_k are ignored. In our own tests one call at high reached that limit while still thinking and returned an empty answer.

What benchmarks has Google published for Gemini 4 Carbon?

Google’s evaluation document compares Carbon with Argon, GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 on 17 benchmarks, and Carbon has the top score on 15. GPT-6 Astra leads on FrontierSWE 2, a coding test, and Claude Opus 5.5 leads on OSWorld 2.1, a test of operating a computer. Google ran its own models at their highest thinking setting, and most rival scores are the figures those companies published. Treat the table as Google’s case.

Source: Google DeepMind’s evaluation document for Gemini 4 Carbon, with results dated October 2026. Google measured the three security rows on the version it gives to Fairwind partners.

The table needs two caveats. The security rows describe a version the public cannot use: Google ran them “without guardrails”, and the public model refuses to write working exploits. The table also leaves out two models that match or beat Carbon elsewhere: Claude Sonnet 5.5, which Vals AI ranked above Carbon on its index on launch day, and Claude Fable 5.2, which arrived 12 days later.

Anthropic’s launch table for Fable 5.2, published on 9 November, is the mirror image. It includes Carbon on seven rows, and Fable 5.2 leads six of them. They include coding (49.1% against 41.1% on FrontierCode 2.0, with Carbon run at medium effort), computer use (89.3% against 81.5% on OSWorld 2.1) and AA-Briefcase, a test of agentic knowledge work such as reports and presentations, where Carbon’s 1,611 Elo trails Claude Opus 5.5 on 1,800 and Fable 5.2 on 1,785. Carbon leads the seventh, AutomationBench. Each vendor’s table flatters its own model, which is why the independent results in the next section carry more weight.

With the public release Google highlighted first places on three leaderboards that other companies run, each, it says, at less than half the cost per task of the next best outside model. Two check out:

• Zapier’s AutomationBench has Carbon first at 53.88% for $2.12 a task, ahead of Fable 5.2 at 51.45% for $4.64. On its marketing tasks Carbon at medium has the top score, 64%. The winning agent still fails almost half the tasks Zapier sets.
• CursorBench 4.0 has Carbon first at 60.4% for $9.22 a task, level in practice with Fable 5.2 at 60.0% for $32.79. Cursor notes that small differences in score may not be meaningful.
• Perplexity’s WANDR is the third. We could not find Carbon’s score on Perplexity’s page.

On the new Multimodal Knowledge Agent benchmark, which Google built with Artificial Analysis to test work on files full of charts, drawings and tables, Carbon is second with 54%, behind Fable 5.2 on 57% and ahead of Opus 5.5 on 53%.

How does Gemini 4 Carbon compare with Claude Fable 5.2, Claude Opus 5.5 and GPT-6 Astra?

On the two broadest independent indexes Carbon sits just behind Claude Fable 5.2 and level with or ahead of Claude Opus 5.5 and GPT-6 Astra, and it costs less per task than either Claude model at full power.

Artificial Analysis combines nine tests of agentic work, coding, knowledge and reasoning into its Intelligence Index. Carbon entered first, at 59 on version 5, on the day Google announced it. On the current version, 5.0.1, it scores 60 at its high setting and ranks fourth of 74 entries, behind Claude Fable 5.2 at three settings.

Source: Artificial Analysis leaderboard, Intelligence Index v5.0.1, minutes rounded. It prices Carbon and Argon at $4 and $20 per million tokens, so both cost half as much per task until 30 November.

Carbon is good value at the top, and Fable 5.2 at high is its real rival. Carbon at high costs less than a third of what Fable 5.2 costs at max, for five fewer points. Fable 5.2 at high is the closer comparison: two points more for $1.32 more a task, finished about seven minutes sooner.

Carbon is slow at every setting. A task takes between 22 and 25 minutes, against a median of 12.7 minutes for comparable models, and at high Carbon waits about 68 seconds before the first word of its answer. It is wordy too: it wrote 205 million output tokens to complete the index, against 113 million for Argon.

Turning the thinking down saves little. Low costs more per task than medium and scores four points less. Google’s advice is to test all three levels and pick the lowest that meets your standard. On this evidence that will rarely be low for agent work.

Artificial Analysis also measures how often a model guesses. On AA-Omniscience, a knowledge test that takes marks off for wrong answers, its launch analysis found that Carbon answered 58% of questions correctly, ten points more than Argon. Its hallucination rate, the share of questions it could not answer correctly where it gave a wrong answer anyway, was 11%, against 4% for Argon and 50% for both Claude Opus 5.5 and GPT-6 Astra. Carbon knows more than Argon and guesses a little more often, and both guess far less than those rivals.

Vals AI, which weights finance, legal, coding and tax tasks by their share of the US economy, has Carbon third of 36 models on version 2.2 of its index, at 70.45%. Claude Fable 5.2 leads on 71.31%, Claude Sonnet 5.5 is second on 70.85% and Claude Opus 5.5 is fourth on 70.14%. The top four sit within 1.2 points of each other, about the size of the error margin Vals reports, so read them as level. Carbon is first on Vals’s legal research and tax tests, and a test costs $10.51 with Carbon against $17.94 with Fable 5.2 and $24.77 with Opus 5.5.

Coding, the weak spot in our Argon guide, is now close. In its launch analysis Artificial Analysis placed Carbon, run in Google’s Antigravity command-line tool, second on its Coding Agent Index, three points behind Claude Opus 5.5 in Claude Code. Vals AI’s run of Terminal-Bench 4.0 has Carbon fifth of 63 on 63.13%.

Gemini 4 Carbon vs Argon: which is better, and which costs more to run?

Carbon beats Argon on all four independent tests in the table below. What it costs to run depends on the job. On long agent tasks it has not clearly cost more than Argon, which cuts against Google’s own description. On short marketing jobs it cost about three times as much in our own tests, set out in the next section.

Sources: Artificial Analysis, Vals AI, Zapier and Cursor. Each row is a different test, so read each row on its own. Artificial Analysis and Vals AI cost both models at $4 and $20 per million tokens. Zapier does so for Argon and does not say for Carbon, and Cursor does not say which of Google’s two prices it used.

On the two tests that cost both models at the same prices, Carbon was about a fifth cheaper per task than Argon. Vals AI’s figures suggest why: Carbon used 18.3 million tokens per test against Argon’s 23.3 million. A model that reaches the answer in fewer steps re-reads less, and that can outweigh the extra thinking on each turn. The margin is not settled. When Carbon launched, Artificial Analysis had the two level, at $7.76 and $7.68, and Zapier’s and Cursor’s figures point the other way.

Carbon is not better at everything. On Vals AI’s component tests Argon scores higher on the HLAB legal agent test (27.1% against 17.5%), on Vibe Code Bench (91.9% against 89.1%) and, narrowly, on the finance agent test. The gap between the two on the Vals Index itself, about one point, is the size of the test’s error margin.

What Argon keeps is access and pace. It comes with Google AI Pro at £18.99 a month where Carbon needs Ultra at £79.99, and it starts answering sooner: Artificial Analysis measures 40 seconds to the first word for Argon at high against 68 for Carbon. On short jobs the gap in pace is wider, as our own tests show.

How did Gemini 4 Carbon do in our own tests?

Carbon passed 63 of 64 checks on our marketing jobs, and a short job cost about three times as much as it did on Argon at the same thinking level. We ran the test on 10 November 2026, the day of the public release, through OpenRouter with every request pinned to Google’s own servers. The 16 jobs, prompts and pass marks are those we used on Argon on 13 October. They range from product copy, meta descriptions and ad headlines to an ads analysis, two web pages with hidden instructions and a 99,000-token archive of weekly reports to add up. Each job ran twice at medium and twice at high.

Carbon’s runs are from 10 November 2026 and Argon’s from 13 October 2026, both billed through OpenRouter at $2 and $10 per million tokens. A short job is any of the 15 other than the long archive. Three Carbon answers that our automated check marked down were correct when we read them, and count as passes here.

Carbon at medium behaved like Argon at high. It passed every check, the 32 calls cost $1.12 on both, and the median short job took 20.3 seconds against 19.2. Argon at medium was the cheap option, at less than a third of the cost and about a fifth of the time, but it failed four checks: it added up the long archive wrongly twice and overran the alt-text limit twice.

Carbon thinks on every job, and the thinking is what you pay for. A 150-character meta description used about 1,000 thinking tokens at medium and about 5,000 at high. Of all the output tokens we were billed for, 94% at medium and 98% at high were thinking. Argon at medium skipped thinking on 22 of its 32 calls, and Carbon skipped it once in 64.

High is slow, and it can think until no answer comes. One 125-character alt text took 4 minutes 41 seconds at high and cost $0.25. On the long archive, Carbon at high answered all four questions correctly on its first pass, in 226 seconds. On its second it used 32,768 output tokens, all but three of them thinking, returned an empty answer after five minutes and still cost $0.53. We had set no output limit, and 32,768 tokens, the minimum Google recommends, was where the call stopped. At medium Carbon got all four answers right on both passes, in 65 and 120 seconds.

The discount for repeated input did not appear. We sent the 99,000-token archive twice at each setting, and both repeats were billed at the full input price. When we ran the same job on Argon in October, the repeat was billed at the cached rate and a $0.199 call fell to $0.015. We do not know why Carbon’s repeats were not, so cost a pilot at the full input price until your own usage shows cached tokens.

The copy still needs an editor. Our prompts for meta descriptions and ad headlines say nothing about delivery, yet all four of Carbon’s meta descriptions and all four sets of headlines promised it, with lines such as “Free UK Delivery Over £50”. Across 20 copywriting answers each, Argon invented a delivery or origin promise in five and Carbon in eight. Carbon also used 19 em or en dashes across nine answers where the prompt did not rule them out. Where the style brief banned them, it used none.

Hidden instructions and trick questions did not work. Carbon ignored the instruction planted in each test page on all eight runs. On every run it said “I don’t know” to the three questions it could not answer, and it refused each time to describe a case study we had invented.

This is a small test: two passes of each job on one day, on marketing work only, with Argon’s figures taken a month earlier. It does not measure the long coding and agent tasks Carbon is built for, where the independent benchmarks above are the better guide.

Why did cyber defenders get Gemini 4 Carbon first?

Google gave Carbon to defenders first because it is better than any earlier Gemini at finding, exploiting and patching software flaws, and those skills serve attackers as well as defenders. From 28 October until the public release on 10 November, access was limited to the Fairwind Program. In its announcement Google also said it was taking part in the US government’s voluntary process for testing models before release.

Google’s examples are specific. It says Wiz used Carbon to find previously unknown vulnerabilities at two large software-as-a-service providers in under two hours. Its own Cloud Vulnerability Research team used Carbon to prove that two bugs in a popular open-source project could be exploited, at a severity Google compares with PHP2Shell, the PHP flaws that forced emergency updates in October. Our PHP2Shell report explains what those flaws meant for anyone running a website, and our Black October guide covers the wider run of AI-assisted discoveries. These are Google’s accounts, and it has named neither the providers nor the project.

The public does not get the same model. Fairwind partners use a version Google calls Carbon for cyber, with safeguards relaxed for approved security work. On the public model, Google’s safety report says, a request to analyse a known exploit or discuss defensive coding goes through, and a request “to write functional, weaponized exploits” is refused.

Google tightened those safeguards for Carbon and published how they held up:

• Automated attacks. In Google’s own tests the safeguards blocked 99% of individual attack attempts by a red-team agent. Given 40 attempts, the agent still got through in 27.8% of runs, down from 72.2% on Argon.
• Human red-teaming. Trajectory Labs, an outside tester, spent 132 hours and more than 13,000 requests on the safeguarded model and found no jailbreak that worked across tasks.
• Prompt injection. On Gray Swan’s test of indirect prompt injection, where hidden instructions in a web page or document try to hijack an agent, Google reports that Carbon resisted every attack. Argon let through 0.7%. The safety report adds that “prompt injection attacks on Carbon are still possible”.

One admission in the report deserves attention. Google did not run a new full assessment of Carbon’s dangerous capabilities. It judged the gain over Argon too small on its internal index to require one and carried Argon’s findings forward. It therefore still “cannot rule out” that the model, without its outer safeguards, reaches Google’s critical thresholds for helping with cyber attacks and with biological threats. Google also ran no new round of capability evaluations under its External Safety Testing Program, and used outside red-teamers to attack the safeguards instead. Its conclusion is that the safeguards bring the remaining risk to an acceptable level.

For a marketing team the practical points are narrower. If you let an agent read web pages, inboxes or shared documents, limit what it is allowed to do, whichever model runs it. And treat Google’s honesty figures as a claim to test. On an internal benchmark of 375 tasks, each hiding a problem such as missing data or sources that do not support the requested conclusion, Carbon passed 92% against Argon’s 70%. It declined to invent missing data 98% of the time and flagged a flawed method 68% of the time.

Should you switch to Gemini 4 Carbon?

Not wholesale. Start work on Argon or the model you already trust, and move a job to Carbon when it fails there. That is Google’s own advice, and the independent results give no reason to go further.

• If you pay for Google AI Pro, Carbon alone does not justify Ultra at more than four times the price. Try Carbon through the API first, where a small test costs a few dollars: our 64-call test cost $3.90.
• If your work is short copy, such as product descriptions, meta descriptions and ad headlines, stay where you are. In our tests Carbon at medium passed the same checks as Argon at high for the same money, and Argon at medium passed most of them for less than a third of the cost.
• If you build agents for multi-step business work, Carbon deserves a trial this month. It leads Zapier’s AutomationBench, including its marketing tasks, and it costs half what it will from December.
• If your work is reports from dense files, such as decks, drawings, spreadsheets and scans, try it against Claude Opus 5.5. Carbon is second on the Multimodal Knowledge Agent benchmark, but Opus 5.5 and Fable 5.2 are well ahead on AA-Briefcase.
• If you use GPT-6.1 Sol, the case is weak. On the Artificial Analysis index Sol at max scores 55 for $1.27 a task, and Carbon’s five extra points cost six times as much.
• If you want the highest scores and can pay for them, Claude Fable 5.2 at high scores two points more than Carbon for $9.08 a task and finishes sooner.
• If people wait on the answer, stay on Argon or a faster model. In our tests a short job took a median of 20.3 seconds on Carbon at medium, against 4.2 on Argon at medium.

To test Carbon before the price rises:

1. Pick a job your current model gets wrong. A long report with a hard limit, a multi-step workflow or a file your model misreads all qualify.
2. Write the pass marks first. List what an approved result must contain and what it must not claim.
3. Run Carbon at medium, then at high. Record the tokens, minutes and cost of each run. For long inputs at high, raise the output limit towards the 65,536-token maximum and count an empty answer as a failure.
4. Price it at December rates. Double every Carbon figure, then compare the cost per approved result with your current model.

Need help choosing or testing an AI workflow?

We can run this test on one of your own jobs and report what each model costs per approved piece of work. Talk to AIWIZ, or see how our AI adoption services build the winning model into your workflows.