Newsartificial-intelligence / Hardware / Products

Gemini 4 Carbon Is Live: Benchmarks, Zero-Days and the Nov 30 Price Jump

Google's most intelligent model opened to every paying developer on 10 November. We checked the benchmarks, the zero-day claims, the $2-to-$4 price change on 30 November and where Carbon still loses, then looked at what it means for your PC.

Words / Pixlboard

Reading time / 14 minutes

Gemini 4 Carbon Is Live: Benchmarks, Zero-Days and the Nov 30 Price Jump
News / Pixlboard JournalImage / Pixlboard

Gemini 4 Carbon went generally available on 10 November 2026. Google's strongest model is now open to every paying developer at $2 per million input tokens, a price that doubles after 30 November. Here is what it really does, where it still loses, and what to do before the deadline.

Narrowly first on the index, looser with facts, and on a price clock that ends 30 November

Direct Answer: Gemini 4 Carbon is Google DeepMind's most intelligent model. It was announced on 28 October 2026 and opened up on 10 November to every paying developer and enterprise through the Gemini API in Google AI Studio and Gemini for business, and to Google AI Ultra subscribers in the Gemini app, AI Mode in Search and Google Antigravity. It costs $2 per million input tokens and $10 per million output tokens until 30 November 2026, then $4 and $20. On the Artificial Analysis Intelligence Index v5 it scores 59, one point ahead of Claude Opus 5.5 (58) and two ahead of GPT-6 Astra (57), and Google says it leads 15 of the 17 benchmarks in its own comparison. The headline is cybersecurity: 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, plus a Google-reported case where Wiz used it to find zero-day vulnerabilities in two major SaaS providers in under two hours. The catch: its measured hallucination rate rose from 4% on Argon to 11%, it trails Claude Opus 5.5 on computer-use and coding-agent tests, and almost every cyber number comes from Google itself. Carbon is cloud-only, so no graphics card can run it. What your PC needs is the hardware to keep up with the agents it drives.


The 30-Second Version

  • What it is: Google's newest frontier model, built for cybersecurity, long-running agents and hard software engineering. Google says its own teams have seen it run for days on a single task.
  • Who can use it today: paying developers (Google AI Studio), enterprises (Gemini for business), and Google AI Ultra subscribers (Gemini app, AI Mode in Search, Antigravity). Vetted cyber defenders have had it since the week of 28 October through Google's Fairwind Program.
  • What it costs: $2 input and $10 output per million tokens until 30 November. After that, $4 and $20.
  • Where it wins: Google's cyber benchmarks, 256K to 1M-token long-context retrieval (GraphWalks 96.2%), and most agent tests in Google's own table.
  • Where it loses: hallucination rate (11% against Argon's 4%), computer use (OSWorld-2.1) and Anthropic's coding agents.
  • Do you need a new GPU? No. It is a closed-weights model that runs in Google's cloud.

How We Got Here: The Gemini 4 Timeline

DateWhat happened
30 Sep 2026Gemini 4 Argon announced, with Fairwind defenders first
13 Oct 2026Argon opens more broadly through the API and paid consumer products
28 Oct 2026Carbon announced, with access starting for cyber defenders in Fairwind
10 Nov 2026Carbon becomes available to all paying developers and enterprises
30 Nov 2026Introductory pricing ends, $2 and $10 becomes $4 and $20

Who Can Use Carbon Right Now

WhereWho gets itStatus
Gemini API in Google AI StudioAll paying developersAvailable since 10 Nov
Gemini for business (API, model drop-down, Antigravity, CodeMender)EnterprisesAvailable since 10 Nov
Gemini app, AI Mode in Search, Google AntigravityGoogle AI Ultra subscribersRolling out from 10 Nov
Fairwind ProgramVetted cyber defenders, by applicationSince the week of 28 Oct, widening gradually
Antigravity top usage tierScientists and researchers, free for six months, by applicationRolling applications

Early write-ups from 29 October to 6 November, including Lindy and NeuralTrust, still describe Carbon as gated to government testers and Fairwind members. They were right on the day they were published. Google's own 10 November announcement is the source for the table above, so check Google AI Studio for what your account can actually select.


The Price Clock: $2 Today, $4 on 1 December

TierInput per 1M tokensOutput per 1M tokens
Introductory, through 30 Nov 2026$2$10
From 1 Dec 2026$4$20

Google also lists a 95% discount on cached input. VentureBeat's worked example makes the deadline concrete: 10 million uncached input tokens plus 1 million output tokens costs $30 at the introductory rate and $60 afterwards. Same job, double the bill.

Two things to keep straight before you budget:

  1. Same price as Argon does not mean same cost per task. Carbon spends more tokens than Argon. Artificial Analysis cost-per-task figures reported by NeuralTrust and Lindy put Carbon at $7.76 per task on the Intelligence Index, against $7.68 for Argon, $10.04 for Claude Opus 5.5 and $6.25 for GPT-6 Astra. Measure cost per finished job in your own workflow.
  2. The end date is Google's footnote, not a guarantee of what comes next. Re-check Google's pricing page before you commit a budget.

Benchmarks: Carbon vs Argon vs the Best Rival

BenchmarkWhat it testsCarbonArgonBest rival listedLeader
AA Intelligence Index v5.0Overall intelligence5950Opus 5.5 58, GPT-6 Astra 57Carbon
SEC-bench ProFinding vulnerabilities88.2%83.4%GPT-6 Astra 85.4%Carbon
ExploitGymProving a bug is exploitable54.8%42.0%GPT-6 Astra 42.4%Carbon
CWE-benchPatching vulnerabilities71.0%68.0%GPT-6 Astra 68.0%, Grok 4.7 68.0%Carbon
GraphWalks 256K to 1MLong-context retrieval96.2%84.2%GPT-6 Astra 71.8%Carbon
Terminal-Bench 4.0Terminal agent tasks65.2%57.1%Opus 5.5 59.6%Carbon
Agents' Last Exam v1.1General agent tasks77.4%69.0%GPT-6 Astra 73.1%Carbon
Vending-Bench 2Simulated business balance$16,484$13,718GPT-6 Astra $15,515Carbon
Vals Index v2.2Mixed professional tasks70.4%69.2%Opus 5.5 70.2%Carbon, by 0.2
SWE-sweepSoftware engineering9.6%4.3%GPT-6 Astra 9.5%Carbon, by 0.1
FrontierSWE 2Software engineering61.2%55.0%GPT-6 Astra 65.5%GPT-6 Astra
OSWorld-2.1 (offline subset)Computer use81.5%74.7%Opus 5.5 87.2%Opus 5.5

Source: Google's launch comparison, reproduced in full below. Rows marked "by 0.2" and "by 0.1" are margins too thin to call without uncertainty estimates, and Google published none.

Google's full Gemini 4 Carbon evaluation table comparing Carbon, Argon, GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 across general capabilities, agentic coding, computer use, knowledge work, multimodal understanding, long context, science and math, and cybersecurity
Google's complete evaluation table for Gemini 4 Carbon, published with the 28 October launch. Image courtesy of Google.

The Real Headline: Three Stages of Cyber

Google splits security work into three stages: find the bug, prove it can be exploited, write the patch. Carbon beat Argon on all three.

Stage 1, finding vulnerabilities. Carbon leads SEC-bench Pro at 88.2%, ahead of GPT-6 Astra (85.4%) and its own predecessor Argon (83.4%).

Google's SEC-bench Pro bar chart: Gemini 4 Carbon 88.2%, GPT-6 Astra 85.4%, Gemini 4 Argon 83.4%, GPT-6.1 Terra 83.3%, GPT-6.1 Sol 78.8%, MiMo-V2.6-Pro 66.3%, DeepSeek-V4.1-Flash 62.8%, DeepSeek-V4-Pro 56.4%
SEC-bench Pro, pass at 1, higher is better. Image courtesy of Google.

Stage 2, proving a bug is exploitable. The widest gap is here. ExploitGym scores Carbon at 54.8% with a 10-hour limit per task, against 42.4% for GPT-6 Astra and 34.5% for Claude Opus 5.5.

Google's ExploitGym bar chart: Gemini 4 Carbon 54.8% at 10 hours and 53.5% at 6 hours, GPT-6 Astra 42.4% with no time cap, Gemini 4 Argon 42.0% at 10 hours and 39.5% at 6 hours, GPT-6.1 Sol 35.1%, Claude Opus 5.5 34.5% at 6 hours, Claude Mythos 5.1 30.4%
ExploitGym causal pass rate with the time cap per task shown under each model. Image courtesy of Google.

Stage 3, patching it. The margin is smallest here: 71.0% for Carbon against 68.0% for Argon, GPT-6 Astra and Grok 4.7.

Google's CWE-bench bar chart: Gemini 4 Carbon 71.0% in Antigravity, Grok 4.7 68.0% in opencode, Gemini 4 Argon 68.0% in Antigravity, GPT-6 Astra 68.0% in Codex, Claude Opus 5.5 67.0% in Claude Code, Claude Fable 5.1 58.0%
CWE-bench pass at 1, ties broken by pass at 4, with the agent harness shown under each model. Image courtesy of Google.

The field results Google attached to the launch are the reason this model is being talked about:

  • Project Fortify: Google's scan-to-remediation system reached 100% recall on scanning and 95% precision on triage on an internal benchmark. On open-source repositories it showed a recall gain of more than 25% over Argon.
  • Google Cloud Vulnerability Research: the team used Carbon to build proof-of-concept exploits for two bugs in a popular open-source project. Google compares their severity to the PHP2Shell exploit. Maintainers reclassified them as critical and patched them. Details are withheld until disclosure finishes.
  • Wiz (Scan for Good program): scanning two major SaaS providers, the model found zero-day vulnerabilities in under two hours. Google describes that as compressing weeks of research into an afternoon. The providers were notified and fixed the issues.

Read these with care. They are Google's reports. The three cyber benchmarks were run without guardrails, so results through the production API can differ. Neither the SaaS providers nor the open-source project were named. Google's own charts also show the fine print. The ExploitGym headline uses a 10-hour limit per task for Carbon, while Opus 5.5 was run at 6 hours. At the same 6-hour cap Carbon still scores 53.5% to Opus 5.5's 34.5%, so the lead holds. The CWE-bench chart shows every model ran inside its own agent harness (Carbon in Antigravity, Opus 5.5 in Claude Code, GPT-6 Astra in Codex), so it measures the model and its tooling together. None of that makes the result small. It makes it vendor-reported and not yet independently replicated.

Google also reports 0.0% attack success for Carbon on Gray Swan's indirect prompt-injection test (0.7% for Argon), and says severe alignment alerts fell by almost 80% in its internal coding environments.


Where Carbon Still Loses

This is the part most launch posts skip.

  • It hallucinates more than Argon. On Artificial Analysis's AA-Omniscience test, Carbon's score is the highest on the index (54), and it answers more questions correctly (58% against 48%). But its hallucination rate is 11%, up from Argon's 4%. More right answers, and more confident wrong ones.
  • Claude Opus 5.5 still leads agentic computer use. 87.2% against 81.5% on OSWorld-2.1.
  • Anthropic's coding agent edges it. On the Artificial Analysis Coding Agent Index, Carbon through the Antigravity CLI scores 59 while Claude Code with Opus 5.5 scores 62.
  • It is wordy. NeuralTrust reports about 200 million output tokens for a full Intelligence Index run, against a median of 64 million across models. That feeds straight into the bill.
  • The lead is narrow. One index point over Opus 5.5, and Lindy notes Artificial Analysis ran Carbon at high reasoning and the rivals at max.
  • It is not the top scorer everywhere on Artificial Analysis. Kimi K3.1 Pro posts 67.2% on Terminal-Bench 4.0 against Carbon's 65.2%, per Lindy.

Should You Switch? A Quick Decision Guide

Your jobOur read
Authorized security research, vulnerability triage, patch draftingStart with Carbon. It leads all three of Google's cyber tests.
Searching 256K to 1M tokens of documents or codeCarbon led GraphWalks by a wide margin. Test it on your data.
Agentic coding in an IDERun Carbon and Claude Opus 5.5 on your own repo. Opus leads the coding-agent index.
GUI and computer-use automationClaude Opus 5.5 led OSWorld-2.1.
Fact-sensitive answers where a wrong claim is costlyWeigh Argon's 4% hallucination rate against Carbon's 11%.
Budget-sensitive production workloadsBenchmark cost per finished task before 30 November, while the $2 rate applies.

This is our reading of the published numbers, not a lab result. Your workload decides.


What Gemini 4 Carbon Means for Your PC

Carbon is a closed-weights model served from Google's data centers. Nothing you can buy, including a 16GB graphics card, will run it locally, and anyone telling you otherwise is selling something. What changes is the machine around the model. Google says its own teams have seen Carbon run for days on a single task. Long agent sessions keep your IDE, language servers, test runners, browser tabs and Docker containers busy on a local PC the whole time.

1. CPU: parallel builds and test runs

Agents compile, lint and run suites constantly while you keep working.

2. Memory: room for the whole toolchain

An agentic IDE, a browser full of docs and a container stack add up quickly. 32GB of dual-channel DDR5 is a comfortable target, and tuning matters too. See DDR5 Sub-Timings: How Memory Tuning Fixes Game Stutters.

3. Storage: fast re-indexing of large repos

Coding agents read and index thousands of files. A Gen5 NVMe drive such as the Crucial P510 1TB M.2 NVMe Gen5 SSD with Heatsink keeps that from becoming the wait.

4. GPU: optional, for local sidecars only

You do not need a GPU for Carbon. A 16GB card such as the Gigabyte RTX 5060 Ti Eagle OC ICE 16GB GDDR7 is worth considering only if you also run local embedding models or small language models next to the cloud API, or game on the same machine.

5. Screen and power

Agent logs, diffs and code side by side need pixels. The AOC Q27G11E 27 Inch 1440p 180Hz Fast IPS Monitor is a sensible 1440p pairing, and a quality supply like the MSI MAG A750GL PCIE5.1 ATX 3.1 750W Power Supply keeps long sessions stable.

Browse the full range at the Pixlboard Store. Every product ships with its manufacturer warranty.


Gemini 4 Carbon FAQ

Is Gemini 4 Carbon available now?

Yes. Since 10 November 2026 it is open to all paying developers through the Gemini API in Google AI Studio, to enterprises through Gemini for business, and to Google AI Ultra subscribers in the Gemini app, AI Mode in Search and Antigravity.

How much does Gemini 4 Carbon cost?

$2 per million input tokens and $10 per million output tokens until 30 November 2026. From 1 December it is $4 and $20. Cached input is 95% cheaper than the input price.

Can I run Gemini 4 Carbon locally?

No. It is a proprietary, closed-weights model served by Google. Artificial Analysis lists it as not open weights.

Is Carbon better than Claude Opus 5.5?

By one point on the Artificial Analysis Intelligence Index v5 (59 against 58), and ahead on most of Google's cyber and long-context tests. Opus 5.5 leads on OSWorld-2.1 and on the Coding Agent Index. The better model depends on your workload.

Did Carbon really find zero-days in under two hours?

That is what Google and Wiz report: zero-day vulnerabilities in two major SaaS providers in under two hours, which were then disclosed and fixed. The providers were not named, so it has not been independently confirmed.

What is the Fairwind Program?

Google's access program for vetted cybersecurity defenders. It is by application, aimed at defensive and research use, and it received Carbon first.

How big is the context window?

Artificial Analysis and Vals list 1 million tokens. Google's launch post does not state a window, so confirm in AI Studio before you design around it.


Sources

Prices, availability and benchmark figures were checked on 11 November 2026 and can change. Confirm current pricing on Google's own pages before you budget.


Related Guides From the Pixlboard Journal

Photo credits and courtesy: the header key art and the evaluation table and charts are official Gemini 4 Carbon launch images from Google, provided courtesy of Google. Gemini, Gemini 4 Carbon, Argon and Antigravity are Google products and trademarks, and Pixlboard is not affiliated with Google. Benchmark figures belong to their respective evaluators.

Explore the hardware

From reading to your next build.

Browse products in the categories covered here. Check each product’s specifications and compatibility with your system before buying.