Gemini 4 Carbon: What Google's New Agent Model Means

Key takeaways

Google announced Gemini 4 Carbon on October 28, 2026, two weeks after Gemini 4 Argon, and says it is built for long, multi-step agent work that can run for days.
Carbon ranks first on the Artificial Analysis Intelligence Index v5.0 with 59 points, just ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57.
Most businesses cannot use it yet: Google is giving access to a subset of cyber defenders first through its Fairwind Program, with no public release date.
Introductory pricing matches Argon at $2 per million input tokens and $10 per million output tokens, rising later to $4 and $20, with no end date announced.
Independent tests show trade-offs, including a higher hallucination rate than Argon on one factual-knowledge test and weaker scores than several rivals on professional knowledge work.

Gemini 4 Carbon is Google's newest frontier AI model, and the company is pitching it as an agent that can keep working on hard, multi-step jobs for a very long time. Google announced it on October 28, 2026, two weeks after Gemini 4 Argon, calling it "our strongest model with state-of-the-art cybersecurity and long horizon capabilities" (Google). The launch reached 929 points and 745 comments on Hacker News, and our count of X posts mentioning the model jumped from 116 on October 27 to 11,843 on October 28 (UTC days), with 5,560 more before 09:35 UTC on October 29. It is also a trending story on X.

The short version for a small business: Carbon looks like a real step up for long-running AI agents, but you can't use it yet, and the price you see today is an introductory one.

What Google announced on October 28

A model built for long agent tasks

Google says Carbon is "built for sustained agentic workflows." Its own teams have seen it "run for days" while working through dead ends on complex engineering jobs such as tracking down and patching subtle bugs in large codebases (Google). The company also says it improved coding, scientific tasks, full-stack web development and multimodal understanding, and that Carbon moved up 89 Elo points to fourth place on WebDev Arena.

Number one on a major AI index

Carbon scores 59 on version 5.0 of the Artificial Analysis Intelligence Index, which Google says puts it in first place (Google). According to VentureBeat, that is ahead of Claude Opus 5.5 at 58, GPT-6 Astra at 57 and Argon at 50, and Artificial Analysis says 60% of the revised v5 test suite uses private tests or withheld answers (VentureBeat).

Cybersecurity is the headline skill

Google reports 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploiting them and 71.0% on CWE-bench for patching them. It notes these external scores were measured "without guardrails" (Google). Reuters reported that Google says Carbon generates secure code at more than twice the success rate of Argon on internal benchmarks (Reuters via Investing.com).

Who can use Gemini 4 Carbon, and when

Cyber defenders first

Google says it will make Carbon available "to a subset of cyber defenders" starting this week, then gradually add more trusted partners through its Fairwind Program. It is also taking part in the U.S. government's voluntary pre-release testing process (Google). The Verge describes current access as limited to "a select group of cybersecurity defense companies" (The Verge).

No public release date yet

Google says developers, businesses and consumers will get Carbon "as soon as possible," but it has not given a date (Google). VentureBeat points out that Argon, which Google opened more broadly on October 13, is still the model customers can actually test today, and that Carbon's announcement includes no full API specification (VentureBeat).

Gemini 4 Carbon pricing

Per million tokens	Introductory	Later
Input	$2	$4
Output	$10	$20
Cached input	95% off the input price	Not stated

Those figures come from Google's announcement (Google). VentureBeat notes Google has not said when Carbon's introductory period ends, while Argon's higher rates take effect December 1. Its worked example of 10 million input tokens and 1 million output tokens costs $30 at the introductory rates and $60 afterwards (VentureBeat). For an agent that runs for hours, the price per finished job matters more than the price per token, and Google hasn't published that yet.

Is it more reliable? The claims and the caveats

What Google says improved

Google says Carbon makes about half as many unsupported claims as Argon, overreaches 26% less and under-delivers 29% less on multi-step plans. It reports high-severity misalignment alerts fell by almost 80% in its internal coding environments, successful cheating on Terminal-Bench 4.0 fell to 0.0%, and the attack success rate on Gray Swan's indirect prompt-injection benchmark dropped from Argon's 0.7% to 0.0% (Google).

What independent tests found

VentureBeat's review of Artificial Analysis results is more mixed. Carbon answers 58% of AA-Omniscience questions correctly, up from Argon's 48%, but its hallucination rate on that test rose from 4% to 11%. On AA-Briefcase, a knowledge-work test, Carbon scored 1,609 Elo, behind Grok 4.7, Muse Wonder and Kimi K3.1 Pro. On AutomationBench, a business workflow test, it improved only modestly, from 51.3% to 53.9% (VentureBeat). In other words, a 0.0% score on one prompt-injection test is not proof the model can't be tricked by a malicious email or web page.

What we did not verify or test

We have not used Gemini 4 Carbon, because it is not open to the public. Every benchmark score here is reported by Google, Artificial Analysis or the outlets linked above, and we did not rerun any of them. We could not confirm when Carbon will reach the Gemini app, Google Workspace or the public API, or whether it will be offered in Canada at launch.

What this means for small businesses

Carbon is a signal of where AI agents are going: longer jobs, less babysitting, better security work. Here is how to get ready without waiting:

Keep using what you can test today. Argon is the Gemini 4 model that is publicly available now (VentureBeat). Use our prompt to compare software tools before you switch providers.
Budget for the later price. Plan on $4 and $20 per million tokens, not the introductory rates. The AI time savings calculator shows whether a long agent task still pays off at that price.
Write agent briefs that survive long runs. Agents that work for hours need clear goals, limits and a definition of done. Check your instructions with the prompt grader or build them with the AI prompt generator.
Make agents check facts against live sources. Given the higher hallucination rate on one test, set rules like those in check the live source for dates, prices and statuses.
Verify what an agent says it did. Use check whether your AI coding agent really did what it says on any website or code change.
Answer the security questions first. Before you give any agent account access, work through the money and security questions before an agent goes live.

Frequently asked questions

What is the Gemini 4 Carbon release date?

Google announced Carbon on October 28, 2026, and is giving it to a subset of cyber defenders first. It says wider access for developers, businesses and consumers will come "as soon as possible," with no date yet (Google).

How much does Gemini 4 Carbon cost?

The introductory price is $2 per million input tokens and $10 per million output tokens, with cached input 95% off. After the introductory period it rises to $4 and $20 (Google). Google has not said when the introductory period ends (VentureBeat).

Is Gemini 4 Carbon better than Claude Opus 5.5?

On the Artificial Analysis Intelligence Index v5.0, Carbon scores 59 against Opus 5.5's 58. But Opus leads on the OSWorld offline computer-use subset, and on Artificial Analysis's Coding Agent Index Claude Code with Opus 5.5 scores 62 against Carbon's 59 (VentureBeat).

What is the difference between Gemini 4 Carbon and Argon?

Carbon builds on Argon, which launched two weeks earlier, and Google says it improves coding, agent work and cybersecurity, with more than twice Argon's success rate at generating secure code on internal tests (Google). Both have the same advertised token prices (VentureBeat).

Sources

How we report: every fact, number and quote above comes from the linked sources, which we read before publishing. Product details can change quickly during previews, so check the vendor page before you buy. Spotted an error? Tell us via the link on our home page.

Gemini 4 Carbon: what Google's new long-running agent model means for small businesses