Gemini 4 Carbon: facts, limits and what SMEs should do now

Google calls Gemini 4 Carbon its strongest model. For now, only cyber defenders can use it. What is verified, what remains open and what Swiss SMEs can do now.

Back to Blog
Editorial review:

Editorial review: 30 October 2026. Access, pricing and benchmarks checked against the primary sources. Carbon is not yet generally available; details may change.

At a glance

Gemini 4 Carbon is Google's new frontier model for cybersecurity and long tasks. It has been announced but is not yet generally available. Independent tests place it narrowly ahead, with more hallucinations than its predecessor Argon. This article puts benchmarks, pricing, access and data protection into context for Swiss SMEs.

  • Google announced Gemini 4 Carbon on 28 October 2026. For now, only selected cyber defenders have access; there is no date for businesses.
  • Measured independently, Carbon is narrowly in first place: 59 points on the Artificial Analysis Intelligence Index v5, ahead of Claude Opus 5.5 at 58.
  • The hallucination rate rises from 4% to 11% compared with Argon. It stays low, but it does not replace your own test.
  • For Swiss SMEs this means: work with Gemini 4 Argon today, watch Carbon and do not promise that data stays in Switzerland.

On 28 October 2026 Google introduced Gemini 4 Carbon and called it its strongest model. For now, only selected cyber defenders can use it. This article separates what Google claims, what independent tests measure and what is still open. It ends with what Swiss SMEs can sensibly do now. [1]

Title graphic for Gemini 4 Carbon: announced on 28 Oct 2026 by Google DeepMind, access for now through the Fairwind Program for selected cyber defenders, introductory API price USD 2 / 10 per 1M tokens, later 4 / 20.
Key facts according to the Google blog of 28 October 2026. Chart: GlasBox Studio.

What Google announced

Koray Kavukcuoglu, Senior Vice President at Google DeepMind and Chief AI Architect at Google, introduced Carbon on the official Google blog. Google describes the model as "our strongest model with state-of-the-art cybersecurity and long horizon capabilities". The two focus areas are cybersecurity and long, multi-step tasks. [1]

According to Google, Carbon builds on Gemini 4 Argon. Argon has been generally available since 13 October 2026. For the Carbon launch, Google says it is taking part in the US government's voluntary process for testing new models. [1] [5]

Google does not explain the name. Only the sequence is known: the first model of the generation is called Argon, the second Carbon. Google does not say whether the name stands for a size, a price tier or anything else. We therefore do not interpret it.

Gemini 4 timeline: 30 Sep 2026 Argon announced, Fairwind only at first; 13 Oct 2026 Argon generally available as gemini-4-argon; 28 Oct 2026 Carbon announced, cyber defenders first; 30 Nov 2026 Argon introductory price ends; Carbon access for businesses without a date.
Simplified timeline based on Google's statements. As of 30 October 2026.

Who can use Carbon today

In short: almost nobody. From the week of 28 October, Google is giving Carbon to a subset of cyber defenders first. Further trusted partners then get access step by step through the Fairwind Program. Developers, businesses and consumers are to get Carbon as soon as possible. Google gives no date. [1] [4]

ChannelGemini 4 CarbonGemini 4 Argon
Fairwind Programstep by step from 28 Oct 2026first from 30 Sep 2026
Gemini API and Google AI Studiono model ID publishedgenerally available since 13 Oct 2026 as gemini-4-argon, paid tier only
Vertex AI (Google Cloud)not announcedavailable to businesses
Gemini appnot announcedfor Google AI Pro and Ultra plans
AI Mode in Google Searchnot announcedselectable for Google AI Pro and Ultra

In practice this means: if you want to use a Gemini 4 model in production today, you use Argon. Carbon is an announcement, not yet a tool. [5] [8]

What Carbon is meant to do

Google describes Carbon as a model for long, agentic workflows. Internally, it reportedly kept working on complex software tasks for days, for example finding and fixing hidden bugs in large codebases. [1]

  • Cybersecurity across the whole cycle: finding vulnerabilities, checking exploitability, fixing the root cause. [1]
  • Secure code: in internal tests, Carbon produces secure code more than twice as often as Argon, according to Google. [1]
  • Patches: the Chrome team reportedly measured more than ten times better results for high-quality patches than with a large model from another lab. Google does not say which model. [1]
  • Web development: fourth place on WebDev Arena with 1768 points, 89 points more than before. [1]
  • Collaboration: when a task is unclear, Carbon is meant to retrieve sources and show discrepancies to the user first instead of guessing. [1]

These are Google's claims. They show the direction. They do not replace a test with your own tasks.

Benchmarks: Google's figures

Google compares Carbon in a table of 17 tests with Gemini 4 Argon, GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1. The chart shows six of these tests in a direct comparison with Argon. [1]

Bar chart with Google's figures, Carbon versus Argon: SEC-bench Pro 88.2 versus 83.4 per cent, ExploitGym 54.8 versus 42.0, CWE-bench 71.0 versus 68.0, Terminal-Bench 4.0 65.2 versus 57.1, GraphWalks 96.2 versus 84.2, OSWorld-2.1 81.5 versus 74.7.
Vendor figures from the Google blog of 28 Oct 2026. Google measured the three cyber tests without guardrails.

Two points matter for interpretation. First, Google says it measured the three cyber tests without guardrails. In everyday use, safeguards are active, so results may differ. Second, Google chooses the comparison models itself. Models that beat Carbon in individual independent tests are missing from the table. [1] [2]

Even in Google's own table, Carbon does not lead everywhere. [1]

TestAhead of CarbonGemini 4 Carbon
FrontierSWE 2 (software tasks)GPT-6 Astra 65.5%, Claude Opus 5.5 62.3%61.2%
OSWorld-2.1 (operating a computer)Claude Opus 5.5 87.2%81.5%

What independent tests show

Artificial Analysis compares AI models with its own tests. On the new Intelligence Index v5, Carbon takes first place with 59 points. The lead is narrow: Claude Opus 5.5 scores 58, GPT-6 Astra 57 and Gemini 4 Argon 50. [2]

Bar chart, Artificial Analysis Intelligence Index v5, as of 28 Oct 2026: Gemini 4 Carbon 59, Claude Opus 5.5 58, GPT-6 Astra 57, Gemini 4 Argon 50.
Independent measurement by Artificial Analysis. Higher is better.

Carbon leads 4 of the index's 9 tests: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA. On agentic coding it is narrowly behind Claude Opus 5.5. On the Coding Agent Index, Carbon with Antigravity CLI scores 59, while Claude Code with Opus 5.5 reaches 62. [2]

On agentic knowledge work (AA-Briefcase), Carbon reaches 1609 Elo points, 126 more than Argon. Kimi K3.1 Pro, Muse Wonder and Grok 4.7 are still ahead there. Vals AI ranks Carbon second of 29 models on its Vals Index, at 70.36%. [2] [11]

Hallucinations: the figure that matters most in daily use

On the AA-Omniscience knowledge test, Carbon answers 58% of questions correctly, 10 points more than Argon. At the same time, its hallucination rate rises from 4% to 11%. According to Artificial Analysis, this is still the second-lowest value of any model, after Argon. Claude Opus 5.5 and GPT-6 Astra are both at 50%. [2]

Bar chart, hallucination rate in AA-Omniscience v1.1: Gemini 4 Argon 4 per cent, Gemini 4 Carbon 11 per cent, Claude Opus 5.5 50 per cent, GPT-6 Astra 50 per cent.
Independent measurement by Artificial Analysis. Lower is better.

The hallucination rate shows how often a model gives a wrong answer instead of admitting a gap in its knowledge. So 11% does not mean that 11% of all answers are wrong. For your use case, only your own test with known correct answers counts.

Google reports about half as many unsupported claims internally as with Argon. The independent test measures something else. Both statements can be true at the same time. [1] [2]

How to read these figures

  • Version 5 of the index has been rebuilt. Values from earlier versions are not comparable. [2]
  • A one-point lead does not prove that a model is better for your task.
  • Artificial Analysis tests Carbon with a high reasoning setting (High). Other settings produce other values. [2]
  • Benchmarks measure test tasks. Your quotes, contracts and codebases are different tasks.

Pricing and costs

Google announces the same introductory price for Carbon as for Argon. Cached input tokens cost 95% less than regular ones. After the introductory period, prices double. Google does not say when this period ends for Carbon. [1] [2]

USD per 1M tokensGemini 4 Carbon (announced)Gemini 4 Argon (Gemini API)
Input, introductory price2.002.00 until 30 Nov 2026
Output, introductory price10.0010.00 until 30 Nov 2026
Cached input95% cheaper than input0.10, from 1 Dec 2026 0.20
Input, standard price4.004.00 from 1 Dec 2026
Output, standard price20.0020.00 from 1 Dec 2026
End of introductory pricenot stated30 Nov 2026

Prices are in US dollars, as Google states them. We deliberately do not convert them to Swiss francs. [1] [6]

The same token price does not mean the same cost per task. A model that reasons longer produces more output tokens. Artificial Analysis therefore calculates the cost per test task. [2]

Bar chart, cost per Intelligence Index v5 task at standard pricing: GPT-6 Astra USD 6.25, Gemini 4 Argon USD 7.68, Gemini 4 Carbon USD 7.76, Claude Opus 5.5 USD 10.04.
Independent measurement by Artificial Analysis, standard pricing without introductory discount.

At standard pricing, one index task with Carbon costs USD 7.76. That is about 23% less than with Claude Opus 5.5, but about 1.2 times as much as with GPT-6 Astra. Carbon uses an average of 123,000 output tokens per task, 29% fewer than Opus 5.5. [2]

A simple example: 10 million input tokens and 1 million output tokens cost USD 30 at the introductory price. At the standard price, it is USD 60. Caching, batch processing and grounding with Google Search change the calculation.

Technical specifications: what is verified

There is no model card and no API documentation for Carbon yet. Artificial Analysis reports some specifications from its own tests. For Argon, Google documents them itself. [2] [3]

FeatureGemini 4 CarbonGemini 4 Argon
Model IDnot publishedgemini-4-argon
Context window1M tokens (according to Artificial Analysis)1,048,576 tokens
Inputtext, image, video, speech (according to Artificial Analysis)text, image, video, audio, files
Outputtexttext
Max. outputnot published65,536 tokens by default
Speed at high reasoningabout 100 tokens per second, first answer after about 97 secondsnot part of this comparison

The roughly 97 seconds to the first answer apply to difficult test tasks at a high reasoning setting. They are no benchmark for simple questions. Still, expect waiting time for complex jobs. [3]

Safety and data protection

According to Google, Carbon uses the same multi-layered protection as Argon: training, real-time misuse defences, prompt-injection resilience and chain-of-thought monitoring. Google cites internal measurements for this. [1]

  • Successful cheating on Terminal-Bench 4.0 drops to 0.0%, cheating attempts to 0.4%.
  • High-severity misalignment alerts in internal coding use drop by almost 80%.
  • 26% less overreach beyond the task and 29% less work left undone.
  • 0.0% successful attacks in Gray Swan's prompt-injection test (Argon: 0.7%). This applies to the attacks tested, not to every conceivable one.

Google has not yet published a separate model card or safety report for Carbon. The announcement refers to the safety report for Argon. [1]

What happens to your data

For the Gemini API: with paid services, Google does not use prompts and responses to improve its products. For users in Switzerland, the EEA and the United Kingdom, this also applies to free services such as Google AI Studio. Anyone offering API applications to users in Switzerland may only use paid services. [7]

Even with paid services, Google logs prompts and responses for a limited period to detect abuse. [7]

On Vertex AI, you choose the location. Stable Gemini models run in the global location and often also in the us and eu multi-regions. With the global location, Google can process a request anywhere in its worldwide pool. Google has not yet listed a Swiss location for these models. [9] [10]

What this means for Swiss SMEs

Today, Carbon is a signal, not a tool. Plan with Argon and keep your solution open for a later change of model.

Test, wait or apply?

  • Test now: Gemini 4 Argon, if you have demanding code or document tasks. The introductory price applies until 30 November 2026. [6]

  • Wait: with Carbon, until there is a model ID, a model card, documented locations and an end date for the introductory price.

  • Do not apply: the Fairwind Program is aimed at cyber defence organisations. For most SMEs it is not a route to Carbon. [4]

  • Prepare for a switch: write down prompts, test cases and expected answers. Then compare a new model against the same cases.

Four applications you can test with Argon today

Code maintenance

An online shop has a plugin checked for outdated functions and known vulnerabilities. A person reviews and tests every change.

Security check of your own website

An agency has forms and dependencies analysed for typical gaps. Only on its own systems and with written permission.

Long documents

A fiduciary firm has contracts or quotes summarised and deviations highlighted. The specialist decides.

Internal knowledge

A team answers questions from approved internal documents. That requires roles, sources and review rules, not just a strong model.

What this means for SEO and GEO

Google has not announced that Carbon will be used in AI Overviews or in AI Mode in Google Search. Only the predecessor is confirmed: since 13 October 2026, Gemini 4 Argon can be selected in AI Mode by Google AI Pro and Ultra subscribers. [1] [8]

A new model is not a new ranking factor. Nobody can guarantee you a position or a citation in AI answers. According to Google, Carbon retrieves sources and surfaces discrepancies when things are unclear. Clear, evidenced and consistent content therefore remains a solid foundation. [1]

  • Every service has its own crawlable page with an unambiguous name.

  • Figures carry a date and a source.

  • Company name, address and service terms are the same on every page.

  • You regularly check your most important customer questions in AI Mode and note which sources appear.

What this article does not claim

  • No claim that Carbon is the best model for your task.
  • No adoption of vendor figures as verified fact.
  • No promise of rankings or mentions in AI answers.
  • No assurance that data stays in Switzerland.
  • No legal advice on the nDSG/FADP.
The next step

Review your website foundation for Google and AI?

Start with your website. The free short check provides a four-page report from publicly accessible data: priorities, sources and measurement gaps. It includes no AI-answer series, complete audit or ranking promise.

Request the free short check

Sources

As of 30 October 2026. Vendor figures are labelled as such. Independent values come from Artificial Analysis and Vals AI.

Frequently Asked Questions

What is Gemini 4 Carbon?

An AI model from Google DeepMind, announced on 28 October 2026. Google calls it its strongest model, focused on cybersecurity and long, multi-step tasks.

Can I use Gemini 4 Carbon today?

Usually not. For now, selected cyber defenders have access through the Fairwind Program. Google has not given a date for developers, businesses or consumers.

How much does Gemini 4 Carbon cost?

Google announces USD 2 per million input tokens and USD 10 per million output tokens. After the introductory period, USD 4 and USD 20 apply. There is no end date for the introductory price yet.

Is Carbon better than Claude Opus 5.5 or GPT-6 Astra?

On the Artificial Analysis overall index, Carbon scores 59, narrowly ahead of Opus 5.5 at 58 and GPT-6 Astra at 57. Opus 5.5 leads on agentic coding. For your task, your own test decides.

Does Google use my data for training?

According to Google, not for paid Gemini API services. For users in Switzerland this also applies to free services such as Google AI Studio. Time-limited logging for abuse detection remains.

Does my data stay in Switzerland?

That is not documented. For Gemini models on Vertex AI, Google lists locations such as global, us and eu, but no Swiss location.

Is Carbon used in AI Overviews?

Google has not announced this. The only confirmed point is that Gemini 4 Argon can be selected in AI Mode by Google AI Pro and Ultra subscribers.