Gemini 4 Carbon: what it is, benchmarks, what it means

Gemini 4 Carbon is Google's new flagship model, announced on 28 October 2026: for now only a few cybersecurity specialists can use it, and it is not a voice model. If you run an AI phone assistant, nothing changes tomorrow morning. Two figures in the announcement, though, matter to anyone who lets an AI talk to their customers: Carbon makes about half as many unsupported claims as its predecessor, and on the Gray Swan benchmark, as measured by Google, not a single attempt to manipulate it with hidden instructions succeeded.

Here is what was announced, what the benchmarks say and what to ask the provider of your AI phone assistant.

What Google announced

In the launch post on blog.google, Koray Kavukcuoglu of Google DeepMind describes Carbon as "our strongest model", with leading results in cybersecurity and long-horizon work. The key points:

• Four weeks after Argon. Gemini 4 Argon was announced on 30 September; Carbon builds on it and improves coding, agentic and scientific tasks, and understanding of images and documents.
• Tasks that run for days. Google says its internal teams have seen Carbon work for days on complex jobs, such as tracing the root cause of a bug in a large codebase and patching it.
• Security first. Carbon is built to find, validate and fix software vulnerabilities, which is where Google reports its strongest results.
• Gradual access. This week it goes to a subset of cyber defenders. Access will then expand to other trusted partners through the Fairwind Program and, later, to developers, businesses and consumers. No date has been given.

As of 30 October the official Gemini API model list does not include Carbon. Developers building on Google today can use Argon, generally available since 13 October according to the Gemini API changelog.

Carbon vs Argon

	Gemini 4 Argon	Gemini 4 Carbon
Announced	30 September 2026	28 October 2026
Availability	all developers since 13 October; in the Gemini app for Google AI Pro and Ultra subscribers	a few cyber defenders, then Fairwind Program partners; "coming soon" for everyone else
Stated strength	office work, automation, very long documents	cybersecurity, long multi-step tasks
Artificial Analysis Intelligence Index v5.0	50	59 (first)
Unsupported claims (Google internal testing)	baseline	about half

Sources: blog.google and the Google DeepMind Gemini page, accessed 30 October 2026.

The benchmarks Google published

Benchmarks are standardised tests: the same set of tasks is given to different models and the results are compared. Below is a selection from the table published by Google DeepMind. None of these tests measures a phone call.

Gemini 4 Carbon compared: benchmarks reported by Google

Benchmark	What it measures	Gemini 4 Carbon	Gemini 4 Argon	GPT-6 Astra	Claude Opus 5.5
AA Intelligence Index v5.0	general capability (Artificial Analysis index)	59	50	57	58
Vals Index v2.2	economically valuable office work	70.4%	69.2%	65.0%	70.2%
AutomationBench	business process automation	53.9%	51.3%	41.4%	42.5%
Terminal-Bench 4.0	command-line work	65.2%	57.1%	59.1%	59.6%
FrontierSWE 2	advanced software engineering	61.2%	55.0%	65.5%	62.3%
Agents' Last Exam v1.1	computer use as an agent	77.4%	69.0%	73.1%	71.6%
OSWorld-2.1 (offline)	computer use as an agent	81.5%	74.7%	80.9%	87.2%
GraphWalks 256k–1M (F1)	reasoning over very long inputs	96.2%	84.2%	71.8%	66.8%
CWE-bench	patching vulnerabilities	71.0%	68.0%	68.0%	67.0%
SEC-bench Pro	finding vulnerabilities	88.2%	83.4%	85.4%	—

Best result in each row in bold. Source: Google DeepMind, Gemini page, Performance section, accessed 30 October 2026. Methodology: deepmind.google/models/evals-methodology/gemini-4-carbon. A dash means no value was reported.

Three caveats:

1. These are the vendor's numbers. Google ran the cybersecurity tests without safety filters, and for competitors it partly relies on figures published by others. Independent checks will come once the model opens up.
2. Carbon doesn't win everywhere. In the full table (17 rows) it leads 15 tests: it trails GPT-6 Astra on advanced software engineering (FrontierSWE 2) and Claude Opus 5.5 on computer use (OSWorld-2.1).
3. The biggest jump is on very long inputs: on GraphWalks it goes from Argon's 84.2% to 96.2%. That is the kind of gain an agent needs to work for a long time without losing track.

The figure that matters on the phone

The most useful part of the announcement for a business is not the table but the section on reliability. According to Google's internal testing, compared with Argon, Carbon:

• makes about half as many unsupported claims, meaning answers given with confidence but not backed by the information available;
• overreaches 26% less and under-delivers 29% less, two failure modes Google links to errors in multi-step work;
• resists hidden instructions better. On Gray Swan's indirect prompt injection benchmark, which plants malicious instructions in the content a model reads, the attack success rate drops from Argon's 0.7% to 0.0%.

On the phone, these are exactly the failures that cause damage. An assistant that makes up your opening hours, confirms a service you don't offer or promises a call-back nobody will make creates a problem you then have to fix. And an assistant connected to a calendar, CRM or inbox reads text written by other people: notes, messages, web pages. Resistance to hidden instructions is what stops one of those texts from making it do something it shouldn't.

Why it isn't a voice model

Carbon's announcement says nothing about audio, voice or real-time conversation. For voice Google has a dedicated family of models, Gemini 3.8 Live, introduced on 15 September 2026 and built to answer in a fraction of a second, cope with callers who talk over it and take actions while the conversation goes on.

It's the same pattern we saw with Grok Voice 3.0 and GPT-Live-1.1: flagship models handle long, complex work, voice models handle the call. Over time the gains of the former reach the latter, and that is where a business will notice them: fewer calls ending with "we'll call you back", fewer wrong answers, more requests settled in a single call.

Questions to ask your provider

For a restaurant, a medical practice or a professional firm, the Carbon announcement is a good excuse to check the assistant you have, or the one you are considering:

• What happens when it doesn't know the answer? Make a test call with a question it can't answer: it should say so and offer a next step, not make something up.
• Where does its information come from? Opening hours, services, prices and rules should come from what you configured.
• What does it read from outside? If it is connected to your calendar, CRM or inbox, ask how it is protected against instructions hidden in that content.
• Which voice models does it use, and how are they updated? A good provider switches models when a better one comes out for its use case, without making you redo the setup.
• What can it do during the call? Booking appointments, logging requests, sending a summary: features and integrations matter more than scores.
• Does it hand the call to a person when needed, and say up front that it is an AI, as the AI Act requires?

For a comparison of the options available in Italy, see our guide to the best AI phone assistants of 2026.

In short

Gemini 4 Carbon is Google's strongest model and tops the Artificial Analysis index, but for now only a few security specialists can use it and it isn't built for voice. For anyone using AI on the phone, the real news is the direction: models that make things up less and are harder to manipulate. An assistant should still be judged on real calls. If you want to hear how an AI phone assistant would answer for your business, book a demo.