Gemini 4 Carbon: what Google's new top model means at home

Google has released Gemini 4 Carbon, which it calls its strongest model so far. It is built for two things: finding and fixing security flaws in software, and long, multi-step jobs that can run for hours or days. This piece explains what that means for ordinary households rather than retelling the launch, which our news story already covers.

The short answer: most people at home will notice little this week. The model matters most if you write or maintain your own smart-home code, or if you pay for Google's top consumer plan and use Gemini for research. For everyone else it is a sign of where AI assistants are heading, not a reason to change anything.

The practical advice is simple. Try it on non-critical work first, and check every security fix it suggests before it goes near a live system.

What is Gemini 4 Carbon?

Gemini 4 Carbon is the newest and most capable model in Google's Gemini 4 family. Google says it is designed for cybersecurity work and for long tasks, such as engineering projects that take hours or days to complete rather than a single question and answer.

The rollout came in stages. According to Google DeepMind's announcement, access first went to a small group of cyber defenders, meaning people who protect systems from attack. On 10 November 2026 it opened to developers, businesses, scientists and consumers, as described on Google's product blog. Engadget reported that availability is worldwide.

The previous model, Gemini 4 Argon, has not gone away. 9to5Google reported that Argon stays available, which matters if you have something that works and would rather not change it.

What do the benchmark numbers actually tell you?

A benchmark is a standard test that lets you compare models on the same task. Google reported three security results, all run without guardrails, meaning without the safety limits a public product would normally apply.

SEC-bench Pro checks how well a model finds real flaws in software. ExploitGym checks whether a flaw can actually be exploited. CWE-bench measures whether a model fixes the root cause of a flaw rather than patching over the symptom. Google also said that on its internal tests, Carbon writes secure code with more than twice the success rate of Argon.

Those are Google's own figures, so treat them as the maker's claims until others reproduce them. The independent numbers are more modest. Tom's Hardware reported that Carbon scored 59 on version 5 of the Artificial Analysis Intelligence Index, first place but only one point ahead of Anthropic's Claude Opus 5.5. That is a narrow lead, not a gap.

Who is affected at home?

Three groups have a real reason to pay attention.

The first is the tinkerers. If you write automations, maintain a Home Assistant setup with custom code, or run your own scripts that talk to cameras, locks and sensors, a model that is better at spotting weaknesses in code could be useful. It could also be tempting to trust too much. Code that controls a door lock deserves more scrutiny than code that dims a lamp.

The second is paying subscribers. Carbon is available now in the Gemini app for AI Ultra subscribers, and Digital Trends reported that developers and businesses get it at the same price as Argon. If you are on a free or lower tier, the material does not say when Carbon will reach you, so assume it will not, for now.

The third group is everyone else, who will meet Carbon indirectly. Software makers, including those behind the apps and firmware on your devices, may use models like this to audit their own code. That is a possible benefit, but nothing in the material says any smart-home maker is doing so.

Is it really better at long tasks and agents?

Google positions Carbon for long, multi-step work, which is the territory of AI agents: programs that plan steps, use tools and carry on without you typing each instruction. Our assistants and agents section follows that shift.

The Verge highlighted an agent swarm feature, in which several AI agents work together on a job. Until that is clear, we would not build a home routine around it.

There is a general point here. The longer an agent runs unsupervised, the more a small mistake can compound. A task that runs for days needs checkpoints where a person reviews the work, however capable the model.

Does it make things up less often?

Google says so. It reports that Carbon cuts unsupported claims by about half compared with Argon. Tom's Hardware reported that Carbon led the AA-Omniscience test, which measures both knowledge and hallucinations (confident answers that are simply wrong), with a score of 54 on a scale from -100 to +100.

The same report added an important caveat: even the best models scored well below a level that would mean reliable recall. In other words, first place in a weak field is still not a model you can trust without checking.

For research at home, such as comparing heat pump models, reading a tariff or checking a device's compatibility, that means Carbon may be a little safer than before, not safe. Ask it for sources, then open them. Our AI at home section covers where assistants help and where they still fall short.

What should you do now?

If you want to try Carbon, a cautious route looks like this.

1. Check your access. Carbon is in the Gemini app for AI Ultra subscribers and through the API for developers. If you do not already pay for either, there is nothing to rush.
2. Start with non-critical code. Use a spare script or a test copy of your configuration, not the automation that controls your front door or heating.
3. Keep a backup. Save the working version before you accept any change, so you can roll back.
4. Read every proposed change. Ask the model to explain each fix in plain words. If it cannot, or you do not understand the explanation, do not apply it.
5. Test after each change. Run the automation in a safe state, such as with the heating off or the lock open and you standing next to it.
6. Do not paste secrets. Keep passwords, API keys and access tokens out of any prompt. This is general good practice rather than something specific to Carbon.

If you already use Argon and it does the job, you can stay with it. Argon remains available, and Carbon at the launch price offers no reason to switch in a hurry unless you have a task Argon handles badly.

What does it cost, and who should skip it?

Google says Carbon is available through the API at the launch price, and Digital Trends reported that this matches Argon. In the Gemini app it sits behind the AI Ultra subscription.

Skip it, for now, if you mainly use an assistant for timers, music and lights. Carbon is aimed at demanding work, and a voice assistant on a smart speaker will not behave differently because of it. Skip it too if you are wary of sending home-network details or private code to a cloud service. In that case, our local AI section explains how to run smaller models on your own hardware, though they are generally less capable than a frontier model like this.

It is also worth keeping the dual-use nature in mind. A model that is good at finding and exploiting flaws helps defenders, and in principle could help attackers. That is presumably why Google began with a small group of cyber defenders, though the material does not detail what safeguards apply to the public release.

What to watch next

Several things will decide whether Carbon matters beyond the headlines.

• Independent testing. Google's security scores need outside reproduction, especially as they were run without guardrails.
• Rival responses. Carbon leads Claude Opus 5.5 by one point on one index. A small lead can disappear with the next release.
• Wider access. Whether the model reaches cheaper plans, and when.
• Connected services. Google is also working with Adobe on bringing Acrobat into Drive and Gmail, which we cover in our piece on the Google and Adobe tie-up. Stronger models are what make that kind of integration worth doing.
• Rules and oversight. Models with strong offensive security skills raise questions for regulators. Our AI policy section tracks that.

For now the sensible stance is curiosity with a safety net. Carbon looks like a capable tool for people who write code and do long research, and a modest event for everyone else.