Google announces Gemini 4 Carbon, its new frontier AI model

Google announced Gemini 4 Carbon, a new frontier AI model that arrives just two weeks after Gemini 4 Argon reached the Gemini app.

In a blog post published Oct. 28, Koray Kavukcuoglu, SVP at Google DeepMind, the company’s frontier AI lab, called Carbon Google’s strongest model yet. (Kavukcuoglu is leading Google DeepMind now that former CEO Demis Hassabis has stepped aside.) The post points to cybersecurity and long-horizon tasks, meaning complicated jobs that take many steps or run for a long time, as Carbon’s biggest strengths.

Already, Gemini 4 Carbon has the top spot in the Artificial Analysis Intelligence Index, beating out Claude Opus 5.5.

Credit: Screenshot: Artificial Analysis

The pace at which AI companies are releasing new frontier models has accelerated sharply in recent months, as the frontier AI labs race to have the best model available. Anthropic released Claude Opus 5.5 on Sept. 22, OpenAI shipped GPT-6 Astra earlier that month, and Google announced Argon on Sept. 30 before releasing it to Google AI Pro and Ultra subscribers on Oct. 13. Each lab has claimed a new lead within weeks of the last.

As with Argon, Google is announcing Carbon well before most people can use it. Carbon is starting with a small group of cybersecurity defenders, and Google hasn’t said when it will reach everyone else. (Last week’s PHP2Shell incident shows just how much havoc new AI models could wreak on the internet at large.) Announcing early lets Google claim the top spot on the leaderboards now. It also gives everyone a look at the benchmarks before the next round of releases.

Gemini 4 Carbon: Benchmarks

In Google’s comparison table, Carbon posts the top score on 15 of 17 benchmarks against Argon, GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1. All figures in this section are as reported by Google.

Credit: Google

The two benchmarks Carbon doesn’t win are worth noting. GPT-6 Astra still leads on FrontierSWE 2, a software engineering test. Claude Opus 5.5 leads on OSWorld-2.1, which measures how well a model can operate a computer.

A big change from Argon is that Google now reports Terminal-Bench 4.0 scores run by Artificial Analysis, an independent testing firm, rather than its own. For quick context, Terminal-Bench 4.0 measures how well an AI model can complete real technical tasks — like installing software or fixing broken code — by typing commands into a computer’s terminal on its own. It’s one of the hardest agentic tests for LLMs.

This is notable because the scores companies reported for each other’s models didn’t always line up with independent testing. On Terminal-Bench 4.0, Carbon jumps about eight points over Argon and moves ahead of Opus 5.5.

Google also says Carbon can work on difficult engineering problems for days at a time. Its web development skills improved enough to climb to fourth place on WebDev Arena, a leaderboard where people vote on AI-built websites.

The Artificial Analysis Intelligence Index

The Artificial Analysis Intelligence Index is one of the most closely watched independent leaderboards for AI models. Artificial Analysis runs a range of tests itself and combines them into a single composite score.

On the newly updated version 5.0, Carbon (high) scores 59, one point ahead of Claude Opus 5.5 (max).

Gemini 4 Carbon’s score of 59 is leaps above Argon’s 50.
Credit: Artificial Analysis

The scores can’t be directly compared to the previous version of the index, where Argon (High) scored 53. A one-point lead is also thin enough that the next release from OpenAI or Anthropic could easily erase it. Carbon and Meta’s Muse Wonder Full are both flagged on the leaderboard as having no publicly available API, so they can’t yet be used outside their companies’ own access programs.

Gemini 4 Carbon: Safety and safeguards

Google says Carbon uses the same safety system as Argon, which includes monitoring the model’s reasoning, and adds more work on alignment, or keeping the model behaving as intended. According to Google’s own testing, Carbon is also better behaved than its predecessor.

Google says Carbon never successfully cheated on Terminal-Bench 4.0, meaning it never found a shortcut that passed a task without actually doing the work, and it tried to cheat only 0.4 percent of the time. During internal coding work, high-severity misalignment alerts, which flag a model acting in ways its developers didn’t intend, dropped by nearly 80 percent. Carbon also makes about half as many unsupported claims as Argon. And on Gray Swan’s prompt injection benchmark, which tests whether hidden instructions can hijack a model, Carbon blocked every attack attempt, improving on Argon’s 0.7 percent attack success rate.

All of these figures come from Google’s own testing.

How to try Gemini 4 Carbon

For now, you can’t use Gemini 4 Carbon unless you’re an approved cyber defender in Google’s Fairwind Program. Starting this week, Carbon is rolling out to a small group of those testers, and Google says it will gradually add more trusted partners. Developers, businesses, and consumers will get access after that, but Google hasn’t given an exact public launch date.

If we use Argon to figure out a timeline, about two weeks passed between its announcement and its arrival in the Gemini app for Pro and Ultra subscribers. Though we won’t be able to confirm whether Google will commit to that same schedule for Carbon.

When Carbon does reach the API, Google says it will cost the same as Argon’s introductory price:

• $2 per million input tokens

• $10 per million output tokens

• Cached input tokens at 95 percent off the input price

That keeps Carbon well below its main rivals. Mashable previously reported that Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, while GPT-6 Astra costs $10 per million input tokens and $50 per million output tokens.

Want to learn more about getting the best out of your tech? Sign up for Mashable’s Top Stories and Deals newsletters or get Mashable push alerts today.