Gemini 4 Carbon: Google's Newest World Model

Two weeks after Gemini 4 Argon became broadly available, Google is back with another model. Gemini 4 Carbon, announced today, is what Google calls its "next frontier model," and it has taken first place on the newly updated Artificial Analysis Intelligence Index v5.0.

Most people can't use it yet. Carbon is launching first to a small group of cybersecurity defenders, with developers, businesses, and consumers getting access only later.

In this article, I'll cover what Gemini 4 Carbon is, what's changed since Argon, how it performs on benchmarks (both Google's and independent ones), what it will cost, and when you can expect to start using it.

Finally, before we get started, I recognize you may have landed on this article after hearing generally about Google AI news. To be clear, Gemini 4 Carbon (what this article is about) is a text- and code-focused frontier large language model; Genie 4, which was also just released, is an interactive, physics-based world model for real-time digital simulations. Genie 4 is also worth reading about.

What Is Gemini 4 Carbon?

Gemini 4 Carbon is the newest model in Google's Gemini 4 family, sitting above Gemini 4 Argon. That's a big story because when Argon was released only a couple weeks ago, it was widely talked about as a big comeback for Google on account of the fact that Argon has such impressive benchmark tests.

Carbon is even better. Google positions the model as great for long-horizon agentic tasks, meaning work that takes many steps, many tool calls, and a lot of recovering from dead ends.

According to Google DeepMind, internal teams have seen Carbon run for days on multi-step engineering problems, such as root-causing and patching bugs in large codebases and low-level systems.

Google also claims gains in:

• Coding and full-stack web development
• Agentic tasks and computer use
• Scientific reasoning
• Multimodal understanding

Google hasn't published a technical report, architecture details, or a parameter count. It also hasn't shared an API spec, so we don't know yet whether Carbon keeps Argon's limits (a context window of over 1 million tokens, a 65,536-token default output limit, and text, image, video, audio, and PDF input).

Gemini 4 Carbon vs. Gemini 4 Argon

The Gemini 4 rollout has moved quickly:

The benchmark results are impressive, no doubt. Notice how Carbon improves on Argon in all 17 rows of Google's published comparison.

Gemini 4 Carbon Benchmarks vs. the Competition

Let me divide this section into two smaller parts: Google's own reported results, and independent evals.

Google-reported results

Carbon has the top score in 15 of 17 rows. The two exceptions are:

• Software engineering: On FrontierSWE 2, GPT-6 Astra and Claude Opus 5.5 both beat Carbon.
• Computer use: On the OSWorld-2.1 offline subset, Claude Opus 5.5 leads.

Carbon's clearest lead is on long-context reasoning. On GraphWalks, it scores far ahead of GPT-6 Astra and Opus 5.5.

On Vending-Bench 2, a simulated business benchmark, Carbon finished with a balance of $16,484, ahead of GPT-6 Astra and Argon.

Something to note: Google's results don't include uncertainty estimates, and the security benchmarks were run without guardrails, so real-world results may or will differ.

Independent evaluations

Artificial Analysis's independent numbers broadly back up Google's claims, although with nuance:

• Intelligence Index v5.0: Carbon takes the top spot, barely ahead of Claude Opus 5.5 and GPT-6 Astra, and well ahead of Argon. Version 5.0 is a harder rebuild of the index, with most of its questions coming from private or withheld tests. That’s why scores are lower than in earlier versions.
• Coding Agent Index: Carbon (via Antigravity CLI) scores 59. That's a solid jump from Argon (53), but behind Claude Code with Opus 5.5 (62).
• AA-Briefcase (knowledge work): 1,609 Elo, up 126 points from Argon, but still behind Grok 4.7, Muse Wonder, and Kimi K3.1 Pro.
• AA-Omniscience (factual knowledge): Carbon tops the chart with a score of 54 and answers 58% of questions correctly (Argon: 48%). However, its hallucination rate rose from 4% to 11%.

On WebDev Arena, Carbon ranks #4 with 1,769 Elo, up 89 points.

The takeaway: Carbon is the strongest all-round model on the leaderboards right now, but it isn't the best at everything. If your workload is mostly agentic coding or computer use, things are still close.

Gemini 4 Carbon Cybersecurity

Security is a big story in this launch. Google breaks its cyber work into three stages, and Carbon reports strong numbers on each:

1. Vulnerability discovery: 88.2% on SEC-bench Pro (state-of-the-art)
2. Exploit validation: 54.8% on ExploitGym (state-of-the-art)
3. Patching: 71.0% on CWE-bench

Google also reports:

• More than two times Argon's success rate at generating secure code
• Carbon was over ten times better than another lab's frontier model at generating high-quality patches (no, the other lab wasn't named)

These abilities cut both ways. A model that finds and exploits vulnerabilities this well is valuable to defenders and attackers alike, which might explain the careful rollout.

Safety and Reliability Improvements

Carbon uses the same layered safety stack as Argon: safety training, real-time misuse defenses, prompt-injection resilience, and chain-of-thought monitoring. Google reports these improvements over Argon:

• Fewer unsupported claims: roughly halved
• Misalignment alerts: high-severity alerts in live coding environments fell by almost 80%
• Scope control: overreach down 26%, under-service down 29%
• Benchmark cheating: 0.0% successful cheating on Terminal-Bench 4.0, with attempts down to 0.4%
• Prompt injection: 0.0% attack success on Gray Swan's indirect prompt-injection benchmark (Argon: 0.7%)

Note that the "fewer unsupported claims" result comes from Google's internal testing, while Artificial Analysis measured a higher hallucination rate on AA-Omniscience. The two tests do measure different things, but it complicates the overall story.

Gemini 4 Carbon Pricing

Carbon launches at the same introductory price as Argon:

Google hasn't said when Carbon's introductory period ends. For reference, Argon's introductory pricing ends on November 30, 2026.

As an example, a workload with 10M input tokens and 1M output tokens would cost about $30 at introductory rates and $60 afterward.

How to Access Gemini 4 Carbon

Right now, you probably can't. Here's the rollout plan:

1. Now: A subset of cyber defenders get access.
2. Next: Access expands gradually to more trusted partners through Google's Fairwind Program.
3. Later: Developers, businesses, and consumers. 

In the meantime, Gemini 4 Argon is available now at the same introductory price and is a good way to start building workflows you can later move to Carbon.

Final Thoughts

Ultimately, because Carbon shares the exact same pricing structure as Argon, it is highly likely to quickly overshadow its predecessor for most users. The moment Carbon hits wide public availability, I expect developers and enterprise clients will naturally migrate to it to take advantage of its superior abilities.