Gemini 4 Carbon: Google DeepMind's New #1 AI Model

Gemini 4 Carbon tops the AA Intelligence Index at 59 and leads in cybersecurity. Benchmarks, pricing, rollout and what it means for business AI agents.

Google DeepMind unveiled Gemini 4 Carbon on October 28, 2026, and called it its strongest model yet. It ranks #1 on the Artificial Analysis Intelligence Index v5.0 with a score of 59. It leads on cybersecurity tests and can run long, multi-step engineering jobs.

The catch: most people can't use it yet. Here is what's confirmed, what's still hype, and what it means for businesses building AI agents.

What Is Gemini 4 Carbon?

Gemini 4 Carbon is Google's newest frontier model. It arrived just two weeks after Gemini 4 Argon. Google says Carbon focuses on three things:

Cybersecurity: finding, validating and patching software vulnerabilities

Agentic coding: long-running engineering tasks that span many steps

Reliability: Google reports lower misalignment and fabrication rates in live coding tests with its own teams

Koray Kavukcuoglu, SVP at Google DeepMind and Chief AI Architect, announced the model on the official Google blog.

Gemini 4 Carbon Benchmarks

These numbers come from Google's own published comparison table. Treat vendor benchmarks as a starting point, not a final verdict.

Scores listed as Carbon / Argon / GPT-6 Astra / Claude Opus 5.5 (best in bold):

AA Intelligence Index v5.0: 59 / 50 / 57 / 58

Terminal-Bench 4.0: 65.2% / 57.1% / 59.1% / 59.6%

FrontierSWE 2: 61.2% / 55.0% / 65.5% / 62.3%

OSWorld-2.1 (computer use): 81.5% / 74.7% / 80.9% / 87.2%

AutomationBench (Zapier): 53.9% / 51.3% / 41.4% / 42.5%

GraphWalks (long context): 96.2% / 84.2% / 71.8% / 66.8%

CWE-bench (patching): 71.0% / 68.0% / 68.0% / 67.0%

Cybersecurity results

SEC-bench Pro: 88.2% for finding vulnerabilities

ExploitGym: 54.8% for exploitation

CWE-bench: 71.0% for patching

Google also says Carbon writes secure code at more than twice the success rate of Argon on internal tests. Reuters reported the same claim.

Where it doesn't win

Carbon is not best at everything. Claude Opus 5.5 still leads on computer use (OSWorld-2.1), and GPT-6 Astra leads on FrontierSWE 2. The Register notes Carbon sits fourth on WebDev Arena and second behind Opus 5.5 on Arena's Best Overall Agent board.

What About WANDR and "$5.86 Per Task"?

Posts on X claim Carbon set a WANDR record of 0.857 at $5.86 per task. We could not find that figure in Google's announcement or in major news coverage, so we are not treating it as confirmed.

Gemini 4 Carbon Pricing

$2 per million input tokens

$10 per million output tokens

Cached input tokens: 95% off

That is an introductory price, matching Argon. The Register reports it rises to $4 input / $20 output per million tokens later — the same as Claude Opus 5.5.

Who Can Use It? The Fairwind Program

Google is rolling Carbon out slowly. It goes first to a small group of cyber defenders, then to more trusted partners through the Fairwind Program. Google is also taking part in the U.S. government's voluntary pre-release testing. Broad developer, business and consumer access comes "as soon as possible," with no date given.

Ars Technica summed it up: a frontier model "that you can't use" — yet.

Pros and Cons

Pros

#1 on the AA Intelligence Index v5.0

Clear lead on security benchmarks

Strong long-context and automation scores

Competitive introductory pricing

Cons

Limited access for now

Benchmarks are mostly Google-reported

Price is set to double after the intro period

Trails rivals on computer use and some coding tests

The 88 Labs AI Take

At 88 Labs AI, we build agents for real businesses, and here is our view: the model race is now a monthly event. Argon to Carbon took two weeks. Picking "the best model" is no longer a strategy.

What works instead:

1. Build model-agnostic agents. Your agent should be able to swap Gemini, Claude, GPT or Grok without a rebuild.

2. Match the model to the job. Carbon looks like a strong fit for security reviews, long coding jobs and big-document work. A cheaper model may handle email triage just fine.

3. Test on your own tasks. A benchmark win means little until it beats your current setup on your real workflows.

4. Plan for price changes. Budget for the post-intro rate, not the launch rate.

When Carbon opens to businesses, our clients will be able to plug it in where it earns its keep. Want to see what an agent built this way looks like for your company? See your free demo.

Related reads: Best AI agent for business and our AI tools directory.

FAQ

What is Gemini 4 Carbon?

Gemini 4 Carbon is Google DeepMind's newest frontier AI model, released October 28, 2026. Google calls it its strongest model, with top cybersecurity and long-horizon task skills.

Is Gemini 4 Carbon the best AI model?

It ranks #1 on the Artificial Analysis Intelligence Index v5.0 with 59, just ahead of Claude Opus 5.5 (58) and GPT-6 Astra (57). Rivals still lead on some tests, like computer use.

How much does Gemini 4 Carbon cost?

The introductory API price is $2 per million input tokens and $10 per million output tokens. Reports say it will rise to $4 and $20 after the intro period.

Can I use Gemini 4 Carbon now?

Not broadly. Google is giving access first to cyber defenders and trusted partners through its Fairwind Program, with wider access coming later.

Should my business switch to Gemini 4 Carbon?

Not blindly. Build agents that can swap models, then test Carbon on your own workflows once it is available. 88 Labs AI can help set that up.