Gemini 4 Carbon Frontier Model
Sustained agentic work and state-of-the-art cyber defense in one model
None

None

None

Long Story Video Skill

Turn your idea into 10s–10min video—AI writes scripts, prompts, and generates footage automatically.

Ads Video Skill

Ads Video Skill

Generate professional ads and sales videos—AI auto-generates scripts, prompts, and footage.

3D Science Explainer Video Skill

Convert scientific concepts into stunning 3D explain animations

AI Video Prompt Generator

Feedback

freeTrialImage.bannerPity

freeTrialImage.upgradeUnlock

  • ✓freeTrialImage.benefitHd
  • ✓freeTrialImage.benefitWatermark
  • ✓freeTrialImage.benefitUnlimited

Gemini 4 Carbon

See how Gemini 4 Carbon performs on agentic coding, security and honesty benchmarks, what it costs, and when access opens for developers.

All Tools

Discover our comprehensive AI-powered animation toolkit

Built Different: The Story Behind Gemini 4 Carbon

Designed for workloads that run for days and for security teams on the front line, this model pairs record-setting results with a deliberately cautious release.

  • Runs That Last for Days, Not Minutes
    Engineers built it for problems that span many steps. Early internal use shows multi-day sessions where the model keeps digging past dead ends until it finds the root cause of a subtle bug in a sprawling codebase or a low-level system.
  • Security Coverage from Detection to Fix
    Its scope covers the whole loop of spotting, validating and repairing: 88.2% on SEC-bench Pro for real-world flaw discovery, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for source-level patching.
  • Two Weeks from Argon to Carbon
    Gemini 4 Argon arrived just a fortnight earlier. This release shows what happens when researchers and models improve side by side, pushing the frontier forward while safeguards mature at the same speed.

How Access to Gemini 4 Carbon Unfolds

Rollout happens in stages on purpose. Here is the order in which availability expands.

What Gemini 4 Carbon Can Do, and How It Scores

First place on the Artificial Analysis Intelligence Index and a 0.0% prompt-injection success rate: these are the figures that define the release.

First Place on Intelligence Index v5.0

A score of 59 puts Gemini 4 Carbon at the top of the freshly updated Artificial Analysis Intelligence Index v5.0, with large gains in coding, agentic work and the abilities that matter most for security defense.

Up 89 Elo to #4 on WebDev Arena

Cleaner artifacts and stronger full-stack output lift the model to 1768 points on WebDev Arena, a climb of +89 elo and a fourth-place finish.

Fewer Shortcuts, Less Cheating

Alignment tuning brought successful cheating on Terminal-Bench 4.0 down to 0.0%, with attempts at 0.4%, while accuracy rose; high-severity misalignment alerts in internal live coding environments dropped by nearly 80%.

More Honest About What It Knows

In internal tests, unsupported claims fell by roughly half compared with Argon. The model now folds methodological caveats into the deliverable, or declines a blocked action and lists exactly what would still be required.

Holds the Line on Indirect Prompt Injection

Gray Swan's Indirect Prompt Injection Benchmark recorded a 0.0% attack success rate against Gemini 4 Carbon, resisting every evaluated attempt, against 0.7% for Argon.

Finishes the Job Without Overstepping

Overreach dropped 26% and under-service fell 29%, so multi-step plans now execute end to end more reliably while respecting project constraints and staying inside scope.

FAQ

Gemini 4 Carbon: Your Questions Answered

Straight answers on cost, benchmark scores, access order and how this release compares with its predecessor.

1

So what exactly is Gemini 4 Carbon?

It is the newest and most capable frontier model we have shipped, aimed at top-tier cybersecurity and tasks that run for a long time. It sits at #1 on the updated Artificial Analysis Intelligence Index v5.0 with a score of 59.

2

What does it cost to use?

Introductory rates are $2 per million input tokens and $10 per million output tokens, with cached input discounted 95%. Once the introductory window closes, $4 per 1M input and $20 per 1M output apply.

3

Where does it rank on benchmarks?

Vulnerability discovery reaches 88.2% on SEC-bench Pro, exploitation 54.8% on ExploitGym, and patching 71.0% on CWE-bench, alongside 1768 points and #4 on WebDev Arena.

4

Can anyone try it today?

Not yet broadly. The first group is a hand-picked set of cyber defenders, with more trusted partners added under the Fairwind Program over time; developers, businesses and consumers follow through Google's product surfaces.

5

How far has it moved past Argon?

Internal benchmarks show over 2x improvement in writing secure code, roughly half as many unsupported claims, a 0.0% prompt-injection attack success rate and tighter scope discipline.

6

Which tasks suit it best?

Hard multi-step engineering, such as tracing and fixing subtle bugs in large codebases or low-level systems, plus agentic workflows, scientific work, full-stack web development and multimodal understanding.

Get Set for Gemini 4 Carbon

This release aims at real gains in software engineering and security work while keeping safety and reliability intact. Track the staged rollout through the Fairwind Program, with developer, business and consumer access on the way.