Google unveils Gemini 4 Carbon with big cybersecurity upgrades

Google is moving at a breakneck speed right now. Just two weeks after dropping Gemini 4 Argon, the team is back with an even heavier hitter.

Google announced Gemini 4 Carbon, calling it its strongest frontier model yet. The headline capabilities here are two things the industry has been desperately waiting for: serious cybersecurity chops and true long-horizon agentic workflows.

On the freshly updated Artificial Analysis Intelligence Index v5.0, Carbon landed at #1 with a score of 59, edging out Claude Opus 5.5 and GPT-6 Astra.

A full-lifecycle cyber defender

Carbon’s biggest leap comes from defensive security. Google basically trained it to handle the whole cycle: finding the bug, seeing if it can actually be exploited, and writing the patch.

On external benchmarks, Carbon scored 88.2% on SEC-bench Pro for unearthing vulnerabilities, 54.8% on ExploitGym for testing exploitability, and 71.0% on CWE-bench for root-cause patching. Internally, Google’s Chrome team reported Carbon generated high-quality patches more than 10 times better than another lab's large frontier model. Cloud security firm Wiz is also using it for web penetration testing.

It’s also far less prone to going rogue. Carbon posted a 0.0% attack success rate on prompt injection benchmarks, and high-severity misalignment alerts dropped by nearly 80% during internal live coding tests. When it doesn't know something, it surfaces methodological caveats rather than hallucinating plausible-sounding nonsense.

Agents that stay on the clock for days

Beyond security, Carbon is engineered for persistence. Google claims internal teams have let the model run unattended for days at a time, grinding through dead ends to trace subtle bugs across massive codebases and low-level systems.

That said, it didn't sweep every single coding board. On the Artificial Analysis Coding Agent Index, Carbon scored 59 using the Antigravity CLI, sitting just behind Claude Opus 5.5 running Claude Code at 62. On pure software engineering suites like FrontierCode, it ranks behind Opus as well.

Where it shines is practical reasoning. It topped the Vending Bench evaluation for running a virtual business and led Blueprint Bench, which turns room photos into 2D floor plans.

On pricing, Google matches Argon’s introductory half-price discount: $2 per million input tokens and $10 for output, with cached inputs slashed by 95% to $0.20. Standard pricing will eventually sit at $4 and $20, though no discount end date was given.

Don't expect to plug it into your stack this afternoon, though. Google is running it through the U.S. government’s pre-release testing framework, rolling access out strictly to select cyber defenders via its Fairwind Program first.

Why it matters

Google spent a good chunk of 2026 looking like it was playing defense against OpenAI and Anthropic. Shipping two frontier models in less than a month changes that narrative entirely, and Carbon’s benchmarks give them real bragging rights.

Focusing heavily on cybersecurity is a brilliant wedge. Standard chatbot tasks are commoditized, but an agent that can autonomously audit legacy code and safely patch zero-days saves enterprise teams thousands of hours. At roughly 23% lower cost per task than Claude Opus 5.5 while generating 29% fewer output tokens, the economics make immediate sense.

Still, the obvious catch is availability. Right now, only a tight circle of security partners can touch it. Without independent developers vibe-checking these benchmarks in production, it’s hard to tell how much is benchmark optimization and how much translates to everyday engineering. And with frontier labs shipping on bi-weekly cadences, Google’s lead could look very different by the time Carbon opens to the public.