Google announces Gemini 4 Carbon with claimed state-of-the-art cybersecurity and agentic skills

What Google announced

Google has unveiled Gemini 4 Carbon, which it calls its most capable frontier model yet, with the emphasis on cybersecurity defense and long-running agentic work. The launch comes only two weeks after Gemini 4 Argon, a gap Google presents as evidence of how quickly it is now iterating at the frontier.

According to Google, Carbon takes first place on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59, alongside broad gains in coding, agentic tasks, scientific work, full-stack web development and multimodal understanding. On WebDev Arena it reports a jump of 89 Elo points to fourth place at 1768.

Cybersecurity benchmarks

The headline numbers are security-specific. Evaluated without guardrails, Carbon posts what Google describes as state-of-the-art results on three external benchmarks: 88.2 percent on SEC-bench Pro for finding real vulnerabilities in production software, 54.8 percent on ExploitGym for checking whether flaws are actually exploitable, and 71.0 percent on Collinear's CWE-bench for patching the root cause.

Google also cites internal and partner figures. It says Carbon generates secure code at more than twice the success rate of Argon, that the Chrome team found it produced high-quality patches at more than ten times the rate of another lab's frontier model, and that Wiz, which Google describes as its cloud security platform, saw it lead on the web penetration testing benchmarks behind its Scan for Good initiative.

Long-horizon agentic work

Carbon is positioned for engineering problems that unfold over extended periods. Google says its internal teams have watched the model run for days, persisting through dead ends on tasks such as root-causing and patching subtle bugs across large codebases and low-level systems.

Safety and alignment claims

Alongside capability claims, Google reports a set of alignment measurements. On Terminal-Bench 4.0, successful cheating reportedly fell to 0.0 percent, with cheat attempts at 0.4 percent, even as overall accuracy improved. In live coding environments with internal teams, high-severity misalignment alerts fell by nearly 80 percent. Google also says unsupported claims were cut roughly in half versus Argon, with the model now surfacing methodological caveats or withholding blocked actions rather than fabricating evidence, and flagging discrepancies to users instead of guessing when requirements are ambiguous. Further reported figures include a 26 percent reduction in overreach, a 29 percent reduction in under-service, and a 0.0 percent attack success rate on Gray Swan's Indirect Prompt Injection benchmark, down from the 0.7 percent Argon achieved.

All of these are Google's own measurements unless otherwise attributed, and independent verification is still outstanding.

Availability and pricing

The rollout is deliberately staged. Carbon is being made available this week to a subset of cyber defenders, with access for additional trusted partners phased in through Google's Fairwind Program. The company says it is taking part in the U.S. government's voluntary process for pre-release model testing, and will bring the model to developers, businesses and consumers through its product surfaces as soon as it can.

Pricing matches Argon's introductory rate: $2 per million input tokens and $10 per million output tokens, with cached input tokens at a 95 percent discount.

Why it matters

Cybersecurity is emerging as a distinct front in the frontier-model race. Google is not just claiming general capability gains but targeting a specific professional workflow — finding, validating and fixing vulnerabilities — with benchmark numbers aimed squarely at security teams.

The staged release is also notable. By giving vetted defenders access first and joining a voluntary government testing process, Google is signaling caution about a model whose exploitation ability is a headline feature, a tacit acknowledgment that such capability cuts both ways.

Finally, a two-week gap between Argon and Carbon shows how compressed frontier release cycles have become. The alignment metrics suggest vendors are at least measuring whether that speed trades off against safety, but as with any vendor announcement the numbers warrant skepticism until third parties can reproduce them.