Did Google "preview" the strongest Gemini 4 model again?
Gemini 4 Carbon ranks first on the Artificial Analysis Intelligence Index v5.0 (score 59, versus 58 for Claude Opus 5.5), and claims leads on defense-oriented benchmarks for vulnerability discovery, exploit verification, and patch fixing. But it isn't usable yet: it's first being given to some cybersecurity defenders and trusted partners (the Fairwind program), and is taking part in voluntary pre-release testing with the US government, before being gradually opened to developers, enterprises, and consumers.
https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon
Capability narrative: harder, longer engineering tasks
Carbon targets harder engineering problems that span longer periods, suited to sustained Agent workflows. Internal teams say it can run for days, keep pushing forward after getting stuck, for example locating root causes and writing patches in large codebases and low-level systems. Feedback-driven improvements cover coding, agents, scientific tasks, full-stack web, and multimodal understanding. On web development, it is said that after improved output quality, it rose about +89 Elo on WebDev Arena, reaching 4th place (about 1768 points).
Cybersecurity
Officially, there is a significant leap across the whole chain of "discovery → verifying exploitability → fixing the root cause"; external benchmarks are all evaluated without added guardrails:
· SEC-bench Pro about 88.2% (autonomously finding complex vulnerabilities in real software)
· ExploitGym about 54.8% (judging whether a vulnerability can be exploited, to help prioritize)
· Collinear CWE-bench about 71.0% (turning verified findings into working patches)
Internal metrics also include: compared with Argon, the success rate for generating secure code is claimed to have improved by more than about 2x; the Chrome team says its high-quality patch performance is more than about 10x stronger than "a large frontier model from another lab"; cloud security company Wiz also says it leads on benchmarks related to web penetration testing (tied to Wiz's Scan for Good). These are mostly partner or internal observations, and the claims need to be verified separately.
Safety and alignment
It keeps Argon's multi-layer safety stack (training, real-time abuse protection, prompt injection resilience, chain-of-thought monitoring, etc.), and claims to have taken alignment a step further, especially on high-pressure or complex tasks, doing "what should be done":
· On Terminal-Bench 4.0, successful cheating dropped to 0.0%, with cheating attempts at about 0.4%, while overall accuracy still improved; in internal live coding, high-severity misalignment alerts dropped about 80%.
· Compared with Argon, unsupported assertions are roughly halved; it more often adds methodological qualifiers to its deliverables, or explicitly lists what evidence is missing and why it hasn't acted yet.
When faced with ambiguity, it leans toward first pulling sources and surfacing contradictions to the user before acting.
· Overreach dropped about 26% and under-service about 29%, making multi-step plans more likely to be completed without exceeding scope.
· On the Gray Swan indirect prompt injection (IPI) benchmark: Argon's attack success rate was about 0.7%, and Carbon is claimed at 0.0%.
Release and pricing
Starting this week, it opens to a subset of cybersecurity defenders, then expands gradually to more trusted partners through Fairwind; after collecting feedback and iterating on guardrails, it will reach product surfaces as soon as possible. The introductory pricing is the same as Argon: about $2 per million input tokens, about $10 per million output tokens, with cached input at about a 95% discount off the input price. After the introductory period ends, it will become about $4 input and $20 output (per million tokens).