Google unveils Gemini 4 Carbon with advanced security

Google DeepMind has unveiled Gemini 4 Carbon, the latest flagship in its Gemini line, following the rollout of Gemini 4 Argon. The release positions Carbon as a leading frontier model, with independent evaluations placing it ahead of Claude Opus 5.5 and GPT-6 Astra. In the Artificial Analysis Intelligence Index v5.0, Carbon achieved 59 points (ranked first), compared with 58 for Claude Opus 5.5 and 57 for GPT-6 Astra.

The model also reported only 11 percent hallucinations on the AA-Omniscience evaluation, a figure cited as notably lower than competing architectures, which reportedly exceeded 50 percent error rates on the same tests. In a year-long simulated scenario known as Vending-Bench 2, Carbon generated operational profits of $16,484.

Beyond standard reasoning, Google asserts that Gemini 4 Carbon demonstrates strong performance across defensive and offensive cybersecurity tasks. The official DeepMind release notes the model reached 88.2 percent on SEC-bench Pro for locating complex zero-day defects, 54.8 percent on ExploitGym for assessing actionable attack vectors, and 71.0 percent on Collinear CWE-bench for synthesising secure code patches, with the system producing secure code at more than twice the rate of its predecessor.

The company describes the model as having prompt injection immunity, recording a 0.0 percent breach rate on Gray Swan testing, and notes that engineers curbed speculative behaviours by prompting for clarification when faced with ambiguities rather than overstepping boundaries.

Access to Gemini 4 Carbon is being managed via the Fairwind Program, targeting vetted national defence agencies and critical infrastructure operators, with eventual plans to broaden to enterprise developers and retail subscribers.

The accompanying pricing proposal features introductory rates of two dollars per million input tokens and ten dollars per million output tokens, with standard pricing at four dollars and twenty dollars respectively; these figures are said to align with Claude Opus 5.5 while undercutting GPT-6 Astra. The article attributes these claims to the official Google DeepMind research announcement.