Google Launches Gemini 4 Carbon, Boosting Cybersecurity and Long-Running Task Capabilities With Limited Access

Google announced Gemini 4 Carbon on October 28, emphasizing its ability to carry out multi-step engineering tasks over extended periods, as well as discover, validate, and fix vulnerabilities. Access remains limited: the model is available to select cybersecurity defense teams this week, with access set to expand to trusted partners through the Fairwind program. Google has not announced when it will become available to developers and consumers generally. Official announcement

Google reported scores of 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym for exploit testing, and 71.0% on CWE-bench for vulnerability remediation. These benchmarks respectively measure finding flaws, confirming exploitability, and fixing root causes. Google explicitly notes that the external benchmarks were run with safeguards disabled; success rates in production deployments remain to be validated. Evaluation conditions

Independent evaluations offer another perspective. Artificial Analysis’ Intelligence Index v5 gives Carbon a score of 59, up from 50 for Argon; its coding-agent index ranks Carbon second. That agent evaluation used Antigravity CLI, so the results reflect the influence of the agent tools and execution environment and should not be treated as a measure of the base model alone. The evaluator also notes that Carbon has improved in knowledge work but does not lead across the board. Independent evaluation

For deployment, the evaluation page lists a one-million-token context window and support for text, image, audio, and video inputs, with text output. Google’s introductory pricing is $2 per million input tokens and $10 per million output tokens; after the promotional period, rates will be $4 and $20, respectively. Cached inputs receive a 95% discount. The promotion’s end date has not been set. Model specifications, official pricing

For engineering teams, useful metrics to track next include vulnerability remediation success rates with safeguards enabled, the number of human interventions required for long-running tasks, and the cost per fix that passes testing. These deployment measures will show whether benchmark gains translate into practical engineering outcomes.