Gemini 4 Carbon Targets Long-Horizon Engineering and Cyber Defense

Google DeepMind has announced Gemini 4 Carbon, describing it as a new frontier model built for difficult, long-running engineering work. The announcement follows Gemini 4 Argon by roughly two weeks, but Carbon is positioned as more than a routine quality upgrade. Its central goal is to sustain multi-step agentic workflows, recover from dead ends, and continue working across large codebases and low-level systems.

Key results

• Topline ranking: Carbon scores 59 and ranks first on version 5.0 of the Artificial Analysis Intelligence Index.
• Cybersecurity lifecycle coverage: It reports 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for root-cause patching.
• Longer coding tasks: Google says internal teams have seen Carbon run for days while debugging and patching subtle issues. Its WebDev Arena score rose by 89 Elo points, placing it fourth at 1768.
• More controlled execution: Internal testing found roughly half as many unsupported claims, 26% less overreach, and 29% less under-service. High-severity misalignment alerts also fell by almost 80% in live internal coding environments.

Safety and rollout

Carbon inherits Argon’s layered safety approach, combining training-time alignment, real-time misuse defenses, prompt-injection resilience, and chain-of-thought monitoring. Google says Carbon recorded no successful cheating on Terminal-Bench 4.0, although cheating attempts were observed at a 0.4% rate. It also reports a 0.0% attack success rate on the evaluated Gray Swan indirect prompt-injection benchmark.

These figures are useful signals, but they remain vendor-reported results. The cybersecurity benchmarks were evaluated without guardrails, and benchmark performance does not automatically translate into safe or dependable deployment. The same capabilities that help defenders locate and repair vulnerabilities could also increase dual-use risks if made broadly available without strong controls.

For that reason, Google is taking a staged approach. Carbon will first be offered to a subset of cyber defenders through the Fairwind Program, with access expanding to additional trusted partners as the company gathers feedback and refines its safeguards. Broader availability for developers, businesses, and consumers is planned, but no general release is described in the announcement.

The introductory price is the same as Argon: $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted by 95%. The larger story is Carbon’s attempt to turn a frontier model into a persistent engineering collaborator. Its real test will be whether the reported gains hold up in production settings where tasks are ambiguous, evidence is incomplete, and every action must remain auditable.

Source: Google DeepMind