A release in quick succession and the claimed performance
Google DeepMind made Gemini 4 Carbon available just two weeks after Argon, the model that precedes it. According to The Decoder, Carbon takes the top spot on the Artificial Analysis Intelligence Index in version 5.0 with 59 points, while Argon stops at 50. The improvement covers both agentic tasks and coding, but Google also reports advances in scientific reasoning, multimodal understanding and full-stack web development. On WebDev Arena, the model gains 89 Elo points and places fourth with 1,769 points.
The company says it designed Carbon for more complex, long-horizon engineering problems that require persistent agentic workflows. According to Google, internal teams used it for several consecutive days on root-cause analysis and fixing subtle bugs in large codebases.
The comparison with competitors
The competitive picture remains mixed. In agentic coding on Antigravity, Carbon rises from 53 to 59 points compared with Argon but remains three points behind Claude Code with Opus 5.5. On AA-Briefcase, it gains 126 Elo points over its predecessor, reaching 1,609, driven by presentation quality and analysis, but remains behind Kimi K3.1 Pro, Muse Wonder and Grok 4.7. Response accuracy rises from 48 to 58 percent, while the hallucination rate increases from 4 to 11 percent: a higher figure, the source notes, but still the second-lowest in the ranking. On the Vals AI platform, Carbon is second with 70.36 percent, behind Claude Sonnet 5.5 and ahead of Claude Opus 5.5.
On the cost side, the model consumes more resources than Argon: tasks in the Finance and Tax Agent benchmark take more than twice as long, and in legal research the average exceeds 50 minutes per task, with 2.3 times the reasoning tokens and 64 percent more generated tokens.
Security, gradual access and pricing
Google justifies the gradual rollout with the model's advanced security capabilities, designed to cover the entire cybersecurity defense cycle, from detecting and validating vulnerabilities to remediating them. In tests conducted without safety guardrails, Carbon reaches 88.2 percent on SEC-bench Pro and 54.8 percent on ExploitGym; on Gray Swan's indirect prompt injection benchmark, the company says it reduced the attack success rate to zero, compared with 0.7 percent for Argon, while noting that the test is saturated and this does not mean the problem has been solved. Access begins this week for selected experts and will then expand through the Fairwind program, while opening to developers, enterprises and consumers will come after further feedback and adjustments. The introductory price is identical to Argon's: $2 per million input tokens and $10 per million output tokens.
