Google's Gemini 4 Carbon has claimed the #1 position on Artificial Analysis's Intelligence Index v5 with a score of 59, driven by leading performance in complex reasoning, scientific tasks, and knowledge benchmarks. It tops 4 of the index's 9 evaluations — HLE-Diamond, Terminal-Bench-Science, AA-Omniscience, and AutomationBench-AA — while posting substantial gains over its sibling Gemini 4 Argon across the board. The coding story is more nuanced and arguably more interesting. Carbon scores 59 on the Coding Agent Index using Antigravity CLI, sitting just 3 points behind Claude Code with Opus 5.5 at 62. It ranks second on both Terminal-Bench 4.0 (64% vs 56%) and FrontierCode v1.1 (51% vs 56%). For Google, historically weak in agentic coding, this is a notable inflection — not yet dominance, but competitive parity where none existed. Knowledge work tells a similar story of progress without supremacy. On AA-Briefcase, Carbon posts 1609 Elo, a 126-point jump over Argon's 1483, but still trails Grok 4.7 (1638), Muse Wonder (1659), and Kimi K3.1 Pro (1678). The biggest single gain — a 293-Elo leap in Presentation to 1598 — suggests Google is closing specific capability gaps rather than achieving uniform superiority. The hallucination numbers deserve attention. Carbon scores 54 on AA-Omniscience v1.1, six points ahead of Claude Opus 5.5 and eight ahead of Argon. Accuracy rises 10 points over Argon to 58%. The tradeoff: hallucination rate climbs from Argon's 4% to 11%, still dramatically lower than the 50% posted by both Claude Opus 5.5 and GPT-6 Astra, but a directional warning that pushing accuracy upward pulls hallucination rates with it. The economics are competitive. At standard pricing of $4/$20 per million input/output tokens, Carbon costs $7.76 per Intelligence Index task — 23% less than Claude Opus 5.5's $10.04, though 1.2x GPT-6 Astra's $6.25. It achieves this partly through token efficiency, using 123k output tokens per task versus Opus 5.5's 173k. A 50% launch discount with no confirmed end date sweetens the entry, and cached input tokens receive a 95% discount at $0.20 per million. The competitive landscape this reveals is a three-way price-performance race between Google, Anthropic, and OpenAI where no single model dominates all dimensions. Carbon leads on reasoning and science, Opus 5.5 on coding agents, and GPT-6 Astra on cost efficiency. The 1M-token context window and multimodal input (text, image, video, speech) are table stakes. The real signal is that frontier model differentiation is increasingly about specific task profiles and cost curves rather than aggregate supremacy. Carbon is currently rolling out to selected users and is not publicly available. The 50% launch discount, standard pricing parity with Argon, and 9-point Intelligence Index improvement over its sibling suggest Google is pricing for adoption while it holds the benchmark crown — a window that historically closes within weeks as competitors respond.
