What happened
Artificial Analysis reported on October 28, 2026 that Google's Gemini 4 Carbon is the new leader of its Intelligence Index v5. Carbon scored 59, and the evaluator attributes the result to strong complex reasoning, knowledge, and scientific-task performance. It takes the top score in four of the index's nine evaluations: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience, and AutomationBench-AA. The source is careful not to claim universal leadership: Carbon made substantial gains in coding and knowledge work, but does not lead every category.
The coding-agent comparison is one of the more consequential parts of the result for software teams. With Antigravity CLI and its high reasoning setting, Gemini 4 Carbon scored 59 on the Artificial Analysis Coding Agent Index. That is three points behind Claude Code using Claude Opus 5.5 at its max setting, which scored 62. Carbon was second on Terminal-Bench 4.0 at 64% and FrontierCode v1.1 at 51%; the source says Opus 5.5 led FrontierCode v1.1 with 56%. Artificial Analysis characterizes the improvement as notable because Google had historically underperformed in this area.
The report also provides a more mixed view of agentic knowledge work. On AA-Briefcase, Carbon at high reasoning scored 1609 Elo, a 126-point rise over Gemini 4 Argon at high reasoning, which scored 1483. Its largest named improvement was Presentation, up 293 Elo to 1598; Analytical Quality rose 120 points to 1671. Yet it remained behind Grok 4.7, Muse Wonder, and Kimi K3.1 Pro on the cited AA-Briefcase comparison. On AA-Omniscience v1.1, Carbon scored 54, six points ahead of Opus 5.5 and eight ahead of Argon. Accuracy rose ten points over Argon to 58%, while the hallucination rate rose from 4% to 11%.
Cost is presented in completed benchmark tasks rather than merely token rates. At standard pricing, Artificial Analysis estimates Carbon at $7.76 per Intelligence Index task, compared with $10.04 for Opus 5.5 max and $6.25 for GPT-6 Astra max. Carbon used 123,000 output tokens per task, 29% fewer than the 173,000 reported for Opus 5.5 max. Its standard price is $4 per million input tokens and $20 per million output tokens, identical to Gemini 4 Argon's stated standard price; cached input is priced at $0.20 per million at standard rates. Google is offering a 50% launch discount with no confirmed end date, but the charts use standard pricing. Carbon has a one-million-token context window, accepts text, image, video, and speech input, and produces text output. It is being rolled out to selected users and is not publicly available.
What the archive adds
This is a new fact on a known story: Artificial Analysis launches Intelligence Index v5, with 60% of the score on private test sets covered the index launch and its revised evaluation design. Carbon's result is the first reported leadership change under that v5 framework. The archive also records an AA-Omniscience v1.1 update, a benchmark now among Carbon's four index-leading evaluations.
Trend signal
The report supports Agent economics moves to completed work. Its useful comparison is not only the $4/$20 token price: it estimates $7.76 for an Intelligence Index task and reports output-token consumption per task. That does not prove a production workload will share the same economics, but it is a concrete example of evaluation moving toward work completed under a defined task suite.
Why now
The timing follows the launch of Intelligence Index v5 on October 27. Artificial Analysis says v5 includes new benchmarks and upgrades and, for the first time, places the majority of its scoring on private datasets. Carbon's 59 therefore changes the leading score from the initial v5 result within a day, while also making the new index design immediately relevant to model selection discussions.
Connections to other events
The new result meets the v5 launch story because the same index now has a different leader. It also intersects with the archive's Artificial Analysis releases AA-Omniscience v1.1 with a 3,000-question set and new calibrated public subset: AA-Omniscience v1.1 is one of the four evaluations where Carbon leads, and the source reports both its accuracy and hallucination rate there.
Why it matters
Teams deciding which frontier model to evaluate should treat Carbon as a serious candidate for long-context, scientific, terminal, and coding-agent workflows, not as a generic claim of overall superiority. The reported coding score is close to the leading Claude configuration, while the source's standardized task-cost estimate is lower than that configuration's. Those are useful hypotheses for a local evaluation plan.
But access changes the practical decision. A model that selected users can try but most teams cannot call cannot yet be a dependable production default. Moreover, an aggregate score cannot settle whether Carbon's gains map to a team's repository, tools, policies, data, or review process. The AA-Omniscience result also shows why teams should inspect quality dimensions separately: higher accuracy came with a higher hallucination rate than Argon's.
Limits and uncertainty
This Field Note reports Artificial Analysis' published measurements and comparisons; it does not independently reproduce them. The source does not specify which selected users have access, when broad availability will arrive, or whether the 50% launch discount will remain in place. Its cost-per-task figures are estimates for Intelligence Index tasks at standard pricing, not quotes for a team's own workload. Benchmark and agent-harness choices matter: the coding-agent result specifically uses Antigravity CLI, while the leading comparison uses Claude Code. The source also does not establish that its private or held-out tests predict every production domain.
Skill to build now
Create a small, versioned evaluation pack before Carbon becomes generally available. Include representative terminal tasks, knowledge-work artifacts, long-context inputs, expected outputs, tool permissions, human review criteria, and a budget ceiling. Record task completion, failures, retries, output tokens, latency, and reviewer corrections. Run the same pack against the currently available model and against Carbon when access arrives. This turns an index result into a reproducible procurement and deployment decision rather than a reaction to a leaderboard position.
Reader takeaway
Gemini 4 Carbon is now the top model on Artificial Analysis' Intelligence Index v5 and is close to the leading coding-agent score at a lower reported benchmark-task cost than Opus 5.5. That is strong evidence to prepare an evaluation, not evidence to migrate today: Carbon remains restricted to selected users, and a benchmark lead still needs validation against your own completed work.