Gemini 4 Carbon Tops the Chart, While the Comments First Ask When It Can Be Used

Artificial Analysis has updated Intelligence Index v5. Gemini 4 Carbon ranks first with 59 points, and second on Coding Agent Index with 59 points, behind only Claude Code with Claude Opus 5.5 at 62 points. The model is not yet public and is only available to some users. Google is offering a 50% launch discount with no end date. The leaderboard arrived first; the product has not. That is the most noteworthy part of the whole story.

Artificial Analysis is a model evaluation organization. Intelligence Index v5 is its composite index, and Coding Agent Index looks at how coding agents perform on real tasks. Scores are not the same as everyday experience, but they can show which tasks a model is strong on and where it still falls short.

Gemini 4 Carbon took first place in 4 of the 9 evaluations in Intelligence Index v5: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience, and AutomationBench-AA. It stands out on complex reasoning, knowledge, and science tasks, and has made clear progress on coding and knowledge work, but it is not ahead across the board.

On coding agents, Carbon with Antigravity CLI scores 59, 6 points higher than Gemini 4 Argon and behind Claude Code with Claude Opus 5.5 at 62. It scores 64% on Terminal-Bench 4.0 and 51% on FrontierCode v1.1, the latter trailing Claude Opus 5.5 max at 56%. Google has not historically been strong in coding; this time it has caught up to second place, and the size of the improvement is more striking than the ranking itself. On SWE-Atlas-QnA-AA it scores 64%, while Muse Code with Muse Wonder Full scores 69%, putting Carbon in the middle of the pack.

On knowledge work, Carbon scores 1609 Elo on AA-Briefcase, 126 points above Argon's 1483, but still behind Grok 4.7 xhigh at 1638, Muse Wonder max at 1659, and Kimi K3.1 Pro at 1678. The biggest gain is in Presentation, up 293 Elo to 1598; Analytical Quality rises 120 points to 1671. The scores are rising, but it is not yet at the very front.

On AA-Omniscience v1.1, Carbon scores 54, 6 points above Claude Opus 5.5 max at 48 and 8 points above Argon at 46. Accuracy rises from 48% for Argon to 58%, while the hallucination rate rises from 4% to 11%. At 11%, it is still the second-lowest among all models, higher only than Argon; Claude Opus 5.5 max and GPT-6 Astra max both have hallucination rates of 50%. Accuracy and hallucination rate are rising together, and that trade-off needs to be considered in specific use cases.

On cost, Carbon spends $7.76 per Intelligence Index task, about 23% less than Claude Opus 5.5 max at $10.04 and about 1.2 times more than GPT-6 Astra max at $6.25. It outputs 123k tokens per task, 29% less than Claude Opus 5.5 max at 173k. Standard pricing is $4 per million input tokens and $20 per million output tokens, the same as Gemini 4 Argon; cached input gets a 95% discount at $0.20 per million. At the same price, Carbon's cost per task is close to Argon's $7.68, with an Intelligence Index score 9 points higher.

Model details: 1M-token context window; supports text, image, video, and audio input, with text output.

The comments are not as tidy as the leaderboard. One user directly asks: if the model isn't open, the scores don't count as real. Some suspect the benchmark was specifically optimized for, and some want to wait for engineers they trust to verify it before using it. Jon C's doubt is more specific: a company can publish AA scores before a full release, and this needs standardization; if Fable 5.5 is released before Carbon and scores higher, then Carbon was never truly on top. Another user, Jon Cheney, separates his use of models—one for thinking through problems and one for writing code—and says the people around him who build every day are no longer loyal to a single lab.

For more analysis, see Artificial Analysis's model page: https://artificialanalysis.ai/models/gemini-4-carbon

The signal from this leaderboard is not complicated: Gemini 4 Carbon stands first on knowledge, science, and general reasoning, has closed in on Claude Opus 5.5 in coding, costs less than Opus 5.5, and still keeps its hallucination rate low. The problem is also out in the open: the model is not public, the discount has a deadline but no date, and the comments are waiting for real availability. Between the top of the leaderboard and everyday experience lies a timeline; this time Google has released the scores first, while the product is still on the way.