Gemini 4 Carbon Debuts Top of Intelligence Index at 59
By @sharedot · · 6 pages
- AI Frontier
- Gemini 4 Carbon
- Benchmarks
Google's Gemini 4 Carbon tops the Intelligence Index at 59, narrowly beating Claude Opus 5.5, and posts the lowest hallucination score.
Google's comeback model takes the top spot
Google released Gemini 4 Carbon last week, and the model now sits at number one on Artificial Analysis's Intelligence Index V5 with a score of 59 — one point ahead of Claude Opus 5.5 (max with fallback) and two ahead of OpenAI's GPT-6 Astra (max).
Tom's Hardware frames the release as a comeback for a lab that, in its words, languished with irregular releases for much of 2026.
Carbon ranks first on four of the nine evaluations and does so at $7.76 per Intelligence Index task, several dollars cheaper than Claude Opus 5.5 ($10.04) and slightly cheaper than even Gemini 3.8 Flash (high).
The reliability claim that stands out
The result a devoted benchmark-watcher will talk about is hallucination. Carbon tops Artificial Analysis's AA-Omniscience v1.1 test of knowledge and hallucination with a score of 54, six points clear of Claude Opus 5.5 at 48.
Tom's Hardware attributes to Google the claim that Carbon cuts unsupported claims twice as much as Gemini 4 Argon, refuses actions it cannot complete rather than fabricating answers, and cut model overreach by 26%.
The caveat: 54 is on a -100 to +100 scale, and Tom's Hardware notes even the best 2026 frontier model is far from consistently reliable at recalling information.
A deliberately slow, security-first rollout
Access is the twist. According to Tom's Hardware, Google is taking 'a deliberate approach to rollout' and is engaged in the U.S. government's voluntary pre-release model testing process, giving cyber defenders first access through its Fairwind Program before wider developer availability.
Google says Carbon thwarted all attempts in the Gray Swan Indirect Prompt Injection benchmark to make it do something it shouldn't.
Tech Times notes Carbon had not been made generally available as of its publication — meaning SemiAnalysis tested it in Google's native Antigravity harness despite the limited release.
Where Carbon still trails, and why it matters
Carbon is not dominant everywhere. Tom's Hardware reports it loses to Claude Code with Claude Opus 5.5 at programming and trails Grok 4.7, Muse Wonder, and Kimi K3.1 Pro on knowledge work.
On SemiAnalysis's new GDPx agentic benchmark, Tech Times reports Opus 5.5 led at 12.0% overall while Carbon reached 10.9% — 91% of the score at roughly a third of the per-task cost ($16.29 vs $51.41) — and was the best biology model.
With Gartner finding only 3% of companies achieving cost and productivity gains from AI, per Tom's Hardware, a high-intelligence, low-hallucination, affordable model is a formula built for enterprise buyers.