Gemini 4 Carbon Tops Frontier Rankings With Fewest Hallucinations
By @sharedot · · 6 pages
- AI Frontier
- Gemini 4 Carbon
- AI Benchmarks
Google's new Gemini 4 Carbon narrowly beats Claude Opus 5.5 on the Intelligence Index and posts the best hallucination score, per Artificial Analysis.
A narrow win at the top of the index
Tom's Hardware reports that Google's Gemini 4 Carbon now holds the number one spot on Artificial Analysis' Intelligence Index V5, scoring 59 — a single point ahead of Claude Opus 5.5 (max with fallback) and two points ahead of OpenAI's GPT-6 Astra (max).
It ranks first on four of the nine evaluations in the suite.
Tom's Hardware frames the release as a comeback for a lab that languished with irregular model releases for much of 2026, and notes the model achieves the top ranking at a lower cost per task than its closest rivals: $7.76 per Intelligence Index task, roughly 23% less than Claude Opus 5.5's $10.04 and using 29% fewer output tokens per task.
The hallucination angle could be the real story
Perhaps more consequential for practitioners: Tom's Hardware says Gemini 4 Carbon tops Artificial Analysis' AA-Omniscience v1.1 test tracking hallucination and knowledgeable recall, scoring 54 on a -100 to +100 scale — six points ahead of Claude Opus 5.5 (max, 48).
Google claims the model cuts unsupported claims twice as much as Gemini 4 Argon, refuses actions it cannot complete rather than fabricating answers, and cuts model overreach by 26%.
But Tom's Hardware adds a sobering caveat: even the best score of 54 suggests frontier models in 2026 remain far from consistently reliable at accurately recalling information, and Carbon answers more questions correctly than Argon (58% vs 48% accuracy) only at a higher hallucination rate (11% vs 4%).
Careful rollout: cyber defenders get it first
Google is deliberately staging access, per Tom's Hardware. Its launch statement says the company is 'actively engaged in the U.S. government's voluntary process for pre-release model testing,' echoing Anthropic's cautious 2025 handling of Mythos.
Carbon is initially open to a subset of 'cyber defenders' through the Fairwind Program, with wider developer availability to follow only after guardrails are iterated to prevent misuse for exploits and hacking.
Google also highlights security results: the model thwarted all attempts on the Gray Swan Indirect Prompt Injection benchmark.
Tom's Hardware notes where Carbon falls short — it trails Claude Code with Opus 5.5 at programming, and Grok 4.7, Muse Wonder and Kimi K3.1 Pro on knowledge work — and that arena.ai places it top of the Pareto Frontier only at the expensive end, with open-weight GLM 5.3-Flash and DeepSeek V4 Pro leading lower price tiers.
Context: GDPx already dampened the hype
The benchmark victory lands just as reliability claims face fresh scrutiny.
Per Tech Times, SemiAnalysis' new GDPx benchmark tested 13 frontier models across 206 real professional tasks in 11 domains, and leader Claude Opus 5.5 scored only 12.0% overall — with every model scoring zero under binary grading on finance and legal tasks.
On that leaderboard, Gemini 4 Carbon placed second at 10.9% overall, reaching 91% of Opus 5.5's score at approximately 32% of the per-task cost ($16.29 vs $51.41).