Gemini 4 Carbon tops one AI benchmark as other leaderboards disagree

Google takes the headline score

Gemini 4 Carbon, released last week, ranks first on four of the nine evaluations in the Artificial Analysis Intelligence Index V5 suite. It costs $7.76 per task, against $12.99 for Claude Fable 5.1. It is two points ahead of GPT-6 Astra. Tom's Hardware framed the result as a comeback for Google. Google had languished with irregular releases for much of the year.

Benchmark scores shape which models enterprises and governments trial, and what they pay per task. Artificial Analysis, Semi Analysis GDPx, CloudOpsBench and vulnerability tests point to different leaders. No single score settles a procurement decision. How any of these results translate into production work is still unknown.

On AA-Omniscience v1.1, which tracks hallucinations and recall of information, Gemini 4 Carbon scored 54 on a scale from -100 to +100. That is the best result in the comparison. It is also a long way from consistent reliability, as Tom's Hardware noted. Tom's Hardware says the model does so at lower cost per task than other high-ranking models, including Claude Opus 5.5.

A second benchmark names a different winner

OfficeChai reported on 2 November that Claude Opus 5.5 topped Semi Analysis' GDPx benchmark, with Gemini 4 Carbon in second place. Different boards measure different work, so both results can stand. For a buyer, the crown depends on which board gets quoted.

The Hacker News reported on vulnerability benchmarks. Gemini 4 Carbon led tests for finding, exploiting and patching vulnerabilities.

CloudOpsBench's leaderboard ranks Claude Fable 5.1 first at a 60.6% resolution rate. GPT-6 Astra follows at 58.9%, Gemini 4 Argon, not Carbon, at 57.5%, and Grok 4.7 (high) on cursor-cli fourth at 56.5%. Claude Opus 5.5 sits seventh at 47.8% on terminus-2.

Harness choice matters as much as the model. Claude Fable 5.1 falls to 43.8% on terminus-2. Claude Opus 5.5 gets 46.7% on claude-code.

The same board prices the work. The top entry ran 38.6 hours at $404.30, against 34.2 hours and $100.50 for Gemini 4 Argon.

Task-specific scores split the field

Tom's Hardware reported that Gemini 4 Carbon is not as good as Claude Code with Claude Opus 5.5 at programming. It falls behind Grok 4.7, Muse Wonder and Kimi K3.1 Pro on knowledge work. Its cost per task is slightly cheaper than Gemini 3.8 Flash (high). It is about six times more expensive than GPT 6.1 Sol (max).

The open-weight comparison is harsher. Gemini 4 Carbon is about 13 times as expensive per task as DeepSeek V4.1 Flash, which scores 32 on the Intelligence Index.

Arena.ai has Gemini 4 Carbon top on its Pareto Frontier chart at the expensive end of the spectrum. At lower price points, GLM 5.3-Flash and DeepSeek V4 Pro take over. CloudOpsBench shows the same gap further down its table: Qwen3.8-Max at 33.7%, DeepSeek V4.1 Flash at 21.8% and Kimi K3.1 Pro at 19.1%.

For now, Tom's Hardware writes, Gemini 4 Carbon may be the best at most of the high price points. That could change quickly.

What buyers still cannot verify

Google is offering access to Gemini 4 Carbon for cyber defenders first through its Fairwind program. Wider availability is coming later. The company has not given a date.

Google's announcement talked up collaboration with the U.S. government. It cited model screening and pre-release testing.

Any single index top spot is a claim, not a settled fact. That holds until independent testing on real workloads appears.

Tom's Hardware notes that AI model rankings change regularly in 2026.