Google's Gemini 4 Carbon Tops AI Index by One Point, Claims Lowest Hallucination Rate

Google's Gemini 4 Carbon has taken the top spot on Artificial Analysis' Intelligence Index V5 with a score of 59, one point ahead of Anthropic's Claude Opus 5.5. The margin is thin. Rankings in 2026 change fast, and one point is well within the noise of any benchmark suite.

The model, released last week, leads four of the nine evaluations in the suite. It sits two points ahead of OpenAI's GPT-6 Astra (max), nine ahead of Meta's Muse Wonder Full (Max), and ten ahead of Claude Fable 5.1 (max with fallback). Artificial Analysis, the independent benchmarking service, provides all of these figures.

The release reads as a comeback. Google has spent much of the year with irregular model releases, and Carbon arrives with more fanfare than most.

What does the price look like?

Cost is the second half of the story. Gemini 4 Carbon's cost per task on the Intelligence Index test is $7.76, slightly cheaper than Gemini 3.8 Flash (high). That is several dollars below both Claude Opus 5.5 and Sonnet 5.5, and not far off half the price of Claude Fable 5.1 at $12.99 per task.

The picture is less flattering further down. Carbon remains more expensive than OpenAI's GPT 5 Astra, roughly six times the price of GPT 6.1 Sol (max), and about 13 times the cost per task of the open-weight DeepSeek V4.1 Flash 1017. That DeepSeek model scores just 32 on the index, to be fair.

Arena.ai places Carbon at the top of its Pareto Frontier chart at higher price points. At lower price points, Chinese open-weight models take over — GLM 5.3-Flash and DeepSeek V4 Pro lead there.

How reliable is the reliability claim?

Google's strongest claim concerns hallucinations. Gemini 4 Carbon tops Artificial Analysis' new AA-Omniscience v1.1 test, which tracks hallucinations and recall of information, beating competitors by what the service calls significant margins.

But the top score is 54 on a scale running from -100 to +100. Even the best frontier model of 2026 remains far from consistently accurate recall. A score of 54 is a marketing headline and a caveat in the same number, and buyers should read it as both.

Google also makes claims it has not fully quantified. It says Carbon cuts unsupported claims twice as much as Gemini 4 Argon, a model it debuted two weeks ago. It claims the model refuses tasks it cannot complete rather than fabricating answers, and that it cuts model overreach by 26%. These are vendor figures, not independent ones.

The one external safety result cited: Carbon thwarted all attempts on the Gray Swan Indirect Prompt Injection benchmark to make it act against instructions.

Why is the rollout slow?

Access comes in stages. Google is offering the model first to a subset of "cyber defenders" through its Fairwind program, with wider availability through trusted partners later. Anthropic followed a similar pattern with Mythos in 2025.

"We're taking a deliberate approach to rollout and are actively engaged in the U.S. government's voluntary process for pre-release model testing," Google's launch statement reads.

The company says it wants to iterate on guardrails before opening Carbon to developers, making the model harder to use for exploits and hacking. Voluntary screening is not regulation. But the practice appears to be settling in as a fixture of major releases after a year in which oversight proposals rose and fell among the big labs.

Where does Carbon fall short?

It is not the best at everything. Artificial Analysis notes Carbon trails Claude Code with Claude Opus 5.5 at programming, and falls behind Grok 4.7, Muse Wonder, and Kimi K3.1 Pro on knowledge work.

The commercial context matters here. Gartner's latest research shows just 3% of companies achieve cost and productivity improvements from AI use. Hallucinations make return-on-investment tracking harder still.

That frames Google's bet: high intelligence, low hallucination scores, and a relatively low price might be the winning formula. The formula holds until the next model drops, and by 2026's pattern, that will not take long.