Gemini 4 Carbon Is the New #1 AI Model — and the Least Likely to Hallucinate

Introduction: A New Number One

Google's Gemini 4 Carbon, released at the end of last week, has taken the top spot on Artificial Analysis's Intelligence Index V5 — the benchmark suite that has become the de facto leaderboard for frontier models. It beats Claude Opus 5.5 by a single point and OpenAI's GPT-6 Astra by two, takes the number-one score on four of the nine individual evaluations, and does it at a lower cost per task than its closest rivals.

But the more interesting claim isn't the intelligence crown. Gemini 4 Carbon also tops Artificial Analysis's new AA-Omniscience v1.1 test, which tracks hallucinations and factual recall — making it, by that measure, the frontier model least likely to make things up. For anyone picking an AI assistant, a coding agent, or a research tool in 2026, that combination — smartest and least hallucination-prone — is exactly what everyone says they want. Here's what the numbers actually show, and when you can get your hands on the model.

The Leaderboard: 59 Points, Won by One

On the Artificial Analysis Intelligence Index V5, Gemini 4 Carbon (high reasoning) scores 59. The gaps behind it are narrow enough that the ranking could flip with the next release from any lab:

Two details stand out. First, cost: Artificial Analysis notes Carbon achieves its ranking at a lower cost per task than other high-ranking models like Opus 5.5 — a real consideration for anyone running agents at volume through Claude Code, Gemini CLI, or OpenAI's Codex, where token bills scale with how much work you hand the model. Second, breadth: number one on four of nine evaluations is a stronger claim than a single overall point, and it lines up with SemiAnalysis's GDPx benchmark published the same week, where Gemini 4 Carbon led the biology and software engineering domains.

It's also worth remembering what the one-point margin means in practice: Claude and GPT-6 Astra remain within rounding distance on general intelligence. If Carbon has a practical edge today, it's the price-performance and the hallucination numbers — not a blowout on raw capability.

The Hallucination Story: Best in Class, Still Not Good

The AA-Omniscience v1.1 test scores models on a −100 to +100 scale for knowledgeable recall — how accurately they retrieve facts versus inventing them. Gemini 4 Carbon tops the field. But the sobering part is the absolute number: even the best model scored just 54.

Read that twice. The least hallucination-prone frontier model in existence, in late 2026, is barely past the halfway point on a scale where +100 means consistently reliable recall. As governments and institutions push to extract value — and security — from AI deployment, "lowest hallucination score" is becoming a genuine selling point; Tom's Hardware's coverage notes exactly that. But the benchmark is really telling buyers that no frontier model deserves unsupervised trust on factual recall yet. Verification, citations, and human review remain part of the job — whichever model you pick.

The Cautious Rollout: Fairwind and Government Testing

Don't expect to flip a switch and get Carbon everywhere today. Google is taking a deliberately staged approach — notably similar to how Anthropic handled the Mythos release in 2025:

• U.S. government pre-release testing: Google says it is "actively engaged in the U.S. government's voluntary process for pre-release model testing" for Gemini 4.
• Cyber defenders first: initial access goes to a subset of security professionals through the new Fairwind program, with wider availability through trusted partners later.
• Guardrail iteration before general release: Google wants to harden Carbon against use for exploits and hacking before making it broadly available to developers.

The irony is easy to miss: a model being released cautiously because it's strong on security-adjacent work. GDPx showed frontier safety classifiers sandbagging cyber scores for Anthropic and OpenAI models; Google is instead front-loading security practitioners as the first users. For security teams, that's an early-access opportunity worth watching.

Which Tools Get Carbon First

When Carbon does reach general availability, expect it to flow into the usual places: the Gemini app, Gemini CLI, Google's Antigravity agent harness, and API access for third-party tools. If your workflow is sensitive to factual accuracy — legal research, medical summaries, financial briefings, customer support — it's worth planning an evaluation pass when access opens, rather than assuming your current model keeps the crown.

The Bottom Line

Gemini 4 Carbon is the new number one on the most-watched frontier leaderboard, it wins while costing less per task, and it posts the best hallucination score of any major model. That's a genuinely strong launch. But the fine print matters: the margin is one point, the rollout is deliberately slow, and even the "least hallucinating" model sits at 54 on a 200-point scale. The practical takeaway for 2026 is unchanged — pick models per task, verify outputs that matter, and treat every leaderboard crown as temporary. Comparing the tools built on these models side by side is exactly what a curated directory is for.

Frequently Asked Questions