Gemini 4 Carbon Tops the AI Index—But Google’s New Frontier Model Isn’t Widely Available Yet

Google’s Gemini 4 Carbon has taken the lead on Artificial Analysis’s Intelligence Index—but a benchmark debut is not the same as a public launch. Announced on October 28, 2026, Carbon is Google’s newest frontier model, focused on sustained software engineering and cybersecurity defense. As checked on October 30, its “high” configuration leads the independent index with 59, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57 in their highest-scoring listed configurations.

The important qualification: Google is initially giving access to a subset of cyber defenders, with additional trusted partners to follow through its Fairwind Program. Wider availability for developers, businesses and consumers is still described as coming soon. Carbon is real and independently benchmarked; it is not yet a generally available Gemini upgrade.

What exactly is Gemini 4 Carbon?

“Gemini 4 Carbon” is the name used in Google’s official announcement and on its model page—not an unconfirmed codename inferred from an anonymous leaderboard entry. It follows Gemini 4 Argon, which Google says launched two weeks earlier. The company positions Carbon as its strongest model, with improvements in long-horizon tasks, coding, scientific work and multimodal understanding.

That distinction matters because the two Gemini 4 variants are not interchangeable. Argon has its own availability and benchmark results. A subscription or API route that provides Argon should not be assumed to provide Carbon, and Argon customer testimonials should not be treated as Carbon evaluations.

Which leaderboard is Carbon topping?

The verified lead is on the Artificial Analysis Intelligence Index, version 5.0. The table below records displayed scores checked on October 30, 2026. These are index points, not percentages of human intelligence or task success.

This is a selected comparison, not a consecutive ranking: the full leaderboard includes multiple configurations of the same model. For example, Claude Opus 5.5 at “xhigh with fallback” also displays 57. Reasoning settings and fallback labels are part of the tested entries and should not be silently dropped when comparing results.

Artificial Analysis’s methodology gives equal 25% weight to four categories: knowledge work, coding, science and general capabilities. Its nine evaluations include Terminal-Bench 4.0, FrontierCode 2.0, AA-Briefcase v1.1 and HLE-Diamond. The suite is primarily text-based and English-language, although some evaluations include images. Speech, multilingual performance and broader visual reasoning are assessed separately.

In other words, Carbon leads a meaningful composite test; that does not establish that it is best at every task or modality. A one-point lead over Opus is also not, by itself, proof of a statistically decisive advantage on a particular customer workload. Scores and ordering can change as evaluations and model configurations change.

Its strongest pitch: staying with hard engineering problems

Google says its internal teams have seen Carbon run for days on complex engineering work, persisting through dead ends while investigating bugs across large codebases and low-level systems. That is a claim about sustained agentic work: planning, using tools, interpreting failures and continuing toward an outcome, rather than producing a single convincing response.

For developers, the potential benefit is less supervision across a difficult debugging session. But “ran for days” is not a guarantee of successful autonomous delivery. The announcement does not establish a universal completion rate, and production teams still need tests, review and clear permission boundaries.

Google’s own comparative table also shows why “best coding model” would be too broad a label. On Terminal-Bench 4.0, it lists Carbon at 65.2%, above GPT-6 Astra at 59.1% and Claude Opus 5.5 at 59.6%. On FrontierSWE 2, however, Carbon’s 61.2% trails Astra’s 65.5% and Opus’s 62.3%. These figures come from Google’s published comparison, not a separate replication performed for this article; they should not be confused with the independent index scores above.

Cybersecurity is the headline capability—and the rollout constraint

Google frames Carbon around the defensive lifecycle: finding vulnerabilities, validating whether they can be exploited and patching their root causes. Its announcement reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench.

Crucially, Google says those external-benchmark results were evaluated without guardrails. They describe capability under those evaluation conditions, not a promise that the deployed product will perform every request. Google also reports internal gains in secure-code generation and patch quality, but those vendor-run results do not provide a universal comparison across customers and environments.

The same capabilities that help defenders inspect and repair software can be misused. Google’s restricted initial rollout, participation in the U.S. government’s voluntary pre-release testing process, and phased Fairwind access are therefore central to the story—not a footnote beneath the benchmark scores.

Long context and multimodal work: promising, not a clean sweep

Artificial Analysis lists a one-million-token context window for Carbon. A large window can help an agent work across documents and code, but capacity alone does not prove it will reliably retrieve or reason over every relevant detail.

Google’s comparison offers a more specific signal: Carbon scores 96.2% on GraphWalks at 256,000 to one million tokens, versus 71.8% for GPT-6 Astra and 66.8% for Claude Opus 5.5. Google also reports Carbon ahead on Chartography, at 72.5%, compared with 71.0% and 66.3%, respectively. Those are benchmark-specific vendor-reported results, not evidence of universal superiority in document analysis or visual understanding.

There are clear counterexamples. On the offline subset of OSWorld-2.1, Google’s table gives Opus 87.2%, ahead of Carbon’s 81.5% and Astra’s 80.9%. Selecting a model for computer interaction can therefore lead to a different answer from selecting the highest composite-index scorer.

Reliability claims deserve the same scrutiny as capability claims

Google says Carbon reduces unsupported claims by approximately twofold versus Argon in internal tests, and reports a 0.0% attack success rate on the evaluated Gray Swan indirect-prompt-injection benchmark. It also reports fewer high-severity misalignment alerts in internal coding environments.

These are encouraging reported results, but a zero on a finite benchmark is not immunity to future attacks, and fewer unsupported claims is not the elimination of hallucinations. Buyers should evaluate factual grounding, tool permissions and failure recovery on their own workflows alongside raw benchmark performance.

Price, access and the practical comparison

Google says Carbon will launch at the same introductory token prices as Argon: $2 per million input tokens and $10 per million output tokens, with cached input priced at a 95% discount to the input rate. This is announced pricing, not evidence that a public Carbon API is currently available. Artificial Analysis’s Carbon page explicitly states that there is no publicly available API.

Token rates also do not determine the total cost of an agentic job. Reasoning volume, retries, tool execution and elapsed time all matter. A model that wins a capability benchmark may still be the wrong fit for a fast, low-cost production workflow.

Google additionally reports Carbon at fourth place with 1,768 Elo on WebDev Arena at announcement time. That is a separate category from the Intelligence Index. We could not independently retrieve the live Arena leaderboard during reporting, so this should be read as Google’s dated claim—not a verified current Arena rank.

The bottom line

Gemini 4 Carbon has a substantiated claim to the top of Artificial Analysis’s current Intelligence Index, with a particularly strong pitch around long-running engineering work and cybersecurity. Yet its closest rivals remain competitive, and Google’s own data shows tasks where Astra or Opus performs better.

The useful conclusion is narrower—and more consequential—than “Google wins AI.” Carbon raises the measured frontier while deployment remains limited. The next test is whether that benchmark lead translates into dependable, economical work when broader access arrives.

Sources

• Google: Gemini 4 Carbon announcement, October 28, 2026
• Artificial Analysis: model leaderboard
• Artificial Analysis: Gemini 4 Carbon (high) profile
• Artificial Analysis: Intelligence Index v5.0 methodology
• Google DeepMind: Gemini model comparison

Leaderboard snapshot checked October 30, 2026. Independent index results and Google-reported benchmark claims are distinguished throughout.