Google’s Gemini 4 Carbon Takes the Top Spot on a Key AI Index, but Only Security Teams Can Use It Yet

Google has announced Gemini 4 Carbon, a new frontier AI model built around cybersecurity work and long-running software engineering tasks. The company unveiled it on October 28, 2026, and calls it its strongest model so far. Independent evaluator Artificial Analysis places Carbon first on its newly revised Intelligence Index, but access is tightly restricted. For now, only a subset of vetted cyber defenders can use it.

What Google announced

Carbon builds on Gemini 4 Argon, which Google introduced on September 30 through its Fairwind Program for cyber defenders and opened more broadly on October 13. Google says Carbon is designed for harder, multi-step engineering problems. Its internal teams have watched the model work for days, pushing through dead ends to trace and patch subtle bugs in large codebases and low-level systems.

The rollout is deliberately narrow. Carbon is available first to a subset of cyber defenders, and Google plans to extend access gradually to more trusted partners through Fairwind. The company says it is still gathering feedback and refining safeguards before offering the model to developers, businesses and consumers. It also says it is taking part in the US government’s voluntary process for pre-release model testing. Google has not given a date for wider availability.

The benchmark lead in context

On version 5.0 of the Artificial Analysis Intelligence Index, Carbon scores 59. That is one point ahead of Anthropic’s Claude Opus 5.5 at 58 and two ahead of OpenAI’s GPT-6 Astra at 57. Google’s comparison table lists Argon at 50 on the same index, which puts Carbon nine points above its predecessor.

The lead over rivals is narrow. Carbon ranked first in four of the nine evaluations that make up the index, and several of those margins were small. Artificial Analysis also ranks Carbon second, behind Claude Opus 5.5, on its separate Coding Agent Index.

Google’s own comparison covers 17 benchmarks and four rival models: Argon, GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1. Carbon posted the highest score on 15 of them. Astra and Opus 5.5 both beat it on FrontierSWE 2, and Opus 5.5 led on the OSWorld-2.1 offline subset, 87.2% to 81.5%. The comparison is limited to the models Google chose. On Terminal-Bench 4.0, for example, Artificial Analysis’ own rankings show Kimi K3.1 Pro ahead of Carbon, 67.2% to 65.2%.

The biggest gaps in Google’s table come on long-context reasoning and exploitation testing. On GraphWalks, Carbon scored 96.2%, against 71.8% for Astra and 66.8% for Opus 5.5, margins of roughly 24 and 29 points. On ExploitGym, it scored 54.8%, against 42.4% for Astra and 34.5% for Opus 5.5, roughly 12 and 20 points. These figures were published by Google and have not been independently reproduced. Google also says it ran its cybersecurity benchmarks with guardrails switched off.

A security-first model

Google reports that Carbon scored 88.2% on SEC-bench Pro, which measures vulnerability discovery, 54.8% on ExploitGym, which tests whether a flaw can actually be exploited, and 71.0% on CWE-bench, which measures patching. Together they cover the defensive cycle from finding a weakness to confirming its risk to fixing it. On SEC-bench Pro and CWE-bench, the lead over the closest rival is around three points.

Google adds that Carbon more than doubled Argon’s success rate at generating secure code on internal tests. It also says the Chrome team found Carbon performed more than ten times better than another lab’s large frontier model at producing high-quality patches. Google does not name that model or explain the metric, so the claim cannot be checked from outside. Security firm Wiz, a Google Cloud security partner, reports that Carbon led on its web penetration-testing benchmarks.

Pricing: cheaper for now

Carbon launches at an introductory price of $2 per million input tokens and $10 per million output tokens, the same as Argon, with cached input tokens discounted by 95%. A footnote in Google’s announcement says $4 and $20 will apply once the introductory period ends, but no end date has been announced.

Reported list prices for Claude Opus 5.5 are $4 and $20, and for GPT-6 Astra $10 and $50. On those figures, Carbon’s introductory rate is half of Opus 5.5’s and one-fifth of Astra’s, but it matches Opus 5.5 once the discount expires. Per-token price is not the whole picture. Artificial Analysis estimates the cost of running its Intelligence Index at $7.76 per task for Carbon at standard prices, against $10.04 for Opus 5.5 and $6.25 for Astra, because models differ in how many tokens they use. Since Carbon has no public API yet, these numbers are a preview rather than a price developers can pay today.

Reliability and safety claims

Google says Carbon cuts unsupported claims by about half compared with Argon. It also says high-severity misalignment alerts fell by almost 80% in live coding environments with its own teams, and that Carbon resisted every attack in Gray Swan’s indirect prompt-injection benchmark, down from a 0.7% success rate for Argon. These are Google’s internal results.

One independent data point cuts the other way. Artificial Analysis reports that Carbon’s hallucination rate on its AA-Omniscience test rose to 11% from Argon’s 4%, even as accuracy improved. That rate is still far below the 50% reported for both Opus 5.5 and GPT-6 Astra.

What to watch next

Three things will shape how Carbon is judged. The first is when Google opens an API. Argon took 13 days to go from limited release to broad availability, but nothing says Carbon will follow the same schedule. The second is when the introductory pricing ends. The third is whether independent testers can reproduce Google’s large margins on long-context and exploitation tasks. Until then, what is confirmed is a one-point lead on a widely followed index and strong self-reported security results, in a model that few people outside Fairwind can test.

Sources

• Google: Gemini 4 Carbon: our next frontier model
• Artificial Analysis: Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5
• XenoSpectrum: Google’s Gemini 4 Carbon Tops One AI Index by Just 1 Point, With Big Leads Only in Select Tests