Gemini 4 Carbon tops an independent index, but only vetted cyber defenders can use it
What happened
On 28 October 2026 Google announced Gemini 4 Carbon in a blog post by Koray Kavukcuoglu, chief AI architect at Google and a senior vice president at Google DeepMind. It arrived two weeks after Gemini 4 Argon, which remains Google’s publicly available model. Google describes Carbon as its strongest model, emphasising long-running agentic coding and cybersecurity work.
Unusually for a flagship model, Carbon was not released to the public. Google said access would start with “a subset of cyber defenders” and expand to other trusted partners through what it calls the Fairwind Program. The AI news site Nowrap reported that the programme covers vetted governments, national cyber agencies, critical-infrastructure operators and large technology platforms, with background checks. Google said it would make Carbon available to developers, businesses and consumers “as soon as possible”, without a date, and that it is taking part in the US government’s voluntary pre-release model testing. It listed introductory pricing of $2 per million input tokenstoken A small chunk of text, often a word or part of a word, that a language model reads or writes as one unit. and $10 per million output tokens, rising later to $4 and $20.
What is claimed vs. what is verified
Most of the headline figures are Google’s own. The company says Carbon scored 88.2 percent on SEC-bench Pro, a test of finding software vulnerabilities, 54.8 percent on ExploitGym, which tests exploiting them, and 71.0 percent on CWE-bench for patching. Google notes these external cyber benchmarksbenchmark A fixed set of tasks used to score and compare AI systems. were run without guardrails. Nowrap reports that Google ran SEC-bench Pro and ExploitGym itself in its own Antigravity harnessevaluation harness The software setup that feeds a test to an AI model and scores its answers, whose settings can change the results., and points out that harness and settings choices can move agentic scores by many points. Those numbers are claims until someone outside Google reproduces them.
The safety figures are internal too. Google says Carbon makes roughly half as many unsupported claims as Argon, and that high-severity misalignment alerts in live coding environments fell by almost 80 percent. It also cites a claim from its Chrome team that Carbon was more than ten times better at producing high-quality patches than an unnamed rival model. VentureBeat’s Carl Franzen noted that Google gave no uncertainty estimates and no technical report, and that the unnamed rival makes the Chrome comparison impossible to check. The Google post links only a safety report that is presented in the context of Argon.
Some independent evidence does exist. Artificial Analysis, a third-party benchmarking group, ranks Carbon first on its Intelligence Index v5.0 with 59, ahead of Claude Opus 5.5 at 58, GPT-6 Astra at 57 and Argon at 50, while noting that Carbon has no public API. On the same group’s other measures, as reported by VentureBeat, the picture is mixed: Carbon scored 59 on a coding-agent index where Claude Code with Opus 5.5 scored 62, and it ranked below several rivals on an Elo-style comparison called AA-Briefcase. On AA-Omniscience it answered more questions correctly than Argon, but its hallucination ratehallucination rate How often an AI model gives a confident answer that is false, measured in a specific way by each test. on that test rose from 4 to 11 percent. Nowrap reports a separate third-party index, Vals, at 70.4 percent for Carbon against 70.2 for Opus 5.5.
Why it matters
Carbon is a test of a release strategy as much as a model. Google is treating strong vulnerability-finding and exploitation skills as dual-usedual-use Useful both for protection and for harm, such as a tool that can find software flaws to fix or to attack., giving defenders a head start before wider access. If that becomes common, the most capable systems may increasingly be available first to governments and large institutions rather than to the public or independent researchers.
The release also shows how thin independent checking can be at launch. With no public API, outside evaluators can test only through whatever access they are granted, and the cybersecurity claims, which motivate the restricted release, are among the least verified.
Open questions and other views
Google’s position is that staged access lets defenders patch software before attackers gain similar tools, and that it will widen access once guardrails are refined. The case for caution rests partly on the jump in exploitation scores, from 42.0 to 54.8 percent on ExploitGym by Google’s figures, as VentureBeat reported.
The reporting raises other points. A one-point lead on a composite index is a narrow margin, and Carbon trailed rivals on some coding and computer-use tests in Google’s own comparison table. Developers and smaller security firms outside the programme cannot verify the claims or benefit from the model, which concentrates both capability and evaluation in a few hands. Without a model card or technical report for Carbon, which VentureBeat noted is missing, outside researchers have little against which to assess the safety claims.
What remains open: when, or whether, Carbon reaches general release; whether independent groups can reproduce the security scores with guardrails on; and what the US pre-release testing found, which Google has not described.
Sources
- Google, “Gemini 4 Carbon: our next frontier model”, Koray Kavukcuoglu, 28 October 2026. https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/
- VentureBeat, “Google unveils Gemini 4 Carbon, retaking the world’s no. 1 spot on updated, harder benchmarks, but access is limited”, Carl Franzen, 28 October 2026. https://venturebeat.com/technology/google-unveils-gemini-4-carbon-retaking-the-worlds-no-1-spot-on-updated-harder-benchmarks-but-access-is-limited
- Nowrap, “Google’s Gemini 4 Carbon takes No. 1, but access is limited”, 29 October 2026, updated 3 November 2026. https://nowrap.ai/news/google-gemini-4-carbon-launch
- 9to5Google, “Google announces Gemini 4 Carbon with ‘significant’ coding gains”, Abner Li, 28 October 2026. https://9to5google.com/2026/10/28/gemini-4-carbon-announcement/