Google DeepMind introduces Gemini 4 Carbon, aimed first at cyber defenders
Google DeepMind's Gemini 4 Carbon targets cybersecurity and long-running engineering work, with access starting for a subset of cyber defenders.
Google DeepMind has introduced Gemini 4 Carbon, which it calls its strongest model. Rather than opening it to developers, businesses or consumers, the company is starting with a subset of cyber defenders. In a post dated October 28, 2026, Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google, described Carbon as our strongest model with state-of-the-art cybersecurity and long horizon capabilities. Google says access starts this week for that subset of cyber defenders and will be phased to additional trusted partners through the Fairwind Program. The company also says it is taking part in the U.S. government's voluntary process for pre-release model testing.
Google reports 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench, which the post credits to Collinear, for patching. All three were run without guardrails, meaning the scores reflect what the model can do when its safety filters are not in the way. The three tests map onto a defense pipeline: find a flaw, check whether it can be exploited so the most dangerous ones get fixed first, then write a patch that addresses the root cause. Google's argument is that one model can handle all three steps. Google also cites internal results from the Chrome team, which it says found Carbon performed more than 10X better than another lab's large frontier model when generating high quality patches. Wiz says Carbon leads on web penetration testing benchmarks. The post doesn't name the other lab or give the underlying patch counts.
Google says internal teams have seen Carbon run for days, persisting through dead ends on tasks such as root-causing and patching subtle bugs across large codebases. These are internal observations. The post gives no task list, success rate or number of runs behind them.
Independent comparisons are less uniform than the announcement. Artificial Analysis ranks Carbon first on its Intelligence Index v5 with a score of 59, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57. The firm's own write-up is more qualified. Carbon has the top score in four of the nine Index evaluations, including HLE-Diamond and AA-Omniscience, but it is second to Claude Opus 5.5 on the Coding Agent Index, scoring 59 against 62 for Claude Code with Opus 5.5 at max effort. Cognition's FrontierCode 2.0 puts Carbon at 40%, behind Opus 5.5, GPT-6 Astra and Claude Sonnet 5.5. Artificial Analysis's 51% on FrontierCode v1.1 is a different version, so the two results should not be read against each other.
Google's alignment claims carry several qualifications. The post says Carbon reduced successful cheating on Terminal-Bench 4.0 to 0.0%, but attempts to cheat fell only to 0.4%. That is a different measure from a model that never tries. The nearly 80% drop in high-severity misalignment alerts comes from live coding environments run by Google's own teams and has not been independently verified. On the Gray Swan indirect prompt injection benchmark, Argon had a 0.7% attack success rate and Carbon reports 0.0%. The post describes this as resisting every evaluated attack attempt, a claim bounded by the attacks that were tested.
Artificial Analysis's hallucination results show a tradeoff. Carbon scores 54 on its AA-Omniscience v1.1, six points ahead of Claude Opus 5.5 at max, and its accuracy rose 10 points over Argon to 58%. Its hallucination rate, though, rose from 4% to 11%. The firm notes that this is still the second lowest of any model, behind only Argon. By comparison, Claude Opus 5.5 at max and GPT-6 Astra at max are both at 50%.
On cost, Artificial Analysis measures Carbon at $7.76 per Intelligence Index task at standard pricing. That is about 23% less than Claude Opus 5.5 at max effort at $10.04, but about 1.2 times the $6.25 for GPT-6 Astra at max. Carbon uses 123k output tokens per task, 29% fewer than Opus 5.5 at max. The model has a 1M-token context window and accepts text, image, video and speech input. Google's introductory price is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after the introductory period, with cached input tokens at 95% off the input price. Artificial Analysis says Google is offering a 50% launch discount with no confirmed end date, and that its charts use standard pricing. Carbon is not yet publicly available. On the WebDev Arena it reached #4 with a score of 1768, up 89 Elo points.
What changed
Google DeepMind introduces Gemini 4 Carbon, initially available to a subset of cyber defenders and to be phased to additional trusted partners through the Fairwind Program.
Type: New capability · Applies to: Gemini 4 Carbon
Why it matters
Companies running security scanning, vulnerability patching or long autonomous coding work now have another frontier option, though the benchmark claims need confirmation from Google's own figures and third-party scores. The published per-task cost and the post-introductory price rise will matter for budgeting.
What you can do
Security teams should check whether they qualify for the Fairwind Program early access and test Gemini 4 Carbon on their own vulnerability-finding and patching workloads before the introductory pricing ends.
How the story developed
Adds third-party Artificial Analysis and Cognition benchmark results, including first place on the AA progress index and 40% on FrontierCode 2.0.
Latent Space reports leaderboard placements in Text Arena, Agent Arena and WebDev, plus Perplexity's WANDR state-of-the-art claim and a Net Score lead in an independent index.
Original source
Further reporting
Where we saw it
DAIR.AI (Elvis Saravia), Forward Future (Matthew Berman), Latent Space
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works