Google DeepMind unveils Gemini 4 Carbon for cybersecurity and long-horizon coding

An October 28 frontier model on the Gemini line

Google and DeepMind announced Gemini 4 Carbon on October 28, 2026, about 01:02 on October 29 in Asia/Singapore. Google frames Carbon as the strongest Gemini model yet, centered on end-to-end cybersecurity and long-horizon autonomous engineering. This report was published on October 29 after checking the Google blog and DeepMind’s announcement posts.

Carbon follows Gemini 4 Argon by about two weeks. Google describes the release as part of a cadence that advances capability and safeguards together. Carbon sits on the Gemini frontier language and agent line, a different product family from Genie, Google’s world-model series.

Index leadership, with access still limited

On the Artificial Analysis Intelligence Index v5.0, Google reports a score of 59 and a current first-place rank. With safeguard evaluations disabled, it reports SEC-bench Pro at 88.2%, ExploitGym at 54.8% and CWE-bench at 71.0%. Google also says Carbon can sustain multi-day progress on complex engineering tasks.

Introductory prices match Argon’s opening tier: $2 per million input tokens and $10 per million output tokens, with cached input about 95% off. After that introductory period, the listed rates are $4 and $20. Google says Carbon went through U.S. voluntary pre-deployment testing, and that the rollout runs from a small defender cohort to feedback and harder safeguards, then a wider release.

Two weeks after Argon, three sharper bets

The official story is a change in emphasis, as well as a change in score. In Google’s internal evaluations, Carbon can run for multiple days, work through dead ends, and handle root-cause work and patches across large codebases and lower-level systems.

A second emphasis is a full defensive cybersecurity loop: find vulnerabilities, validate exploitability, then fix root causes. A third is alignment against shortcuts, with fewer cheat or shortcut behaviors alongside higher coding quality. DeepMind’s posts say Carbon can work through complex tasks over multiple days without getting stuck or taking shortcuts, and can ship high-quality, reliable code. Early defender feedback is intended to shape further safeguard work before broader access.

Three external cybersecurity scores, as Google published them

Google presents the three external scores as results with safeguard evaluations disabled, restated from the official blog. SEC-bench Pro at 88.2% is described as autonomously finding complex defects in real software. ExploitGym at 54.8% is framed as helping defenders prioritize which issues to fix first. CWE-bench, from Collinear, at 71.0% is described as turning verified findings into working code fixes.

Google and partner notes sit beside those scores and are separate from a public leaderboard anyone can rerun from this page. Versus Argon, Google says internal secure-code-generation success improved by more than 2×. Feedback attributed to the Chrome team says that, when producing high-quality patches, Carbon outperformed another lab’s large frontier model by more than 10×. Wiz, a cloud security company, says Carbon leads on web penetration-testing benchmarks tied to its Scan for Good program. These are vendor and partner claims. This article reports the framing and the published numbers. It does not include attack steps, and my ai daily has not independently reproduced an exploit or rerun the tests.

Long-horizon coding beyond a short answer

Carbon is aimed at longer agentic workflows. Google also cites gains on coding, agents, scientific tasks, full-stack web development and multimodal understanding. On WebDev Arena, it reports an Elo increase of 89 points, to fourth place at 1768.

Taken with the blog and DeepMind’s posts, the pitch is that the model holds up across multi-day work, takes fewer shortcuts and produces more reliable code and artifacts. That framing matters more for multi-repository debugging, long-cycle patching and full-stack delivery than for a one-shot short answer. The comparisons are Google’s.

The same launch tightens measured alignment

Carbon keeps the multi-layer safety stack described for Argon, including training, real-time abuse defenses, prompt-injection resistance and chain-of-thought monitoring. Google says alignment is tighter on the named tests. On Terminal-Bench 4.0, successful cheating is reported at 0.0% and cheat attempts at 0.4%, with accuracy up at the same time. Internal high-severity misalignment alerts are down about 80%, and unsupported assertions are down about 2× versus Argon.

In the same Google framing, over-refusal is down 26% and under-serving is down 29%. On Gray Swan’s indirect prompt-injection evaluation, reported attack success moves from about 0.7% on Argon to 0.0% on Carbon. When a request is ambiguous, Google says Carbon is more likely to retrieve sources and surface conflicts before acting. These figures describe the named evaluations.

Who can use it this week

The Fairwind Program starts this week with a subset of cyber defenders and is due to expand to more trusted partners. Developer, enterprise and consumer surfaces are described as arriving as soon as possible. Google has not given a full launch date. The price schedule remains Argon’s introductory rates: $2 and $10 per million input and output tokens, cached input about 95% off, then $4 and $20.

Readers outside that first defender set can read the benchmarks now. Calling the model from a general API or consumer product still waits on a broader product announcement.

Editorial context: a controlled pilot beside a public score

The launch pairs a first-place published index score with a narrow rollout. An index rank, a cybersecurity benchmark and a partner comparison answer different questions. A defender still has to see how the model behaves on the systems they actually operate.

The cybersecurity figures carry conditions: safeguard evaluations were disabled, and the Chrome and Wiz comparisons are partner statements. my ai daily is reporting those claims and is not publishing a reproduced exploit. A useful follow-up is whether defender feedback changes the safeguards before wider access, and whether independent evaluations match the published scores once more teams can test the model.

For security-engineering agents and long-cycle autonomous coding, this is the Google frontier update to follow while Fairwind is the available door. Anyone who needs Carbon in an API or a consumer product is still waiting on the wider release.