Google DeepMind announces Gemini 4 Carbon, its strongest model

Google DeepMind has announced Gemini 4 Carbon, which it calls its strongest model, with state-of-the-art cybersecurity and long-horizon capabilities. It follows Gemini 4 Argon, launched two weeks earlier. DeepMind says the pace shows how it is pushing the frontier, with its researchers and models "working hand in hand" to advance capabilities while strengthening safeguards in tandem.

On headline numbers, DeepMind reports that Carbon ranks #1 on the recently updated v5.0 of the Artificial Analysis Intelligence Index, at 59. On cybersecurity it reports a "firm lead" on three external benchmarks, evaluated without guardrails: SEC-bench Pro for finding vulnerabilities (88.2%), ExploitGym for exploitation (54.8%) and Collinear's CWE-bench for patching (71.0%). In web development, Carbon gains 89 elo points to reach #4 (1768) on WebDev Arena. DeepMind also says it improved coding, agentic and scientific tasks, full-stack web development and multimodal understanding.

Rollout is deliberate. Google says it is engaged in the U.S. government's voluntary process for pre-release model testing. Starting this week, Carbon goes to a subset of cyber defenders, and access will be phased in gradually to more trusted partners through the Fairwind Program. Google is still gathering feedback and iterating on guardrails before making Carbon available to developers, businesses and consumers through its product surfaces "as soon as possible". No date is given. Carbon will launch at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off the input price.

For engineering work, DeepMind says Carbon is built for sustained agentic workflows. Its internal teams have seen it run for days, persisting through dead ends on multi-step tasks such as root-causing and patching subtle bugs across large codebases and low-level systems.

In security, the post says Carbon covers the defense lifecycle of detection, validation and remediation. On internal benchmarks it shows more than 2x uplift in success rate for generating secure code over Argon. The Chrome team reportedly found Carbon performed more than 10X better than another lab's large frontier model (unnamed) at generating high quality patches. Wiz, which Google describes as its leading cloud security platform, observes Carbon leading on web penetration testing benchmarks, a core capability for Wiz's Scan for Good initiative.

The post then turns to safety. Carbon builds on the same multi-layered stack as Argon: model training, real-time misuse defenses, prompt-injection resilience and chain-of-thought monitoring. DeepMind says it trained Carbon to go a step further on alignment, and lists results. Successful cheating on Terminal-Bench 4.0 fell to 0.0%, with attempts at 0.4%, while overall accuracy improved. In live coding environments with internal teams, high-severity misalignment alerts dropped by almost 80%. In internal testing, unsupported claims are cut by approximately 2x over Argon; the model puts methodological caveats into the deliverable or withholds blocked actions with a list of what is still required, rather than fabricating. When facing ambiguity, it retrieves sources and surfaces discrepancies to the user before acting, and the post says users "should observe" a much better collaborative experience. Overreach fell 26% and under-service 29%. On Gray Swan's Indirect Prompt Injection benchmark, where Argon had a 0.7% attack success rate, Carbon scores 0.0%.

DeepMind closes by saying it wants to deliver leaps in software engineering and cyber defense without compromising safety or reliability, and that it is eager to see how the first cohort of cyber defenders uses the model.

Key facts

Gemini 4 Carbon ranks #1 on Artificial Analysis Intelligence Index v5.0 at 59, per Google DeepMind's own post.
Cybersecurity results, evaluated without guardrails: SEC-bench Pro 88.2% (finding vulnerabilities), ExploitGym 54.8% (exploitation), CWE-bench 71.0% (patching).
Access starts this week for a subset of cyber defenders via the Fairwind Program; developers, businesses and consumers come "as soon as possible", with no date.
Introductory price matches Argon: $2 per million input tokens, $10 per million output tokens, cached input tokens 95% off the input price.
Safety claims include 0.0% successful cheating on Terminal-Bench 4.0, almost 80% fewer high-severity misalignment alerts internally, and 0.0% attack success on Gray Swan's prompt injection benchmark.

Why it matters

Carbon arrives only two weeks after Gemini 4 Argon, and Google presents it as a step up in both agentic engineering and cyber defense. The claimed lead is concrete: #1 on the Artificial Analysis Intelligence Index v5.0 at 59, and top results on benchmarks covering finding, exploiting and patching vulnerabilities. Google also positions Carbon for long-running work, with internal teams seeing it run for days on complex engineering tasks. The restricted first release, aimed at cyber defenders, shows a model launched first to defensive security users.

Who it affects

First in line are the cyber defenders and trusted partners getting access through the Fairwind Program. Developers, businesses and consumers are next, through Google's product surfaces, but no date is given. Security teams that triage vulnerabilities, validate exploitability and write patches are the audience the post addresses most directly. Web developers may care about the WebDev Arena gain to #4 (1768, up 89 elo points).

How to use it

There is no general access yet. Carbon is available this week only to a subset of cyber defenders, with more trusted partners added gradually through the Fairwind Program. Google says it will launch at an introductory price of $2 per million input tokens and $10 per million output tokens, the same as Argon, with cached input tokens at 95% off the input price. Everyone else has to wait for the wider release that Google promises "as soon as possible".

How solid is it

This is a first-party announcement. The Artificial Analysis, WebDev Arena and external cybersecurity figures are reported by Google in its own post, and no independent verification is described. The internal results (secure code uplift, misalignment alerts, unsupported claims, overreach and under-service) come from Google's internal benchmarks and teams. The Chrome team's "more than 10X" comparison names neither the other lab nor its model. Competitor scores on SEC-bench Pro, ExploitGym or CWE-bench are not given, so the size of the lead cannot be judged from the post.

Risks and caveats

The cybersecurity benchmark numbers were evaluated without guardrails, so they describe the raw model rather than what restricted users will get. Google says it is still iterating on guardrails and gathering feedback before broader release, which is why access is limited. The post says the "much better collaborative experience" is something users "should observe", not a measured result. Carbon builds on the same safety stack as Argon, and Google says it is taking part in the U.S. government's voluntary pre-release testing process. The post does not say what Fairwind Program membership requires or how many partners have access.

“our strongest model with state-of-the-art cybersecurity and long horizon capabilities”

— Google DeepMind, describing Gemini 4 Carbon