Google DeepMind announces Gemini 4 Carbon with cybersecurity focus

Google DeepMind has formally introduced Gemini 4 Carbon. The company calls it its next frontier model and its strongest to date.

Google is pitching the model on two fronts: state-of-the-art cybersecurity and long-horizon capabilities.

What Google is claiming

The announcement centers on cybersecurity performance and on tasks that unfold over long stretches of work. Google cited results on the Artificial Analysis Intelligence Index, a third-party benchmark that aggregates model performance. It also cited a separate evaluation focused on patching.

“Long horizon” refers to multi-step work that plays out over many actions rather than a single prompt.

Carbon’s arrival follows a stretch of quiet testing. Internal testing of Carbon reportedly began in early October 2026 on Google’s Jetski platform. That work compared its coding performance against Anthropic’s Claude Opus 5.5.

Around October 21, 2026, reports surfaced of an external early access program. Early feedback, largely anecdotal, praised Carbon’s 3D scene generation. The same reports flagged a weakness: output speed.

The Gemini 4 family so far

Carbon is not the first Gemini 4 model to lean into security. Gemini 4 Argon was publicly announced on September 30, 2026. It launched through the Fairwind Program, an initiative aimed at cybersecurity partners.

Argon supports output of up to 1 million tokens in tasks. Argon also tied on multiple coding benchmarks. It posted a maximum score of 77.9% for long-horizon engineering on DeepSWE v1.1, an evaluation of how well a model handles extended software engineering work.

Earlier reporting described Carbon as a variant of Gemini 4 geared toward coding. It was framed as Google’s answer to Anthropic’s Claude Opus 5.5. Google’s own framing now puts cybersecurity and long-horizon work front and center.

What this means

The early access feedback on Carbon’s slow output is worth watching closely. Carbon’s pricing has not been disclosed, and neither has a clear picture of how widely it will be available or through which channels.

Google cited third-party benchmarks for its security claims. Google has not provided any launch statements or timelines for Gemini 4 Carbon; its eventual public name and pricing remain unknown.

Argon was announced on September 30, 2026. Testing of Carbon reportedly started within days. Google’s rollout strategy for Argon hints at how it may approach Carbon: routing Argon first through a partner program for cybersecurity firms suggests a controlled release. How Google gates access to Carbon, and whether it follows the partner-first approach used for Argon, will say a lot about how seriously the industry takes that risk as these tools grow more capable.