United States • • • 🌿 Progressive

Centro de operaciones de ciberseguridad con analistas ante pantallas con escudos digitales y un núcleo luminoso de inteligencia artificial
Imagen generada para EPM

Google unveils Gemini 4 Carbon with a focus on defensive cybersecurity

Google presented Gemini 4 Carbon as its new frontier model, emphasising defensive cybersecurity, performance in science, agentic and coding tasks, and a staged…

--- The context --- Google has presented Gemini 4 Carbon as its new frontier artificial intelligence model, with a marked emphasis on defensive cybersecurity. It follows the pattern of Gemini 4 Argon, presented two weeks earlier and available in mid-October. --- The facts --- According to the company, Carbon offers strong performance in science, agentic and coding tasks, as well as multimodal understanding and full-stack web development. Google maintains that it leads version 5.0 of the Artificial Analysis Intelligence Index with a score of 59, and that in vulnerability detection, on SEC-bench Pro, it reaches 88.2%, in exploitation, on ExploitGym, 54.8%, and in code patching, on CWE-bench, 71%. On long context, in the GraphWalks test, the company attributes 96.2% to Carbon, compared with 71.8% for GPT-6 Astra and 66.8% for Claude Opus 5.5. On Vending-Bench 2, under those same company results, Carbon totals 16,484 dollars and surpasses Claude Opus 5.5, which remains at 9,235 dollars. Google also says that, in its internal evaluations, the model stands out in secure code generation and reduces alignment errors and hallucination in real development environments. --- The positions --- Google has stressed, in terms attributed to its own announcement, that Carbon illustrates the pace at which the company is pushing the frontier forward, building on what was launched with Argon. The firm has also underlined its prudent approach and its collaboration with the United States Government's voluntary pre-release testing process. --- What remains unknown --- The company has not provided the exact date of the general rollout. It has specified the order of access: this week it will reach a group of cyber defence experts, later it will open to trusted partners through the Fairwind programme, and afterwards it will reach developers, businesses and final consumers. Nor is there independent verification yet of the published figures, which are the company's own results. --- Unanswered questions --- When will Google set the exact date for the general rollout of Gemini 4 Carbon? Will independent evaluations confirm results that today come only from the company itself? Which cyber defence experts make up the first group with access this week? At what point will developers, businesses and final consumers outside the Fairwind programme be able to use it in practice? --- EPM analysis --- The largest gaps against GPT-6 Astra and Claude Opus 5.5 appear, under Google's data, in long context and on Vending-Bench 2. The introductory price matches Argon's: 2 dollars per million input tokens and 10 per million output tokens, with a 95% discount on cached input tokens. 📌 EPM Take: In EPM's view, Gemini 4 Carbon matters less for the headline of a new model than for where it places the emphasis: verifiable cyber defence, secure code and a staged rollout under prior scrutiny. If the company's own figures hold up elsewhere, Google will have moved the competition onto ground where mistakes are costly. Until then, prudence should rule: welcome the advance, demand independent proof, and watch who gains access first and under what effective controls in daily practice.
📤 Share on Telegram

¿Te gustó este artículo? Recibe cobertura global en tu correo.

Suscríbete gratis / Subscribe free