Gemini Carbon

Gemini Carbon is not an all-purpose entertainment assistant for mass consumers; in everyday use in the Gemini app, users will still use lighter and lower-cost models.

Gemini Carbon is designed for continuously running agent workflows, long-horizon software engineering, and cybersecurity defense.

Before publicly releasing Gemini Carbon, Google proactively took part in the relevant model testing process and chose to open it only to a limited group of professional cybersecurity teams and security researchers in the first week.

Gemini Carbon retains a long context window of 1 million tokens and natively supports multimodal processing of text, images, video and speech.

Gemini Carbon is designed to support workloads that run for several days and to autonomously adjust its strategy and try new paths when it encounters an obstacle in its reasoning.

In automated cybersecurity testing, Gemini Carbon can uncover hidden vulnerabilities, analyze their risk levels and attempt to automatically generate fix patches.

Official data shows that Gemini Carbon's fabrication of unfounded information is reduced by about 5% compared with the previous-generation model.

When evidence is insufficient, Gemini Carbon explicitly lists the missing information rather than forcing out an uncertain answer.

According to comprehensive evaluations by organizations such as Artificial Analysis, Gemini Carbon achieved a very impressive ranking on the composite intelligence index, demonstrating top-tier reasoning capability.

Gemini Carbon's output speed exceeds 100 tokens per second, an excellent level among large models of the same size.

Gemini Carbon's API usage price is relatively high, and when handling complex tasks it produces longer output text, further raising the actual cost of use.

Google DeepMind released its new-generation flagship model Gemini Carbon, only two weeks after the launch of the previous-generation model Gemini Argon.