Fast Technology News, November 13 — Google today officially launched its new AI model, Gemini 4 Hydrogen, which focuses on efficient coding and AI agent tasks, delivering significant upgrades in running speed, coding capabilities, and reliability.

According to Google, compared with the previous-generation Gemini 3.8 Flash, Gemini 4 Hydrogen achieves 2.5x the task completion speed of its predecessor in the default High thinking mode, while coding test scores have improved by more than 3x.

Specifically, in the Terminal-Bench 4.1 coding benchmark test, Gemini 4 Hydrogen scored 57%, whereas Gemini 3.8 Flash scored only 18.2% — a remarkably substantial improvement.

Meanwhile, the new model generates fabricated content more than 5x less often and can complete tasks more reliably even when encountering tool call failures, empty search results, and other issues.

In terms of coding capabilities, Gemini 4 Hydrogen can handle tasks such as code writing, program debugging, and vulnerability fixing, and can make targeted modifications to specific issues in a codebase without affecting unrelated code.

Google has also optimized the model for large codebases, supporting parallel file reading and code searching to further boost development efficiency for complex projects.

According to early tests, in complex charts, SQL, and data-intensive enterprise tasks, the new model's running speed can reach 2 to 2.8x that of the previous generation, while reducing token consumption by 12% to 23%.

Notably, although Gemini 4 Hydrogen is positioned as an efficient execution model, it has already surpassed Google's own high-end models in some tests.

In the Artificial Analysis Intelligence Index test, its default High mode score reached 51 points, exceeding Gemini 4 Argon's 50 points, and when the highest xhigh thinking mode is enabled, the score can be further improved to 54 points.

Gemini 4 Hydrogen supports a context window of up to 1 million tokens and can handle various content types, including text, images, audio, video, and complex documents.

Currently, Gemini 4 Hydrogen has begun rolling out to global users, covering platforms including the Gemini app, Google Search, Google AI Studio, and the Gemini API, with free users also having the opportunity to try it out.

Among them, free users of the Gemini app can access the model through an automatic model selection mechanism, while developers can call it through the Gemini API, with an input price of $1.5 per million tokens and an output price of $7.5 per million tokens.