AI News

Google Releases Gemini 4 Hydrogen Model, Focused on Everyday Coding and Agentic Workflows

English text is being prepared.Updated Nov 13 21:47

AI summary

Google has officially launched the Gemini 4 Hydrogen model, aimed at everyday coding and agentic workflows, priced at $1.50 per million input Tokens and $7.50 per million output Tokens, covering a 1 million Token context window with no tiered pricing. The model offers four thinking depth levels: low, medium, high (default), and xhigh. On the Artificial Analysis Intelligence Index, it scores 51 at the default high thinking level and can reach 54 at the xhigh level. Google has also simultaneously launched the Orchestrator intelligent scheduling feature on the Google Antigravity platform, which can automatically switch between models such as Carbon and Hydrogen within the same session; in Terminal-Bench 4.0 tests, average quota consumption was reduced by 24% and completion speed improved by 19%.

Why read this

SelectedAI score 90

The original text provides pricing, four levels of thinking depth, and Terminal-Bench comparison data, which can be used to assess its positioning in everyday coding and agentic workflows.

Topics

Tags

Body · AI translation

Google Releases Gemini 4 Hydrogen Model: Focused on Everyday Coding and Agentic Workflows, $1.50 per Million Tokens of Input

English text is being prepared.

English text is being prepared.

English text is being prepared.

English text is being prepared.

Output: $7.50 per million Tokens (about 50.1 yuan RMB at current exchange rates)

This pricing covers the full 1 million Token context window, with no tiered pricing. On the Artificial Analysis Intelligence Index, Hydrogen scores 51 at the default high thinking level, higher than Gemini 4 Argon's 50, and can reach 54 at the highest xhigh level, while its task completion speed is often better than models of the same class.

Google has also launched the Orchestrator intelligent scheduling feature, built into the Google Antigravity platform, which can automatically switch dynamically between models such as Carbon and Hydrogen within the same session. In the Terminal-Bench 4.0 test, this scheduling approach achieved a task success rate close to Carbon's while reducing average quota consumption by 24% and improving completion speed by 19%.

Compared with the previous generation Gemini 3.8 Flash, Hydrogen at the default high thinking level completes tasks 2.5 times faster, raises its Terminal-Bench 4.1 score from 18.2% to 57%, and reduces the hallucination error rate by more than 5 times.

The model offers four thinking depth levels: low, medium, high (default), and xhigh, allowing flexible adjustment of reasoning depth, latency, and cost according to task requirements; it supports up to 1 million Tokens of native input and 1 million Tokens of output, and can handle multimodal content such as large PDFs, tables, charts, and audio/video.

On safety, Hydrogen reduces unauthorized actions in agentic tasks by up to 81%, reduces under-service issues by 79%, and reduces inaccurate behavior and task shortcut-taking in tool-calling and coding scenarios by 12 times; it scores 96.7% on internal agent honesty evaluations, retains its chain of thought 99.9% of the time, and improves resilience against indirect prompt injection on the Gray Swan IPI benchmark by more than 2.3 times.

Starting on the launch day, Hydrogen is being rolled out globally: developers can call it through the paid tiers of the Gemini API, AI Studio, and Android Studio; enterprises can deploy it through Gemini for Business, Google Antigravity, and Gemini in spreadsheets; ordinary users can select the model from the dropdown in the Gemini app, and free users in automatic model selection mode can also get appropriately matched Hydrogen capabilities, while Google Search is also integrating the model to upgrade the experience.

Source

English text is being prepared.