Google’s Gemini 4 coding agent now switches models mid-task — and finishes 19% faster

Google unveiled Gemini 4 Hydrogen on Thursday, alongside a new system that dynamically switches AI models during coding tasks. Rather than relying on the most capable model throughout a session, Google combines Hydrogen with the more powerful Gemini 4 Carbon, switching between the two models as the work progresses. This combination achieves results nearly as good as using Carbon alone, but at 24% lower cost and 19% faster task completion.

The switching process occurs through Orchestrator, a new feature in Google's Antigravity development environment, where Carbon handles complex reasoning tasks while Hydrogen manages code changes and testing. Google's AI engineering VP, Clement Farabet, describes the pairing as "a huge unlock for agentic coding," and Antigravity team lead Varun Mohan notes that this combination is "very popular internally."

However, Google's benchmark reporting shows discrepancies, as the main Gemini announcement refers to version 4.0, while the Antigravity announcement mentions version 4.1. The company has not released Carbon's standalone success rate, task breakdowns, or detailed routing information for Orchestrator, making it difficult to fully understand the efficiency gains.

Hydrogen is positioned as a "daily driver for agentic workflows" and costs $1.50 per million input tokens and $7.50 per million output tokens, with these rates applying across its full million-token context window. The model is accessible through Google AI Studio, the Gemini API, the Gemini app, and Google Search. In the Artificial Analysis Intelligence Index, Hydrogen scores 51 at its default high thinking level, slightly ahead of the more expensive Gemini 4 Argon, which scores 50.

When set to Google's "xhigh" reasoning level, Hydrogen's score increases to 54. Although the one-point lead is modest, it highlights the rapid progress of Google's lower-cost models, which are increasingly matching the performance of top-tier models just a few weeks ago. The company also claims that Hydrogen completes tasks 2.5 times faster than Gemini 3.8 Flash at the default setting, with a Terminal-Bench 4.1 score of 57% versus Flash's 18.2%.

However, Google has not disclosed how Orchestrator determines when to switch models or the additional costs associated with these model transitions, which are currently limited to Google AI Ultra subscribers.