Google Antigravity gives coding work to 2 Gemini 4 models in turn

Google added a tool to its AI coding software. The tool, Orchestrator, makes a planning model, Gemini 4 Carbon or Argon, and a small new model, Gemini 4 Hydrogen, work together in one session. The price of Hydrogen is $1.50 for each million input tokens (small pieces of text) and $7.50 for each million output tokens.

How it works

The planning model reads the code, finds the parts that connect and writes a plan. Hydrogen then edits files, writes code and does the tests. If a test does not give a good result again and again, Hydrogen gives control back to the planner, which changes the plan. Google tells that a developer does not change models.

The numbers

Google gave no test results for Orchestrator. Analyst Pareekh Jain tells that it can give more speed and more code quality. Amit Kumar Jena, an AI expert at a consulting firm, tells that it can lower AI costs. The 2 persons think this, but they give no measured results.

What is not known

Jena tells that the documentation does not give per-step records that name the model for each step. An analyst tells that firms can find it hard to check AI code or find the cause of an error.

Insight: Google seems to move model selection from the user to its software

Google launched Hydrogen on 12 November 2026. It tells that Hydrogen gets a score of 51 on the Artificial Analysis Intelligence Index, above the Argon model at 50. On 8 October 2026, Google showed a Gemini Agent that also selects the model for each task. This reads as a plan to keep the model selection in Google software, and Google then controls which model does the work. Look for records of each step.