Gemini 4 Hydrogen: the cheap worker, and a manager that hands off mid-task
Google shipped the cheap tier of Gemini 4 on Thursday, and the model is only half the story. Gemini 4 Hydrogen costs $1.50 per million input tokens and $7.50 per million output, flat across a 1M context, with up to 1M output tokens. Google says it scores 51 on the Artificial Analysis Intelligence Index at the default high thinking level, one point above Gemini 4 Argon, and 54 at a new xhigh level. Against 3.8 Flash, Google claims 2.5x the speed and 57% on Terminal-Bench 4.1 versus 18.2%. It is live in AI Studio, the Gemini API, the Gemini app, Search and Antigravity: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/
The more interesting launch is Antigravity Orchestrator (https://antigravity.google/blog/gemini-4-hydrogen-in-google-antigravity/). It switches between Carbon and Hydrogen inside a single task, in both directions. Carbon plans in read-only mode, Hydrogen edits and runs the tests, and when Hydrogen hits friction the task goes back to Carbon to re-plan. Google argues openly against static model routers that look at the first prompt, guess the difficulty and commit. Its claim: near-Carbon success with 24% less quota and 19% faster. One wrinkle worth noting: the Google blog says that number was measured on Terminal-Bench 4.0, the Antigravity post says 4.1. Both are Google's own figures, and no independent run exists yet. Orchestrator is Ultra-only for now.
This is the third day in a row the same idea shipped. Monday, Anthropic pitched Fable 5.2 as an Advisor and Orchestrator over cheaper models. Tuesday, Google put /swarm on Carbon, and Antigravity now says tens of thousands of swarms have run since. Thursday, the handoff became automatic and comes with a cost number. The frontier model is turning into a manager that is billed by the minute it spends thinking, while a cheaper model does the typing.
The router argument is the part to watch. Last week Amazon's Strands Decider and Google's own Decisions API sold small, fast models that pick between fixed options. Orchestrator says that the first prompt is a bad place to make that pick, because how hard a task is only shows up halfway through. If that holds outside Google's benchmarks, the routing layer moves from the front door into the middle of the loop, and the useful number for buyers becomes cost per finished task, not price per token.
← Back to all articles
The more interesting launch is Antigravity Orchestrator (https://antigravity.google/blog/gemini-4-hydrogen-in-google-antigravity/). It switches between Carbon and Hydrogen inside a single task, in both directions. Carbon plans in read-only mode, Hydrogen edits and runs the tests, and when Hydrogen hits friction the task goes back to Carbon to re-plan. Google argues openly against static model routers that look at the first prompt, guess the difficulty and commit. Its claim: near-Carbon success with 24% less quota and 19% faster. One wrinkle worth noting: the Google blog says that number was measured on Terminal-Bench 4.0, the Antigravity post says 4.1. Both are Google's own figures, and no independent run exists yet. Orchestrator is Ultra-only for now.
This is the third day in a row the same idea shipped. Monday, Anthropic pitched Fable 5.2 as an Advisor and Orchestrator over cheaper models. Tuesday, Google put /swarm on Carbon, and Antigravity now says tens of thousands of swarms have run since. Thursday, the handoff became automatic and comes with a cost number. The frontier model is turning into a manager that is billed by the minute it spends thinking, while a cheaper model does the typing.
The router argument is the part to watch. Last week Amazon's Strands Decider and Google's own Decisions API sold small, fast models that pick between fixed options. Orchestrator says that the first prompt is a bad place to make that pick, because how hard a task is only shows up halfway through. If that holds outside Google's benchmarks, the routing layer moves from the front door into the middle of the loop, and the useful number for buyers becomes cost per finished task, not price per token.
Comments