Antigravity now switches models mid-session for you: 24% less quota, a new default model, and a 45-day retirement clock

Google shipped two connected things this week, and the quiet one changes your config more than the loud one. The loud one is Gemini 4 Hydrogen, a fast execution model announced November 12. The quiet one is Orchestrator, arriving in Antigravity, which switches between models mid-session instead of running one model from start to finish. One session, two models, picked per task by the harness.

I read the announcement post, the AI Studio model docs, and the early coverage from InfoWorld and VentureBeat. Here is what switched, what it saves on Google's own benchmarks, and the two catches buried in the documentation.

What Orchestrator actually does

Orchestrator dynamically switches between models like Carbon and Hydrogen within the same session. That is Google's own description. InfoWorld's writeup spells out the split: a planner model, either Gemini 4 Carbon or Argon, explores the codebase, identifies dependencies, and develops an implementation plan. Hydrogen then handles execution: editing files, writing code, and running builds and tests.

Google's numbers, published in the announcement: on Terminal-Bench 4.1, Orchestrator reached near-Carbon task success rates while consuming on average 24 percent less quota and completing 19 percent faster. For parallel work, Carbon acts as the coordinator while Hydrogen powers worker and verifier agents building and testing changes across isolated workspaces.

The point is not that Hydrogen outthinks Carbon. It is that most turns in a coding session never needed frontier reasoning, and the harness now makes that call per task instead of you making it per session.

Hydrogen alone: fast for its class, honestly ranked

Standalone, Hydrogen fills the 3.8 Flash slot. On the Artificial Analysis Intelligence Index it scores 51 at its default high thinking level, one point ahead of Argon's 50, and 54 at the new xhigh level. Google says it finishes tasks 2.5 times faster than 3.8 Flash at the default setting, cuts fabricated claims by more than 5x, and more than triples the Terminal-Bench 4.1 score (57 percent versus 18.2 for Flash).

Two honest caveats: the speed and benchmark numbers are Google's own, measured against Google's previous models rather than an independent lab, and "near-Carbon" success at 24 percent less quota is an average across their benchmark tasks, not a promise about your repo.

Pricing is flat across the entire 1 million token context window: $1.50 per million input tokens and $7.50 per million output tokens, thinking tokens included. Cached input drops to $0.075 per million, a 95 percent implicit caching discount. The unified window is real, though: input, thinking, and output all draw from the same 1,048,576 tokens. On the API it sits on the free tier, and in Antigravity it requires a paid Google AI Pro or Ultra subscription.

Your default model moved, again

Buried in the AI Studio docs, one line: "The Antigravity agent now uses Gemini 4 Hydrogen (gemini-4-hydrogen) by default."

If you had sessions, snapshots, or cost estimates built on the previous default, they just moved. This is the same class of change Claude Code users hit when Fable 5.2 became the default and saved picks shifted underneath them. Auto-updating defaults are convenient until you are trying to reproduce last month's bill or benchmark.

The 45-day retirement clock

Here is the catch that matters most for anyone writing configs. The Google Cloud docs list Hydrogen as a short-term GA release, subject to a minimum 45-day retirement notice policy.

Read that against your own setup. If gemini-4-hydrogen is baked into agent configs, CI variables, docs, or scripts, retirement arrives on as little as 45 days of notice. Aliases drift silently, exact IDs die loudly. I wrote about this when "haiku" stopped meaning one thing across providers: pin what you can verify, and record what each ID was pinned for. A model that can vanish in 45 days is exactly the thing your config should treat as a version, not a constant.

Rolling the same routing yourself

You do not need Antigravity to run planner/executor routing. The AI Studio docs describe switching models across turns in one multi-turn session on the Interactions API, using previous_interaction_id: route planning turns to Carbon and execution turns to Hydrogen, preserving state between them. Hydrogen also supports 65,536-token hops on both streaming and non-streaming requests, double the 32,768 per non-streaming REST hop on Carbon and Argon, which matters in long execution loops.

The routing decision, sketched as config (my illustration, not Antigravity's format):

{
  "routing": {
    "plan":    { "model": "gemini-4-carbon",   "thinking_level": "high" },
    "execute": { "model": "gemini-4-hydrogen", "thinking_level": "high" },
    "verify":  { "model": "gemini-4-hydrogen", "thinking_level": "medium" }
  }
}

Hydrogen's four thinking levels (low, medium, high as default, and the new xhigh) give you an effort dial inside each stage, the same shift GitHub made when Copilot's model picker became an effort selector.

The catch nobody logs for: attribution

When the harness picks the model, your records need to say which one ran. An analyst quoted in InfoWorld's coverage put it directly: "CIOs need visibility into which model made a particular decision or code change. Without model-level logs and clear attribution, enterprises could find it difficult to audit AI-generated code, investigate errors, establish accountability when something goes wrong."

The savings from routing are real. So is the cost of losing per-turn model attribution. If your observability pipeline logs token counts but not model IDs per turn, this feature quietly breaks your audit trail.

One more math item: Google's own pricing docs move Carbon and Argon from introductory $2/$10 pricing to $4/$20 on December 1, while Hydrogen stays at $1.50/$7.50. Any routing policy you write this week has a different cost curve on December 1. Re-run the arithmetic after the switch.

Four checks for this week

1. Grep your configs and CI variables for gemini- model IDs. Every exact ID pointing at a short-term release needs a retirement plan, not a shrug.
2. Check your Antigravity default. If it moved to Hydrogen under you, update cost estimates and any saved comparison baselines.
3. If you route models yourself, record the model ID and thinking level per turn in your logs, not just token counts.
4. After December 1, redo the planner/executor cost math with Carbon at $4/$20 and see whether Argon or a Hydrogen-heavy split wins.

The direction is clear either way: the model field in your agent config is becoming a routing table. Google just made that literal in one IDE.