Gemini 4 Hydrogen and Orchestrator
TL;DR Highlight
Carbon plans and Hydrogen executes: 24% less quota, 19% faster on Terminal-Bench 4.1
Who Should Read
Backend or agent developers weighing how to split work between expensive and cheap models in coding agents because of cost and latency
Core Mechanics
- Orchestrator swaps models mid-task
- It moves between Gemini 4 Carbon and Hydrogen in both directions within one trajectory.
- It appears as a new option in the model selector.
- Static routers break on long tasks
- A static router guesses difficulty from the first prompt and sends everything to the cheapest adequate model.
- As agents run longer, difficulty changes by stage, so this breaks down.
- Carbon plans, Hydrogen executes
- Carbon starts read-only, explores dependencies, and builds a plan with explicit re-planning triggers.
- Hydrogen edits, updates packages, and runs tests, handing back to Carbon when it hits friction.
- 24% less quota on Terminal-Bench 4.1
- Orchestrator reached near-Carbon task success.
- It used 24% less quota and finished 19% faster on average.
- Hydrogen standalone numbers and thinking levels
- It scores 51 on the Artificial Analysis Intelligence Index versus Argon's 50, and 54 at its highest setting.
- It scores 57% on Terminal-Bench 4.1 versus 3.8 Flash's 18.2%.
- Reasoning depth has four levels: low, medium, high (default), and new xhigh.
- You can wire roles yourself with the SDK
- In the Antigravity SDK issue triage demo, Hydrogen handles triage, search, questions, and UI fixes from PR comments.
- Carbon handles larger refactors, new features, and security checks.
- Availability differs by plan
- Hydrogen rolls out to paid Google AI Pro and Ultra subscribers.
- Orchestrator (Carbon) is Ultra only, while Orchestrator (Argon) is available on Pro and Ultra.
Community Reactions
- Similar designs exist elsewhere, per commenters
- One comment said Anthropic's advisor mode and GitHub's multi-model pipeline are essentially the same idea.
- The commenter also asked how this works with caching.
- Worry about delegating model choice, from experience
- One commenter pins models per subagent and worries models are optimistic about their own abilities.
- They would rather pin models per task domain, such as implementor (hard) and implementor (easy).
- A commenter reported bigger savings
- Using Mjolnir to bridge Opus to DeepSeek Flash over ACP, one commenter saw 60% savings.
- They were surprised the official figure is only about 25%.
- Argon and Carbon felt slow to some users
- One commenter switched to 3.8 Flash because agy loads many files into context even for simple tasks.
- Several commenters found the new naming scheme confusing.
How to Apply
- Break a recent agent task into stages
- In the next 30 minutes, split one recent coding task into explore, plan, edit, and test.
- Write down the difficulty each stage needs and mark where the strong model is essential.
- Separate the planning model from the execution model
- In your subagent config, assign a strong model to planning and a fast model to edits and tests.
- Write explicit re-planning triggers into the plan.
- Add a rule to hand back to the strong model
- Write a rule such as: if the same test fails twice, pass the state to the planner.
- The planner inspects the state, repairs the plan, and hands it back to the executor.
- Compare against your single-model baseline
- Run the same task set with one strong model and with the split setup.
- Record success rate, quota (tokens), and time to see whether savings like 24% and 19% reproduce.
Try It
# Paste into Claude Code or ChatGPT (a simplified prompt version of the Orchestrator idea)
You alternate between two roles: Planner and Executor.
[Step 1: Planner, read-only]
- Do not modify code. Explore dependencies and list ambiguities.
- Write a step-by-step plan and add a re-planning trigger to each step (e.g. the same test fails twice).
[Step 2: Executor]
- Only make edits, update packages, and run tests according to the plan.
- If a re-planning trigger fires, stop, summarize the current state, and hand it to the Planner.
[Step 3: Planner returns]
- Inspect the state, repair the plan, and hand it back to the Executor.
Task: {{task description}}
Next to try: run the same task with the Planner and Executor on different models or subagents.Terminology
Related Papers
Gemini 4 Neon
Gemini 4 Neon cuts cost per task 32% versus Carbon, at $3 input and $15 output per million tokens
AI is killing All About Berlin
After Google AI Overviews, a Berlin guide site lost 70-75% of its traffic.
Not Everything Needs an LLM
PostHog used small decision models to skip 98.7% of expensive LLM calls
GPT‑6.2 Astra and Sol
GPT-6.2 Astra and Sol cut costs sharply, but over-engineering and early-stop reports are mixed
Show HN: TrnRT – An open-source LLM inference engine for AWS Trainium
One trn2.48xlarge serves GLM-5.3-Flash through an OpenAI-compatible API using NKI megakernels.
Know When to Zoom: How Small Is Too Small for Vision-Language Models?
Splitting large photos into tiles raised Claude and Gemini accuracy by 11.1–14.5 points
Understanding Positional Embeddings in Modern LLMs
A walkthrough of how LLMs encode token position, from GPT-2 to RoPE
Show HN: DeepSeek V3 in ~200 lines of pure Jax
DeepSeek V3's core design in 190 lines of JAX, trained on a MacBook Air in 2.6 hours