Gemini 4 Hydrogen
TL;DR Highlight
Gemini 4 Hydrogen runs 2.5x faster than 3.8 Flash and scores 57% on Terminal-Bench 4.1.
Who Should Read
A backend developer running coding agents or tool-calling workflows on Gemini 3.8 Flash who is weighing a switch for speed and cost.
Core Mechanics
- Hydrogen is the execution model
- Carbon and Argon are the frontier models for deep, long-horizon reasoning.
- Hydrogen is built for rapid execution across coding and agentic workflows.
- 2.5x faster than 3.8 Flash
- At the default high setting, tasks finish 2.5 times faster than with 3.8 Flash.
- Terminal-Bench 4.1 went from 18.2% on 3.8 Flash to 57%.
- Fabricated claims are cut by more than 5x.
- Flat pricing across the full context
- The entire 1M-token context costs $1.50 per million input tokens and $7.50 per million output tokens.
- Orchestrator switches models for you
- Orchestrator in Antigravity switches dynamically between Carbon and Hydrogen within one session.
- On Terminal-Bench 4.0 it reached near-Carbon success while using 24% less quota and finishing 19% faster on average.
- For parallel agents, Carbon coordinates while Hydrogen runs worker and verifier agents.
- Four thinking levels now
- The levels are low, medium, high (default), and a new xhigh.
- On the Artificial Analysis Intelligence Index, high scores 51 versus Argon's 50, and xhigh reaches 54.
- Agent safety numbers improved too
- Overreach in agentic tasks drops by up to 81% versus 3.8 Flash.
- Resilience to indirect prompt injection is more than 2.3x on the Gray Swan IPI benchmark.
- It is up to 2x less likely to bypass permission boundaries on blocked tasks.
Community Reactions
- Opinion that the pricing is confusing
- One commenter said the price matches undiscounted 3.8 Flash and is higher than what people pay today.
- They worried price-sensitive enterprise customers may leave when the discount ends at year end.
- Own workload: 37% higher cost, 40-60% of the time
- One user measured 37% higher cost per task than 3.8 Flash.
- Total time per task fell to 40-60%, with median and p90 improving equally.
- Both models completed all eval tasks on medium reasoning, and speed was the bottleneck.
- Good at image-to-HTML builds
- One user said it is the first non-Anthropic small model to capture complex flows.
- They felt the quality sits between Sonnet and Opus.
- Added a wrong fun fact on a visual puzzle
- One user had it solve a Bongard problem and it got the answer right.
- It then appended an incorrect side note and made things up entirely on a harder puzzle.
How to Apply
- Run one current agent task on Hydrogen and compare
- Today, pick 5-10 evaluation tasks you run on 3.8 Flash in AI Studio or the Gemini API and rerun them with the same prompts.
- Record cost per task, total time, and success in a table.
- Compare medium and high thinking levels.
- Measure the cost-latency curve per thinking level
- Run the same task set at low, medium, high, and xhigh.
- Stop at the level where quality stops improving and set that as your default.
- Split planning to a big model and execution to Hydrogen
- If you run parallel agents, assign the coordinator to Carbon or Argon.
- Assign worker and verifier agents to Hydrogen and compare token costs.
- Add tool-failure and empty-search cases to regression tests
- Test whether the model fabricates answers when a tool errors or a search returns nothing.
- Compare 3.8 Flash and Hydrogen responses side by side.
Try It
# Paste into Claude or Gemini chat (simplified version of the idea: audit your model routing design)
I run a coding agent workflow. Classify the task list below into "planning, design, hard debugging" and "writing code, running tests, verifying changes".
Task list:
{{task list}}
Criteria:
- The former goes to a large model that does deep reasoning (coordinator).
- The latter goes to a fast execution model (worker/verifier).
Give a one-line reason for each task, and end with a table of task groups that can run in parallel.
Next step: send the execution tasks from the table to Hydrogen and measure cost and time.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