🇫🇷 Cet article est aussi disponible en français.

Gemini 4 Hydrogen: The Economics of Fast Agent Execution

Gemini 4 Hydrogen: The Economics of Fast Agent Execution
🇫🇷 Cet article est aussi disponible en français.
📑 Table of Contents

TL;DR

Google’s Gemini 4 Hydrogen targets the execution layer of AI agents: writing code, using tools and checking results. Its price and vendor-reported benchmarks make it a candidate for routine work, but the real test is the cost of an accepted result, including retries and human review.

Introduction

On November 12, Google introduced Gemini 4 Hydrogen as a fast model for everyday coding and agentic workflows. It also introduced Orchestrator in Antigravity, which switches between Hydrogen and the more deliberative Carbon during a task. This extends the planner-and-worker pattern discussed in our Carbon and swarm analysis, but applies it to a single evolving session rather than assuming every job needs parallel agents. (Sources : Google — Introducing Gemini 4 Hydrogen; Antigravity — Gemini 4 Hydrogen and Orchestrator)

The token bill versus the completed job

Google prices Hydrogen at $1.50 per million input tokens and $7.50 per million output tokens across its one-million-token context window. (Source : Google — Introducing Gemini 4 Hydrogen)

At those rates, a hypothetical direct API workload of 100 million uncached input tokens and 20 million output tokens costs $300. This is token arithmetic, not an Antigravity subscription price: it excludes caching, tools, taxes and any other charges. Longer context can reduce the need to split a job, but repeatedly sending a large context can also inflate the bill. Capacity is not a guarantee that the model will retrieve the right detail.

A more useful metric is total spend divided by accepted tasks. A model with a higher per-token price can still be economical if it needs fewer retries or less human correction; the reverse is equally possible. Record token usage, elapsed time, pass rate and review minutes under the same task set before changing a production default.

Benchmarks: meaningful, but not portable

Google reports Hydrogen scored 57% on Terminal-Bench 4.1 versus 18.2% for Gemini 3.8 Flash. Google also says Hydrogen completed tasks 2.5 times faster than 3.8 Flash at its default high thinking level. (Source : Google — Introducing Gemini 4 Hydrogen)

Those are provider-reported results, not a promise for a particular repository. The gap is 38.8 percentage points on this test, but success in a benchmark does not establish the quality of patches, the security of tool calls or performance across different environments. Look for evaluation settings and independent replication; test against representative tasks with held-out acceptance checks.

There is independent context for the intelligence score: Artificial Analysis lists Hydrogen’s xhigh setting at 54 on its Intelligence Index. That is an aggregate across several evaluations, not a standalone measure of an agent’s reliability in your workflow. (Source : Artificial Analysis — Gemini 4 Hydrogen release)

Dynamic routing and the handoff problem

Antigravity describes a bidirectional handoff: Carbon may plan in a read-only phase, Hydrogen executes edits and tests, and Carbon returns if execution hits friction. Google reports Orchestrator used 24% less quota and completed tasks 19% faster while reaching near-Carbon success. (Sources : Antigravity — Gemini 4 Hydrogen and Orchestrator; Google — Introducing Gemini 4 Hydrogen)

There is a reporting inconsistency worth preserving: Google’s launch post calls that evaluation Terminal-Bench 4.0, while Antigravity calls it 4.1. Treat the savings as vendor-reported and the benchmark version as unresolved. Quota savings are not necessarily dollar savings, and a handoff can lose context or repeat work. Capture the state, failed tests and remaining constraints at each transition; compare the routed run with Carbon-only and Hydrogen-only baselines.

Access and guardrails

Antigravity says Hydrogen is rolling out to paid Google AI Pro and Ultra subscribers, while Orchestrator is rolling out to Ultra subscribers. API usage and Antigravity access are separate billing paths. (Source : Antigravity — Gemini 4 Hydrogen and Orchestrator)

For a first deployment, use isolated workspaces, restrict credentials and require human approval for external writes or deployment. Faster execution multiplies both useful actions and mistakes; an agent’s own verifier should not replace independent tests or review of the final diff.

FAQ

Is Hydrogen available through the Gemini API?

Google says it is rolling out through the Gemini API and AI Studio, as well as consumer and enterprise products.

Is Orchestrator the same as /swarm?

No. Orchestrator switches models during a single session; /swarm coordinates parallel agents on larger tasks.

Do benchmark scores guarantee production reliability?

No. Results depend on the tasks, tools and evaluation setup; run your own held-out tests.

Does lower quota usage mean a lower API bill?

Not necessarily. Antigravity quota and direct API token charges are different measures.

Further Reading