---
type: analysis
slug: the-model-router-is-becoming-the-product
title: "The model router is becoming the product"
description: "Gemini 4 Hydrogen matters less as another fast model than as part of a workflow that can assign routine execution and harder reasoning to different models. That makes routing quality—not any single benchmark—the bet builders should test."
published_at: 2026-11-13T14:00:00.000Z
topics: ["google","routing","gemini"]
source_urls: ["https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen","https://blog.google/innovation-and-ai/technology/developers-tools/build-with-gemini-4-hydrogen","https://simonwillison.net/2026/Nov/12/llm-gemini","https://ai.google.dev/gemini-api/docs/models/gemini"]
author: "Andrey Reshetnikov"
author_profile: https://newruntime.com/owner-profile.md
---
# The model router is becoming the product

Gemini 4 Hydrogen matters less as another fast model than as part of a workflow that can assign routine execution and harder reasoning to different models. That makes routing quality—not any single benchmark—the bet builders should test.

## Summary

Google introduced Gemini 4 Hydrogen for everyday coding and agent workflows, with API pricing and multiple thinking levels. It also added Orchestrator to Antigravity, which can switch between Hydrogen and Carbon during a session; Google says this approach approached Carbon’s Terminal-Bench success rate while using less quota and finishing faster. The larger shift is toward selling a managed division of labor between models, not simply a new model endpoint.

## What happened

Google’s November 12 announcement positions Hydrogen for rapid execution across coding, debugging, web development, and multimodal work. The Gemini API price is $1.50 per million input tokens and $7.50 per million output tokens, and Google says the model supports four thinking levels. In Antigravity, Orchestrator dynamically switches between Hydrogen and Carbon; Google reports near-Carbon task success on Terminal-Bench 4.0 with 24% less quota on average and 19% faster completion. Google also describes Carbon coordinating parallel Hydrogen worker and verifier agents. These figures are company-reported, not independent proof that routing improves every workload.

## What the archive adds

New Runtime’s earlier coverage described Gemini Agent as one interface and API for delegating enterprise work. This release adds a named execution model and an in-session switching mechanism, a practical extension of that product direction. An earlier New Runtime pattern note, “harness architecture outlives model choice,” argued that the orchestration layer can outlast a particular model choice; Orchestrator makes that proposition concrete, while leaving its real-world savings to be tested.

## Trend signal

Model launches increasingly package a role as well as a capability: one model handles routine execution, another handles harder judgment, and a harness decides which is needed. Google’s framing makes the router part of the user-facing value. This is a move from model selection at deployment time toward model allocation at task time. The commercial advantage depends on the router doing more than shifting tokens: it must preserve task quality while avoiding extra handoffs, retries, and review.

## Why now

Google is segmenting work by reasoning depth: its announcement positions Hydrogen for fast everyday execution and Carbon for deep, long-horizon reasoning (c1). The API price and multiple thinking levels make that division testable rather than purely conceptual (c2). Orchestrator brings model switching into Antigravity, giving Google a way to sell coordination alongside model access (c1). The underlying incentive is to improve completed-work economics, not merely reduce token prices; the launch’s quota and speed claims are framed at workflow level (c1).

## Connections to other events

The consequential unit is the workflow, not the token or the benchmark point. Hydrogen’s lower listed API rates and Orchestrator’s reported quota reduction support a plausible division of labor: route bounded execution to Hydrogen and reserve Carbon for coordination or difficult reasoning. But a cheaper step is not a cheaper completed task if routing errors trigger retries or human correction. The public announcement does not describe how Orchestrator chooses a model or expose controls that would let users inspect its decisions, so the savings claim is best treated as a hypothesis for production evaluation.

## Why it matters

Builders should evaluate mixed-model agents as systems. Log task success, elapsed time, total tokens, retries, tool failures, and reviewer edits on the same representative task set, then compare routed runs with a fixed-model baseline. Buyers should ask whether routing decisions are visible and configurable, and whether usage accounting includes failed attempts and verification. If the quality holds, teams can spend stronger-model capacity on the decisions that need it; if it does not, routing becomes another opaque layer to debug.

## What happens next

- **Google’s Gemini API pricing page will list a published Hydrogen API price by 2027-02-12.** (3 months, 75%). Watch for: A stable public price listing gives teams a basis for comparing Hydrogen’s per-token rates with routed task cost. Resolves on 2027-02-12, yes if: The Google Gemini API pricing page lists Gemini 4 Hydrogen input and output prices on 2027-02-12.
- **Google will publish an Antigravity release note or documentation update describing Orchestrator model selection or controls by 2027-02-12.** (3 months, 55%). Watch for: Documentation that exposes routing behavior or controls would let users evaluate whether the workflow gains are portable to their own tasks. Resolves on 2027-02-12, yes if: Google Antigravity documentation or release notes describe a model-selection rule or user control for Orchestrator by 2027-02-12.

## Limits and uncertainty

The speed, quota, benchmark, and quality figures in Google’s announcement are company claims. They do not establish independent performance across coding repositories or agent workloads, and the release does not explain the Orchestrator policy, its user controls, or the accounting behind quota savings. The strongest counterargument is that Google is showing a narrow product demo rather than a general architectural shift. That is credible: broad adoption depends on local quality and transparent controls, not the launch framing.

## Competing readings

- **The release signals a shift toward orchestration as a durable product layer: model routing can matter more than choosing one winning model** (likely, about 65%). What would tell it apart: Orchestrator switches models within a session, with Google reporting workflow-level quota and speed gains ([blog.google](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen)); Hydrogen is positioned as an execution model alongside deeper-reasoning peers ([blog.google](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen)).
- **This is primarily a routine model launch; Orchestrator’s reported result is a product demonstration whose savings may not generalize** (roughly even, about 50%). What would tell it apart: The reported routing results come from Google, and the public description does not specify the selection policy ([blog.google](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen)).

## Base rate

Of 83 earlier release stories about companies, models or products in the New Runtime archive over the last 12 months that are at least 3 months old, 71 (86%) were followed within 3 months by a access-change or pricing story on the same company or product line. The archive records 71 of 83 comparable release stories followed within three months by an access or pricing story. That makes a near-term access or pricing follow-up common in this set, but does not establish adoption or the durability of routing gains.

## What would change this view

- Antigravity docs or release notes state that users cannot inspect or control Orchestrator routing and provide no workflow-level evaluation (signal): The feature is more likely a bundled convenience than a durable, auditable routing layer.
- An independent task benchmark finds routed Hydrogen/Carbon runs do not preserve success while reducing total cost or elapsed time (signal): The central claimed benefit does not generalize beyond Google’s reported test.

## Skill to build now

Build a task-level evaluation harness. First exercise: select 20 past coding or tool-use tasks, define a pass condition and reviewer checklist for each, then record success, elapsed time, retries, tokens, and reviewer edits for fixed-model and routed runs.

## Reader takeaway

This week, build a small routing evaluation: choose 20 representative repository or tool-use tasks, define pass criteria before running them, and compare a fixed model with a routed Hydrogen/Carbon workflow. Record completion, retries, elapsed time, token use, and human edits. Keep the route only if it lowers total effort without lowering verified success.

## Open question

Will Google expose enough of Orchestrator’s routing policy for teams to diagnose bad assignments and reproduce the claimed savings outside Antigravity?

## Sources

- [Introducing Gemini 4 Hydrogen for Agentic Workflows](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen)

- [Gemini 4 Hydrogen: everyday coding and agentic workflows for developers](https://blog.google/innovation-and-ai/technology/developers-tools/build-with-gemini-4-hydrogen)

- [llm-gemini 0.37](https://simonwillison.net/2026/Nov/12/llm-gemini)

- [ai.google.dev](https://ai.google.dev/gemini-api/docs/models/gemini)

- [Gemini 4 Hydrogen | Gemini Enterprise Agent Platform documentation](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/4-hydrogen)

- [Google launches Gemini 4 Hydrogen — a faster, smarter, pricier Flash successor](https://venturebeat.com/technology/google-launches-gemini-4-hydrogen-a-faster-smarter-pricier-flash-successor)

- [Gemini 4 Hydrogen AutomationBench Results: 51.1% | Zapier](https://zapier.com/benchmarks/gemini-hydrogen-4-report)

- [AI Coding Agent Leaderboard - The Joule Index](https://joule.blankline.org/leaderboard/)

- [Gemini 4 Hydrogen Benchmark Results for Data Analysis](https://plotly.com/blog/gemini-4-hydrogen-plotly-data-analytics-bench/)

- [Expanding Google Antigravity for enterprise customers](https://cloud.google.com/blog/products/ai-machine-learning/expanding-google-antigravity-for-enterprise-customers)

- [SWE-Bench Pro V2: A Cleaner, Harder-to-Game Coding Benchmark](https://arxiv.org/abs/2610.24206)

Author: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Google](https://newruntime.com/topics/google.md): Explore the google topic hub.
- topic hub: [Gemini](https://newruntime.com/topics/gemini.md): Explore the gemini topic hub.
- collection: [Analysis](https://newruntime.com/analysis.json): Open the latest authorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
