Skip to content

Gemini 4 Hydrogen arrives with confusing pricing

Google's new cheapest Gemini 4 model undercuts its siblings by 25% today, but the gap widens to 75% once Carbon and Argon double in price on 1 December.

Draftship Editorial3 min read

Google launched Gemini 4 Hydrogen on 12 November 2026 as the lowest cost member of the Gemini 4 family. Simon Willison marked the launch with version 0.37 of llm-gemini, his LLM plugin for accessing Google's Gemini models. The Gemini 4 Hydrogen pricing is the part builders should read closely, because the headline discount is small until a scheduled price rise for the two larger models lands on 1 December.

You can read the full release note and price breakdown on Simon Willison's weblog.

The numbers as published

All figures are per million tokens:

Model Dates Input Cached input Output
Gemini 4 Carbon Through Nov 30, 2026 $2.00 $0.10 $10.00
Gemini 4 Carbon From Dec 1, 2026 $4.00 $0.20 $20.00
Gemini 4 Argon Through Nov 30, 2026 $2.00 $0.10 $10.00
Gemini 4 Argon From Dec 1, 2026 $4.00 $0.20 $20.00
Gemini 4 Hydrogen (no date change listed) $1.50 $0.075 $7.50

Willison says he finds the scheme genuinely confusing, and the table shows why. Hydrogen, the budget option, is only about 25% cheaper than its more capable peers while the introductory pricing holds. After 1 December it will sit at roughly a 75% discount, simply because Carbon and Argon double.

Two models sharing one rate card

Carbon and Argon carry identical prices on both sides of the December change, which raises the obvious question of why Google sells two of them. Google's answer, as quoted in the post, is that Argon "delivers Gemini 4 intelligence with faster responses and lower token spend per turn compared to Gemini 4 Carbon". The same rate per token can still mean a different bill per task if one model burns through fewer tokens getting there.

Where Hydrogen is meant to fit

Google positions Hydrogen as an execution model rather than a strategist. In the description quoted by Willison, teams use Carbon and Argon for architectural judgment and for untangling subtle bugs, while Hydrogen handles writing code, running tests and verifying changes. It can also serve as the execution engine inside a multi-model workflow alongside the larger models.

Two further details matter. Hydrogen is the new free tier model, and it is the model that will power AI features in Google Search. Willison also generated pelican images with all three models at their three supported effort levels, and found the quality difference for Hydrogen stark given how close the prices are.

What this changes if you are paying for tokens

A cheap tier inside the Gemini family is not automatically cheap in the market. The post compares Hydrogen to Claude Haiku 5.5 and GPT-6 Luna, both priced at $0.10 input and $0.50 output, which makes Hydrogen about 15 times those rates. If your app makes many small calls, that multiple will dominate your bill well before the Gemini 4 price change on 1 December does. The practical move is to treat the December date as a deadline for rechecking any cost estimate that assumed current Carbon or Argon pricing.

Limits and open questions

The post does not give context windows, rate limits, regional availability or benchmark scores for any of the three models. It lists no future date for Hydrogen, so whether its $1.50 rate is fixed or introductory is unclear. The quality assessment comes from one informal image test. And while llm-gemini 0.37 is published as the release accompanying the launch, the source does not spell out which changes the version contains.

What it means for building apps with AI

Model choice is now a real design decision rather than a detail, and pricing that shifts on a known date makes it a recurring one. If you are assembling an app that calls a model, Draftship's paid plans include a choice of AI model, and its guide to using the AI chat and choosing a model covers how that selection works. Check current provider prices yourself before committing to a budget, since rate cards move.

Primary sources

This article was drafted with AI assistance from the primary sources listed below, then reviewed and approved by a Draftship editor.