Gemini 4 Hydrogen Surpasses Argon, AA Evaluation and Pricing

Key Takeaways

1Gemini 4 Hydrogen is globally available with four reasoning levels and a context window of one million tokens

2Artificial Analysis rates it at 51–54 depending on the level, 8% hallucinations, and performance lagging on terminal tasks, with costs per task ranging from $0.29 to $3.16 depending on the setting

3Google claims significant gains on Flash, an advantage over Argon, and improvements in the Gemini app and Search

Why it matters : Gemini 4 Hydrogen aims to combine versatility, speed, and reduced cost while positioning itself against competing models for everyday and professional use.

Gemini 4 Hydrogen is offered by Google, providing four levels of reasoning and a context capacity of up to one million tokens. Artificial Analysis assigns it a score between 51 and 54 depending on the level, along with an 8% hallucination rate, while noting a lag in terminal tasks. The model is priced at $1.50 for input and $7.50 for output per million tokens and is accessible via the API, the Gemini app, Search, and Antigravity.

Artificial Analysis measures 51–54 points and 8% hallucinations

The evaluation by Artificial Analysis places Gemini 4 Hydrogen at the level of Claude Sonnet 5.5 at its peak, with a score of 51 at the default "high" level, trailing behind GPT-6.1 Sol and Claude Opus 5.5 at 52. On AA-Omniscience, the model only invents elements 8% of the time and surpasses other Gemini models, including Carbon, in presentation quality assessed by AA-Briefcase. In agentic terminal work, the measures show 54% on Terminal-Bench 4.0 and 50% on Terminal-Bench-Science 0.1, while Sonnet 5.5 (max) scores 64% and 53%, and GPT-6.1 Sol (max) scores 56% and 58%. In terms of cost per task, AA estimates $3.16 at the "xhigh" level, about one-third the cost of Sonnet 5.5 (max) for similar intelligence, but approximately 2.5 times that of GPT-6.1 Sol (max). Moving from "high" to "xhigh" raises the index from 51 to 54 at the expense of nearly doubling output consumption, increasing the average cost from $1.87 to $3.16. The "medium" and "low" levels show costs of $0.80 and $0.29, with indices of 41 and 27. AA indicates that at high, medium, and low levels, the model reaches the Pareto frontier for observed rates.

Google claims gains over Flash and an advantage over Argon

Google presents Hydrogen as a standalone overhaul compared to Gemini 3.8 Flash. At the default reasoning level, the company reports tasks completed 2.5 times faster and a Terminal-Bench 4.1 score rising from 18.2% (Flash) to 57%, with hallucinations reduced by a factor of five and improved recovery after tool or search failures. On the Artificial Analysis Intelligence Index, Google indicates 51 points for Hydrogen compared to 50 for Argon at the "high" level, and up to 54 at "xhigh", trailing behind Claude Sonnet 5.5 and GPT-6.1 Sol at 55. Hydrogen ranks fifth on Agent Arena, and Google claims it is first on Vals Tax Agent Bench with 52.3% and co-first on Vals Legal Research Bench with 55.8%. This superiority over Argon is asserted just four weeks after the launch of Hydrogen.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Four levels of reasoning and context up to one million tokens

Hydrogen offers four levels of reasoning—low, medium, high (default), and a new "xhigh"—while Argon only presents two, medium and high. The model can accept and produce up to one million tokens, with explicit support for multimodal knowledge use cases. The range of covered benchmarks includes Terminal-Bench 4.0 and 4.1, Agents’ Last Exam v1.1, AutomationBench, Vending-Bench 2, OSWorld-2.1, MKA, Blueprint Bench 2, GraphWalks, as well as Terminal-Bench-Science 0.1. On Google's Multimodal Knowledge Agent Benchmark, Hydrogen is measured at 53.4%, and at 71.6% on Agents’ Last Exam v1.1.

Token pricing and announced cache costs

The announced billing is $1.50 per million tokens for input and $7.50 per million tokens for output for Hydrogen. Cache impacts are priced at $0.075 per occurrence.

Widespread access and integrations with Google products

Hydrogen is being rolled out globally and is accessible via the Gemini API in AI Studio and Android Studio, through Google Antigravity for Pro and Ultra subscribers, via Gemini for Business for enterprise clients, and for everyone in the Gemini app and in Search. Google reports improved adherence to tone, format, and guidelines during long exchanges in the app, and the ability to integrate context from Gmail, Drive, as well as third-party applications like Airtable or Dropbox. In terms of search engine capabilities, Google claims improved understanding of queries, including those with typos, composite questions, or very specific requests, more visual results, and a broader range of relevant links. In its comparisons, Google positions Hydrogen ahead of Claude Sonnet 5.5 and GPT-6.1 Terra in knowledge work and multimodality, while acknowledging Sonnet's superior scores on terminal tasks. Hydrogen is described by Google as versatile and fast, oriented towards everyday tasks, whereas Carbon and Argon aim for deep reasoning; it is also presented as the counterpart to the previous Flash lineage.