Gemini 4 Hydrogen: Google's Budget Model Nearly Matches Claude Sonnet 5.5

What's going on

Google DeepMind unveiled Gemini 4 Hydrogen on November 12, 2026. It's the third model in the Gemini 4 family after Argon (September 30) and Carbon (October 28) – three new models in six weeks. The models are named after chemical elements. Hydrogen is the lightest element, and that's the idea behind the model: a fast, cheap everyday workhorse. In the past, Google would have called a model like this "Flash." According to 9to5Google, Google deliberately dropped that name this time.

The model is available worldwide starting today: in the Gemini app, in Google Search (AI Mode), in Google AI Studio, via the Gemini API, via Vertex AI and for businesses through "Gemini for business." Paying Google AI Pro and Ultra subscribers can also use it in the Google Antigravity coding environment, and paid users get it in Android Studio as well. Free Gemini app users get Hydrogen through automatic model selection whenever the system decides it's a good fit for the request. According to Business Insider, Argon and Carbon had so far been limited to paying users.

The launch wasn't a total surprise. According to Cryptobriefing, the community had been testing the model since late October under the name "gemini-3.8-flash." Business Insider reported the day before launch that Google employees could try it on the internal development platform "Jetski," where one engineer described it as "very capable, fast and efficient."

How Hydrogen stacks up

Independent research firm Artificial Analysis rates models with its Intelligence Index. It's an overall score built from several tests – in version 5.0.1, nine individual evaluations that rely heavily on non-public datasets. Hydrogen has four reasoning levels: low, medium, high (the default) and the new xhigh. The higher the level, the longer the model "thinks" before answering. That costs more compute and more tokens, the chunks of text you're billed for.

On xhigh, Hydrogen ranks seventh overall. Ahead of it are Claude Fable 5.2 (65), Gemini 4 Carbon (60), Claude Opus 5.5 (59), GPT-6 Astra (57), and GPT-6.1 Sol and Claude Sonnet 5.5 (55 each). That puts it ahead of Gemini 4 Argon (50), Meta's Muse Wonder Full (50) and OpenAI's GPT-6.1 Terra (51). What stands out is that even on the default setting, the cheap Hydrogen scores 51 – one point ahead of Argon, Google's own flagship.

Strengths: few fabrications, solid document work

Hydrogen's most striking feature is its low hallucination rate – the tendency to make up answers instead of admitting it doesn't know. The AA-Omniscience test penalizes confidently wrong answers. According to Artificial Analysis, Hydrogen fabricates something only 8 percent of the time there. Claude Sonnet 5.5 does so 43 percent of the time, GPT-6.1 Sol 52 percent. Only Argon and Carbon guess less often. The rate does rise at lower reasoning levels, though. Google says fabricated claims are more than five times less frequent than with its predecessor, Gemini 3.8 Flash.

The GDP.pdf test measures visual understanding of long, complex documents. Hydrogen hits 33 percent there on just the medium setting, tying for first place. On AA-Briefcase, a test of autonomous office work, evaluators rated its presentation of results higher than any other Gemini model, Carbon included.

Google published its own numbers pitting the model against Claude Sonnet 5.5 and GPT-6.1 Terra. By Google's count, Hydrogen beats Sonnet 5.5 on seven out of ten tests. A few examples:

• Vending-Bench 2: The model runs a simulated business. Hydrogen earns $14,549, nearly $3,400 more than Sonnet 5.5.
• SWE-sweep: The model has to find and fix bugs without breaking existing code. Hydrogen scores 13.5 percent, more than four times Sonnet 5.5's result.
• HLE-Diamond: Hydrogen reaches 58.8 percent. Argon gets 69.1 percent, while Sonnet 5.5 trails Hydrogen by more than five points.
• Vals Tax Agent Bench: First place with 52.3 percent. On the Vals Legal Research Bench, Hydrogen ties for first with 55.8 percent.
• Agents' Last Exam v1.1: 71.6 percent on end-to-end professional workflows.

Google also cites figures on the reliability of agents – AI systems that carry out multiple steps on their own and use tools. Compared with 3.8 Flash, Hydrogen overreaches on a task up to 81 percent less often and underdelivers up to 79 percent less often. Shortcuts and inappropriate actions are twelve times rarer. On Google's internal honesty evaluation for agents, the model scores 96.7 percent. All of these are the vendor's own figures.

Weakness: working in the terminal

Hydrogen is weaker on the command line (the text-based window where developers type commands and launch programs). That was already true of Argon. According to Artificial Analysis, Hydrogen scores 54 percent on Terminal-Bench 4.0 and 50 percent on Terminal-Bench-Science 0.1. Claude Sonnet 5.5 hits 64 and 53 percent, GPT-6.1 Sol 56 and 58 percent. The xhigh setting helps on the science tasks (from 41 to 50 percent) but does nothing on the regular Terminal-Bench.

Google's own table tells the same story. On Terminal-Bench 4.1, it's 57.0 versus 63.9 percent in Sonnet's favor. On OSWorld-2.1 (offline portion), which tests operating a computer, it's 77.6 versus 83.9 percent. On Terminal-Bench-Science, Sonnet leads 59.9 to 51.4 percent, and Hydrogen also trails Terra there. Still, the leap over its predecessor is big: 3.8 Flash managed just 18.2 percent on Terminal-Bench 4.1.

Pricing and speed

Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens. Cached input (caching is useful when you keep sending the same long text) is 95 percent cheaper, at $0.075. The price stays the same across the full 1 million token context window – the amount of text the model can take in at once. According to 9to5Google, the model can also output up to 1 million tokens. It accepts text, images, video and speech as input and outputs text.

The list price is 25 percent below Sonnet 5.5's ($2 and $10). In practice the gap is even wider, because Sonnet burns through a lot of tokens per task: on xhigh, Artificial Analysis says Hydrogen costs only about 32 percent of what Sonnet 5.5 costs per task. Compared with GPT-6.1 Sol, however, Hydrogen on xhigh is about 2.5 times as expensive.

The default setting probably offers the best value. xhigh nearly doubles output tokens for just three extra points. On high, Hydrogen matches Argon at roughly 19 percent of its cost per task. The model outputs around 200 tokens per second. According to Artificial Analysis, on the low through high settings there's no model that is both smarter and faster. Google says Hydrogen completes tasks 2.5 times faster than 3.8 Flash. Partners report similar numbers: Pixelcut measured 1.8 times the speed, Box roughly double with 12 percent fewer tokens. Cartwheel's chief scientist called the model "thorough and skeptical," saying it spotted two problems in an approach his team had spent ten days on.

Gemini app, Search and the Orchestrator

In the Gemini app's model picker, Hydrogen is labeled "Everyday help." Above it sit 3.1 Pro for in-depth research, 4 Argon for code and analysis, and 4 Carbon for the most complex tasks. Google promises Hydrogen sticks better to tone, format and instructions and more often calls the right tools, such as connected apps. In Google Search, it's supposed to better understand multi-part queries or ones with typos, deliver more visual answers and show more links.

For developers, Antigravity gets a new Orchestrator. Instead of running an entire session on one model, it switches dynamically between Carbon and Hydrogen, handing routine work to the cheaper model. Google says this came close to Carbon's success rate on Terminal-Bench 4.0 while using 24 percent less quota and running 19 percent faster. With parallel agents, Carbon can coordinate the work while multiple Hydrogen instances build and test changes in separate workspaces.

The bigger picture

Google canceled Gemini 3.5 Pro this year and was long seen as falling behind. With Gemini 4, it's back at a rapid clip. Hydrogen is the model with the widest reach: free users, Search and developers get a model that keeps pace with Anthropic's and OpenAI's mid-tier models on many tests, while costing significantly less and making up answers less often.

A few points deserve skepticism, though. Google's own table deliberately compares against Sonnet 5.5, not Anthropic's top models like Fable 5.2. Artificial Analysis results can't be directly compared with older scores: Sonnet 5.5 scored 56 in an earlier index version and now scores 55. For coding agents that spend a lot of time in the terminal, Anthropic stays ahead. And there's a question Cryptobriefing raises: if the cheap model beats the flagship Argon on its default setting, when is Argon still worth it? Google's answer is the Orchestrator, which routes each task to the right model. How well that works day to day remains for independent testing to show.