Google's Gemini 4 Hydrogen outscores the bigger Argon just four weeks after its launch

Google has released Gemini 4 Hydrogen, a fast general-purpose model aimed at developers and everyday users. In Google's positioning it succeeds Gemini 3.8 Flash and sits below the frontier models Gemini 4 Carbon and Gemini 4 Argon, which the company built for deep, long-horizon reasoning. Hydrogen is meant to handle routine work quickly and cheaply. It is available now globally through the Gemini API in AI Studio and Android Studio, the Gemini app, Google Search, Google Antigravity for paying Pro and Ultra subscribers, and Gemini for Business.
The pricing is $1.50 per million input tokens and $7.50 per million output tokens, with cache hits billed at $0.075. The context window is one million tokens on the input side, and the model can also produce up to one million output tokens — a large output ceiling that matters for long documents, summaries and multimodal knowledge work such as digesting big PDFs, spreadsheets or videos.
Hydrogen ships with four reasoning or thinking levels: low, medium, high (the default) and a new xhigh tier. Argon offers only medium and high. That choice of levels turns out to be central to how the model is priced in practice, because higher tiers buy accuracy with more output tokens.
On benchmarks, Google says Hydrogen beats the larger Argon on the Artificial Analysis Intelligence Index at the default high level, 51 versus 50, and reaches 54 at xhigh. That is a striking result for a model positioned as the cheaper, faster sibling: the smaller model outscores the bigger one. At xhigh, Hydrogen lands just behind Claude Sonnet 5.5 and GPT-6.1 Sol, which each score 55 at their maximum thinking level. Argon trails at 50.
That benchmark reality creates some awkwardness in Google's lineup. When the company announced Carbon two days earlier, it recommended staying on Argon for daily coding work. Hydrogen makes that advice largely moot, leaving Argon looking obsolete less than four weeks after its launch. This suggests buyers evaluating the Gemini family should compare on measured task performance and price rather than on model tier names, at least until Google clarifies the intended roles.
Against the previous Flash generation, Google claims a substantial step up. At the default thinking level it says Hydrogen finishes tasks 2.5 times faster. On Terminal-Bench 4.1 the score rises from 18.2 percent for Flash to 57 percent for Hydrogen. Hallucinations fell by a factor of five, and the model recovers better when a tool call fails or a search returns nothing — a practical property for agentic pipelines where errors compound.
On open-ended agent tasks such as research, data analysis and code navigation, Hydrogen ranks fifth on the Agent Arena leaderboard. Google says it takes first place on the Vals Tax Agent Bench at 52.3 percent and ties for first on the Vals Legal Research Bench at 55.8 percent. On Google's new Multimodal Knowledge Agent Benchmark it scores 53.4 percent, and 71.6 percent on Agents' Last Exam v1.1. Across Google's own comparison charts, Hydrogen beats Claude Sonnet 5.5 and GPT-6.1 Terra on the knowledge work and multimodal understanding tests shown.
Independent evaluation from Artificial Analysis broadly agrees on intelligence but finds weaknesses elsewhere. At the default high level the index score falls to 51, a touch behind GPT-6.1 Sol at xhigh and Claude Opus 5.5 at high, each at 52. Hydrogen rarely fabricates: on AA-Omniscience it makes things up 8 percent of the time. It also outperforms other Gemini models, Carbon included, on presentation quality in AA-Briefcase, which measures agentic knowledge work.
Where Hydrogen struggles is agentic terminal work. It scores 54 percent on Terminal-Bench 4.0 and just 50 percent on Terminal-Bench-Science 0.1, against 64 and 53 percent for Claude Sonnet 5.5 at max and 56 and 58 percent for GPT-6.1 Sol at max. For developers running shell-heavy autonomous agents, that gap is the number worth weighing before switching.
Cost per task is where the thinking level really bites. Artificial Analysis puts Hydrogen at about $3.16 per task at xhigh — roughly a third of what Claude Sonnet 5.5 costs at max for nearly the same intelligence, but about 2.5 times as much as GPT-6.1 Sol at max. Dropping to high cuts the average to $1.87, medium to $0.80 and low to $0.29, but intelligence scores fall steeply: medium reaches only 41 on the index and low just 27. The price jump at xhigh comes from using nearly twice the output tokens as high. At the high, medium and low tiers, Artificial Analysis places Hydrogen on the Pareto frontier, meaning no other model is both cheaper and smarter at those price points.
Google is also changing how its coding tool routes work. In Antigravity, a new orchestrator switches dynamically between Carbon and Hydrogen within a session rather than locking onto one model. On Terminal-Bench 4.0, Google says the orchestrator nearly matches Carbon's success rates while using 24 percent fewer resources and cutting processing time by 19 percent. If that holds, it is a template for mixing a cheap fast model with an expensive strong one rather than picking a single default.
In the Gemini app, Google says Hydrogen adheres more closely to tone, format and instructions over long conversations, so drafts, summaries and documents are more often usable first time. It can draw context from Google apps such as Gmail and Drive plus third-party services including Airtable and Dropbox. Google also calls it a significant upgrade for Search, handling typos, multi-part questions and very specific requests better, with more visual results and an improved query fan-out technique that pulls in a wider set of relevant links.
Why it matters: Developers choosing a default model for coding agents and knowledge-work pipelines now face a cheaper option that beats Google's own bigger Argon on published intelligence tests, though it trails Claude Sonnet 5.5 and GPT-6.1 Sol on terminal tasks. The four thinking levels mean the same model spans roughly $0.29 to $3.16 per task, so tier choice matters as much as model choice, and the xhigh tier is the only one that matches the headline scores. Teams already committed to Argon should re-benchmark on their own workloads, and the Antigravity orchestrator hints that mixing models by task may beat standardizing on one.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Google's Gemini 4 Hydrogen replaces the Flash series, but benchmarks suggest it also makes the pricier Argon obsolete. On the independent Artificial Analysis Intelligence Index, Hydrogen nearly matches Claude Sonnet 5.5 for less money but costs about 2.5 times more per task than GPT-6.1 Sol and still falls short on agentic coding. The article Google's Gemini 4 Hydrogen outscores the bigger Argon just four weeks after its launch appeared first on The Decoder .