What changes in the Gemini offering
Google has announced Gemini 4 Hydrogen, a general-purpose artificial intelligence model designed for developers and ordinary users, intended to replace the Flash series. Unlike the frontier models Gemini 4 Carbon and Gemini 4 Argon, designed for long and complex reasoning, Hydrogen aims to perform everyday tasks quickly and economically, according to The Decoder.
Declared performance and independent evaluations
The main novelty concerns positioning: Google claims that Hydrogen beats the larger Argon on the Artificial Analysis Intelligence Index at the default “high” thinking level, with a score of 51 versus 50, and reaches 54 at the maximum “xhigh” level. Hydrogen offers four thinking levels (low, medium, high and the new xhigh), while Argon provides only two. The model has a context window of one million input tokens and can generate as many in output.
On pricing, Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens, with cache hits at $0.075. According to the account, it would therefore be both more capable and less expensive than Argon, effectively rendering obsolete the advice Google had given just two days earlier, at the launch of Carbon, to continue using Argon for daily coding work.
Google presents Hydrogen as a major upgrade over Gemini 3.8 Flash: at the default level the model would complete tasks 2.5 times faster. On Terminal-Bench 4.1 the score would rise from 18.2% for Flash to 57% for Hydrogen, while hallucinations would be reduced fivefold and the model would recover better if a tool call fails or a search returns no results. The company also cites first place on the Vals Tax Agent Bench with 52.3% and a tied first place on the Vals Legal Research Bench with 55.8%.
Artificial Analysis’s independent evaluation places Hydrogen at roughly the same level as Claude Sonnet 5.5 at the maximum thinking level. At the “high” level the score drops to 51, slightly behind GPT-6.1 Sol (xhigh, 52) and Claude Opus 5.5 (high, 52). On AA-Omniscience the model would make up information in only 8% of cases, and it would outperform other Gemini models, including Carbon, in presentation quality as measured by AA-Briefcase.
The weak points emerge in agentic work on the terminal: Hydrogen scores 54% on Terminal-Bench 4.0 and just 50% on Terminal-Bench-Science 0.1, versus 64% and 53% for Claude Sonnet 5.5 (max) and 56% and 58% for GPT-6.1 Sol (max). On value for money, Artificial Analysis estimates about $3.16 per task at xhigh, about one-third the cost of Claude Sonnet 5.5 (max) at similar intelligence, but about 2.5 times the cost of GPT-6.1 Sol (max). At “high” the cost drops to $1.87, at “medium” to $0.80 and at “low” to $0.29, although intelligence scores fall sharply (41 and 27).
Availability and open questions
Google is rolling out Hydrogen globally across all platforms: developers access it via the Gemini API in AI Studio and Android Studio, Pro and Ultra subscribers via Google Antigravity, business customers via Gemini for Business, and all users via the Gemini app and Google Search. In the app, the model is said to show greater adherence to tone, format and instructions in long conversations and can retrieve context from Google apps such as Gmail and Drive and from third-party services such as Airtable and Dropbox. In Antigravity, an orchestrator has been introduced that dynamically switches between Carbon and Hydrogen during a session; on Terminal-Bench 4.0 Google claims it approaches Carbon’s success rates using 24% fewer resources and reducing time by 19%.
Some elements remain to be verified. The better performance compared with Argon and Flash is largely Google’s own claims, while the independent evaluations cited come from Artificial Analysis and do not cover all the benchmarks mentioned by the company. Moreover, the cost picture is nuanced: Hydrogen is inexpensive compared with Sonnet 5.5 but more expensive than GPT-6.1 Sol per task, and the lower thinking levels significantly reduce the model’s intelligence. Finally, it is not clear whether and when Argon will be retired or repositioned, nor how much the Antigravity orchestrator affects overall performance in real-world scenarios.
