Google releases Gemini 4 Hydrogen to replace Flash for coding and agents

A new standard for daily tasks

Google has officially launched Gemini 4 Hydrogen, a general-purpose AI model designed for everyday workflows. While Google's frontier models like Carbon and Argon are built for deep, complex reasoning, Hydrogen is intended to be the fast and affordable workhorse. It is positioned as a successor to the Gemini 3.8 Flash series, aiming to provide a significant leap in performance for common user needs.

The model is already being rolled out globally across various platforms. Users can access it through the Gemini app and Google Search, while business customers can find it in Gemini for Business. For developers, the model is available via the Gemini API in both AI Studio and Android Studio. This wide availability ensures that both casual users and power users can interact with the new technology immediately.

One of the most notable changes is how the model handles user instructions. Google reports that Hydrogen sticks more closely to specific tones, formats, and instructions during long conversations. This improvement is intended to make drafts, summaries, and documents more usable on the first try, reducing the need for repeated prompting by the user.

Outperforming its larger siblings

Despite being a smaller model than Argon, Hydrogen has already begun to outperform it in key benchmarks. As reported by The Decoder, Hydrogen scored 51 on the Artificial Analysis Intelligence Index at the default high thinking level, slightly beating Argon's score of 50. At the new xhigh tier, Hydrogen reaches a score of 54, providing a higher ceiling for complex reasoning than the larger model offers.

The model also introduces a tiered thinking system that allows users to choose the level of intelligence required for a specific task. Hydrogen offers four levels—low, medium, high, and xhigh—whereas Argon only provides medium and high options. This flexibility allows the model to be tuned for speed on simple queries or deeper analysis on difficult problems, making it a more versatile tool for diverse applications.

Furthermore, Hydrogen is significantly cheaper than its larger counterparts. It costs $1.50 per million input tokens and $7.50 per million output tokens, with cache hits costing just $0.075. By offering higher intelligence at a lower price point, Google is effectively making Argon obsolete for many daily coding and agentic tasks just four weeks after its own release.

Technical leaps in coding and agents

For developers, Hydrogen represents a major upgrade over the previous Flash generation. On the Terminal-Bench 4.1, the model's score jumped from 18.2 percent for Flash to 57 percent for Hydrogen. Google also notes that hallucinations have dropped by a factor of five. The model is also better at recovering when a tool call fails or a search query returns no results, which is critical for building reliable autonomous agents.

The model excels in specific agentic benchmarks, such as the Vals Tax Agent Bench where it scored 52.3 percent. It also tied for first place on the Vals Legal Research Bench with a score of 55.8 percent. These results suggest that Hydrogen is highly capable of navigating specific professional domains like law and finance. It also ranks fifth on the Agent Arena leaderboard for open-ended tasks like research and data analysis.

In terms of multimodal capabilities, Hydrogen can process up to one million tokens of input and output. This allows it to handle massive amounts of data, such as summarizing large PDFs, spreadsheets, or long videos. It scored 53.4 percent on Google's new Multimodal Knowledge Agent Benchmark, highlighting its ability to synthesize information from various types of media into a cohesive summary.

Integration and resource efficiency

Hydrogen is being integrated into several core Google products to improve the user experience. In Google Search, the model handles complex queries involving typos or multi-part questions more effectively. It uses an improved query fan-out technique to pull in a wider range of relevant links, resulting in search results that are more visual and easier for users to scan quickly.

For coding specifically, Google's Antigravity tool now features a new orchestrator. This system dynamically switches between the Carbon and Hydrogen models within a single session depending on the complexity of the task. According to Google, this approach nearly matches the success rates of the Carbon model while using 24 percent fewer resources and cutting processing time by 19 percent.

The model also maintains deep integration with the Google ecosystem. It can pull context from various Google apps, including Gmail and Drive, as well as third-party services like Airtable and Dropbox. This connectivity allows Hydrogen to act as a more cohesive assistant that can access and organize information across the different platforms that users rely on for their daily work.

Independent performance and pricing

Independent testing by Artificial Analysis provides a more nuanced view of how Hydrogen compares to other leading models. At its highest thinking level, Hydrogen is roughly on par with Claude Sonnet 5.5. However, at the default high level, it sits slightly behind GPT-6.1 Sol and Claude Opus 5.5. These comparisons help put the model's capabilities into perspective against the current state of the AI market.

The pricing structure of Hydrogen places it in a unique position on the price-performance frontier. At the high level, it costs about $1.87 per task, while the medium level drops to $0.80 and the low level costs just $0.29. Artificial Analysis notes that at these three levels, no other model is currently both cheaper and smarter than Hydrogen at those specific price points.

One caveat noted by independent testers is that Hydrogen can struggle with specific terminal work compared to its rivals. On the Terminal-Bench 4.0, it scored 54 percent, while Claude Sonnet 5.5 (max) reached a higher score. This suggests that while Hydrogen is a powerhouse for general knowledge and multimodal tasks, it may still lag behind specialized models for certain high-level programming and terminal commands.