AI News

Google Releases Gemini 4 Hydrogen, 2.5x Faster and Focused on Low Cost

36Kr · AIUpdated Nov 13 12:37

AI summary

Google has released Gemini 4 Hydrogen, positioned as a fast-execution model for everyday coding and agent tasks, competing with Claude Sonnet 5.5 and GPT-6.1 Terra. Google says it scores 13.5% on SWE-sweep, 71.6% on Agents' Last Exam v1.1, and 76.1% on AutomationBench-AA, all higher than the two competitors; it trails on OSWorld 2.1 and Terminal-Bench-Science 0.1. Artificial Analysis data shows its top-tier score is 54, close to Sonnet 5.5's 55, with an average cost of $3.16 per task, about one-third of Sonnet 5.5's $10.02. The model is being gradually rolled out to the Gemini app, Google Search, AI Studio, and the Gemini API.

Why read this

SelectedAI score 86

The original text provides multiple comparison results and per-task costs for Hydrogen against Sonnet 5.5 and GPT-6.1 Terra, which can be used to judge the selection trade-offs for mid-tier models.

Topics

Tags

Body · AI translation

Just now, Google released Gemini 4 Hydrogen. This new model, which focuses on speed and price, is taking aim at the mid-tier models from Claude and GPT.

For a large number of everyday coding and agent tasks, users want a model that can solve complex problems, but they also want it to run fast and cost reasonably. Google has positioned Hydrogen as a fast-execution model, pitting it against Claude Sonnet 5.5 and GPT-6.1 Terra.

According to the results Google published, Hydrogen outperforms both rivals on multiple tests including coding and knowledge work. On the third-party Artificial Analysis composite index, its average cost per task under the same evaluation is less than one-third that of Sonnet 5.5, yet it achieves nearly the same score.

Head-on clash with mid-tier workhorses

Benchmarks close to Sonnet 5.5, with cost cut by two-thirds.

Google positions Hydrogen as a fast-execution model for everyday coding and agent work. It can handle multi-step tasks independently and can also work alongside frontier models such as Carbon.

Google published a set of results covering coding, knowledge work, computer operation, and multimodal understanding. Among them, Hydrogen beats Claude Sonnet 5.5 and GPT-6.1 Terra on multiple items. And these two models are precisely the mid-tier workhorses of these two leading vendors, handling a large volume of users' everyday tasks.

On SWE-sweep, which involves finding and fixing coding problems, Hydrogen scored 13.5%, higher than 6.1 Terra's 5.8% and Sonnet 5.5's 3.1%. On Agents' Last Exam v1.1, which measures end-to-end professional workflows, Hydrogen scored 71.6%, exceeding 6.1 Terra's 69.8% and Sonnet 5.5's 60.4%; it also leads on the MKA test involving complex visual documents.

Beyond coding, Google also provided results oriented toward real professional work. On the AutomationBench-AA test, Hydrogen scored 76.1%, while Sonnet 5.5 and GPT-6.1 Terra scored 71.8% and 60.2%, respectively. This test focuses on whether a model can use software tools to complete automated workflows.

On the multimodal side, in the MKA test requiring handling of complex visual materials, Hydrogen scored 53.4%, Sonnet 5.5 scored 52.0%, and GPT-6.1 Terra scored 43.8%. Google hopes it can synthesize materials such as PDFs, spreadsheets, charts, audio, and video to produce reports and other work outputs.

The leaderboard also includes items where rivals lead. On the computer-operation benchmark OSWorld 2.1, it trails Sonnet 5.5 slightly. On the science task test Terminal-Bench-Science 0.1, it trails both Sonnet 5.5 and GPT-6.1 Terra.

Overall, these results demonstrate Hydrogen's ability to match the mid-tier models of the two leading companies.

Google also emphasized its speed. According to Google, at the default high thinking level, Hydrogen completes tasks 2.5 times faster than Gemini 3.8 Flash. For developers who need to repeatedly modify code and check results, shorter wait times also affect the experience of the entire workflow.

Following Gemini convention, Hydrogen also offers four thinking levels—low, medium, high (default), and a new extra-high (xhigh)—providing tiered choices among quality, latency, and cost, scaling gradually in a smooth and predictable way, and paying time and cost for a more reasonable depth of reasoning.

Third-party evaluations provide a cross-reference. Artificial Analysis's Intelligence Index aggregates tests in knowledge work, coding, science, and more. When checked on November 13, Hydrogen's highest xhigh level scored 54, close to Sonnet 5.5's highest max level of 55, and higher than GPT-6.1 Terra's highest max level of 51.

Beyond capability, there is also a price advantage. The AA leaderboard also provides the average cost per task. As can be seen, GPT-6.1 Terra is cheaper, but its top score is not as high as Hydrogen's. And in the comparison with Sonnet, at the highest reasoning level, the gap is obvious. They are $3.16 and $10.02, respectively, and Hydrogen's price is only one-third of Sonnet 5.5's.

Code and physics engine in one shot

Google demonstrated a Minecraft-like game called Geminisland. With a single prompt from the user, Hydrogen can fully generate the game through instant 3D programming. The island has roads, buildings, and large animals, and users can also walk into the scene in first person to explore.

Netizen kimmonismus also used Hydrogen to make a canyon flying game. The author said it uses Three.js, procedural terrain, and dynamic water surfaces, and set up 25 scoring gates.

In graphics tasks requiring understanding of physical rules and real-time rendering, Hydrogen also performs steadily. Hydrogen can create 3D interactive models according to specific visual requirements. In a water-and-glass visualization program, the model simulates refraction, reflection, dispersion, and absorption as light passes through water and glass.

This case turns complex visual requirements into real-time imagery, demonstrating the model's ability to write graphics code, understand physical rules, and build complete systems.

Likewise, netizens have also done many 3D hands-on tests. For example, developer Tony used Hydrogen and Google's newly released flagship Carbon, respectively, to create a 3D scene of a boat traveling between glaciers.

Hydrogen took 1 hour 23 minutes 40 seconds and cost $4.27; Carbon took 2 hours 39 minutes 10 seconds and cost $8.83. With both time and cost halved, Hydrogen still maintained extremely high presentation quality. It did not significantly lag behind its own flagship.

Multi-step video editing and dynamic SVG

In another case, Google directly had the AI complete the full production of a video. Following the creative requirements, Hydrogen used a dedicated video assembly skill to find public materials, cut the materials into a continuous video, and add narration.

This case demonstrates the model's ability to call tools to complete multi-step creation according to requirements. It shows the model's multimodal capabilities and aesthetic level, as well as its ability to handle a complete video production workflow.

Also around visual effects, netizen Marcel made an advanced version of the pelican test. He had the model generate an SVG animation of a pelican riding a bicycle through five dynamic seasons. And he compared Hydrogen with Anthropic's strongest Fable 5.2.

Fable 5.2 took about 25 minutes, while Hydrogen took only 5 minutes. With a 5x time difference, Hydrogen also delivered quite good presentation results.

These tasks demonstrate Hydrogen's capabilities across multiple areas, from coding to 3D model creation to multimodality and aesthetics. Google's official release also includes other cases showing website creation, document editing, data analysis, and many other scenarios. Hydrogen completed the tasks excellently in all of them.

Hands-on and selection advice

From 3D interactive scenes and physical graphics rendering to automated video editing, Gemini 4 Hydrogen proves one thing: in high-frequency engineering scenarios, models no longer need to stack flagship-level parameters to gain usability.

As everyday coding and multi-step Agent automation become pay-per-use consumables, the combination of "close to the top tier's ceiling + 2.5x speed + one-third the cost" has more practical value than simply ranking first on a leaderboard. If your daily workflow relies heavily on Claude Sonnet 5.5 or GPT-6.1 Terra, Hydrogen already has the strength to be a direct drop-in replacement.

Currently, Google has begun gradually rolling out Gemini 4 Hydrogen to the Gemini app, Google Search, AI Studio, and the Gemini API, and free users can also experience it in the Gemini app through automatic model selection.

This article comes from the WeChat public account "AI Frontline" (ID: ai-front), authors: Feiye, Siyue, editor: Siyue, published by 36Kr with authorization.

Source

36Kr · AI