Gemini 4 Hydrogen has emerged as a credible challenge to Google’s larger frontier models, outscoring Gemini 4 Argon on the Artificial Analysis Intelligence Index at Google’s default high thinking level. Google released the fast general-purpose model on November 12, 2026, positioning it for developers and everyday users rather than the deepest, longest-horizon reasoning tasks assigned to Carbon and Argon.
Gemini 4 Hydrogen: Speed and price challenge Google’s frontier tier
According to Google, Hydrogen scored 51 against Argon’s 50 on the Artificial Analysis Intelligence Index at the default high setting. Its new xhigh setting raised the score to 54. Google Cloud’s public SKU catalogue listed the model before launch at 1.50 dollars per million input tokens and 7.50 dollars per million output tokens for text, and those rates matched Google’s pricing page after release.
FourWeekMBA calculated that 1 million input tokens plus 1 million output tokens would cost 9 dollars on Hydrogen. That is 37.5 percent of the 24 dollars charged for Gemini 4 Carbon and Gemini 4 Argon from December 1. The Hydrogen rates also match the prices Google lists for Gemini 3.8 Flash starting January 1, 2027, giving the new model a position between the current Flash offer and Google’s more expensive frontier tier.
Cache hits cost 0.075 dollars. The model id is gemini-4-hydrogen, and its listed context window is 1,048,576 tokens. Hydrogen can accept 1 million input tokens and output as many as 1 million tokens, a combination The Decoder said supports work involving large PDFs, spreadsheets and videos.
Google claims stronger agentic performance
Google presents Hydrogen as both a standalone model and a major upgrade over Gemini 3.8 Flash, which arrived in September. At the default setting, Google says Hydrogen completes tasks 2.5 times faster. On Terminal-Bench 4.1, a command-line work test, its score reached 57 percent, more than tripling Gemini 3.8 Flash’s 18.2 percent.
Google also claims hallucinations fell by a factor of five and that Hydrogen recovers better when a tool call fails or a search returns nothing. Artificial Analysis found that it made things up only 8 percent of the time on AA-Omniscience. The model scored 53.4 percent on Google’s Multimodal Knowledge Agent Benchmark and 71.6 percent on Agents’ Last Exam v1.1.
On open-ended agent work involving research, data analysis and code navigation, Hydrogen ranks fifth on the Agent Arena leaderboard. Google says it took first place on the Vals Tax Agent Bench with 52.3 percent and tied for first on the Vals Legal Research Bench with 55.8 percent. The Decoder reported that the Gemini app also follows tone, format and instructions more closely over long conversations, while Hydrogen can draw context from Gmail, Drive, Airtable and Dropbox.
Cost advantage depends on the thinking level
Artificial Analysis placed Hydrogen at about 3.16 dollars per task when using xhigh. That is roughly one-third the cost of Claude Sonnet 5.5 at maximum, which reaches a similar intelligence level, but about 2.5 times the cost of GPT-6.1 Sol at maximum. At high, Artificial Analysis puts the cost at 1.87 dollars per task, falling to 0.80 dollars at medium and 0.29 dollars at low.
Those savings come with lower intelligence scores at the cheaper settings. The index score falls to 41 at medium and 27 at low. Artificial Analysis says xhigh uses nearly twice as many output tokens as high, which explains the price jump. At high, medium and low, Hydrogen sits on the Pareto frontier, meaning no compared model is both cheaper and smarter at those price points.
At xhigh, Hydrogen’s score of 54 was just behind Claude Sonnet 5.5 and GPT-6.1 Sol, which each scored 55 at maximum thinking. At the default high level, Hydrogen’s 51 trailed GPT-6.1 Sol and Claude Opus 5.5, both at 52. Artificial Analysis assessed Hydrogen as roughly on par with Claude Sonnet 5.5 at its highest thinking level.
The model is not uniformly ahead in specialist terminal work. It scored 54 percent on Terminal-Bench 4.0 and 50 percent on Terminal-Bench-Science 0.1. Claude Sonnet 5.5 at maximum reached 64 percent and 53 percent respectively, while GPT-6.1 Sol at maximum scored 56 percent and 58 percent.
Immediate access changes the competitive picture
Distribution is another sharp point. SiliconANGLE reported that Hydrogen skipped the staged rollout used for Argon, which spent nearly two weeks with members of Google’s Fairwind Program, a group of vetted cybersecurity defenders, before paying developers gained access on October 13. Hydrogen was available to anyone from launch day, including users of the free Gemini app.
Koray Kavukcuoglu, who became head of Google DeepMind in August when Demis Hassabis moved into a chairman role, wrote in the announcement that Hydrogen was built for rapid execution. Three Gemini 4 models have shipped in just over six weeks, with Hydrogen arriving two days after Carbon. The Decoder noted that Google had recently recommended Argon for daily coding work, advice Hydrogen appears designed to make less relevant.
Google’s coding environment also illustrates how the models may be used together. In Antigravity, an orchestrator dynamically switches between Carbon and Hydrogen within a session. Google says the approach nearly matched Carbon’s Terminal-Bench 4.0 success rate while using 24 percent fewer resources and reducing processing time by 19 percent.
The central competitive point is not that Gemini 4 Hydrogen replaces every frontier model. Artificial Analysis noted that the xhigh configuration is slower than average, taking 13.5 minutes per Intelligence Index task against a 13.0-minute median, and described it as verbose. Google has nevertheless put a capable, broadly available model into a price tier that could reset expectations for developers choosing between speed, reasoning depth and cost. For those evaluating Gemini 4 Hydrogen, the evidence points to a model whose practical advantage lies in combining high intelligence, immediate access and lower unit costs, even though Argon and Carbon retain specialist strength in deeper reasoning and terminal benchmarks.