Google's Gemini 4 Hydrogen replaces Flash and edges past Argon on a key benchmark
Google's new fast model for coding and agents beats Gemini 4 Argon on one independent index, but costs twice Flash's current introductory API rate.
Google released Gemini 4 Hydrogen on Thursday, a fast model for coding and multi-step agent work that replaces Gemini 3.8 Flash as Google's fast everyday model across its developer tools, the Gemini app and Search.
The pitch is that teams no longer have to choose between a cheap, quick model and a smart one for routine work. At its default "high" reasoning setting, Hydrogen scores 51 on the Artificial Analysis Intelligence Index, one point above Google's own larger Gemini 4 Argon, and 54 at a new "xhigh" setting. According to VentureBeat, the same index put Gemini 3.8 Flash at 31.
What Hydrogen is built to do
Google positions Hydrogen below its two frontier models, Gemini 4 Carbon and Gemini 4 Argon, which it reserves for long, deep reasoning jobs. Hydrogen is meant for the execution layer: writing code, running tests, reading documents and calling tools quickly. Google says it finishes tasks 2.5 times faster than 3.8 Flash at its default setting, more than triples Flash's score on the Terminal-Bench 4.1 agent benchmark (57% versus 18.2%) and cuts fabricated claims by more than five times. Those are the company's own figures. VentureBeat noted that Google's launch table still shows Anthropic's Claude Sonnet 5.5 ahead on Terminal-Bench 4.1, at 63.9%.
The model offers four reasoning levels, low, medium, high and xhigh, so developers can trade answer quality against latency and cost per request. Like Carbon and Argon, it has a 1 million-token context window, and Google's migration notes say that window is now shared across input, reasoning and output tokens, with a 65,536-token cap on a single request. Developers moving from Flash are told to change the model ID to gemini-4-hydrogen and to raise any hard-coded low output limits.
Google is also using Hydrogen as a building block for multi-model agents. A new Orchestrator feature in its Antigravity coding tool switches between Carbon and Hydrogen inside one session. On Terminal-Bench 4.0, Google says that setup came close to Carbon's success rate while using 24% less quota and finishing 19% faster. In parallel-agent setups, Carbon can coordinate while Hydrogen runs the worker and verifier agents.
The price question
On the paid Gemini API tier, Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens, flat across the full context window, with cached input at $0.075 per million, according to Google's pricing page. It is also available on the free, rate-limited tier.
That is twice what 3.8 Flash costs today. Google lists Flash at an introductory $0.75 input and $3.75 output through Dec. 31, rising to the same $1.50 and $7.50 from Jan. 1, 2027. Argon, the larger model Hydrogen narrowly beats on the index, is priced at $2 and $10 through Nov. 30 and $4 and $20 from Dec. 1.
Against rival small models, Hydrogen is far more expensive per token. VentureBeat calculated that its rates are 15 times the lowest standard rates for Anthropic's Claude Haiku 5.5, OpenAI's GPT-6 Luna and Meta's Muse Ember, though Hydrogen outscored all three on the index even at its medium setting. Artificial Analysis estimates Hydrogen's cost at $1.87 per index task at high and $3.16 at xhigh, versus $7.77 for 3.8 Flash at high. The firm's figures also show the trade-off at the top setting: xhigh used far more output tokens and took 13.7 minutes per task on average, against 7.9 minutes at high.
Among open-weight models, Moonshot's Kimi K3.1 Pro comes closest, scoring 46 on the same index at an estimated $3.68 per task, according to the VentureBeat comparison. Xiaomi's hosted MiMo-V2.6-Flash costs about $0.13 per task, with a score 24 points below Hydrogen's.
Search and early customers
For consumers, the bigger change may be in Search. Google calls Hydrogen its "biggest model improvement to Search yet," TechCrunch reported, with the update aimed at multi-part questions, typos and highly specific requests, and it is rolling out to all Search users globally. Google also says it has widened the "query fan-out" technique Search uses to pull in a broader set of links.
One early enterprise data point comes from Box. VentureBeat reported that Box CEO Aaron Levie posted a chart showing Box's agent rising from 74% to 79% accuracy on the company's Complex Work Eval when it switched from Flash to Hydrogen, and said a typical task took less than half as long and used about 12% fewer tokens.
The release follows a busy few weeks for Google DeepMind, which shipped the Genie 4 world model in late October, and comes days after Anthropic cut prices on its Claude Fable and Mythos 5.2 models. Hydrogen is rolling out now in AI Studio, the Gemini API, Gemini for business and the Gemini app, and Google's deprecation schedule lists no shutdown date yet for 3.8 Flash.
