Hydrogen
A training codename for a Google AI model; specific details have not been disclosed.
Core Facts
To use Gemini 4 Hydrogen, you only need to specify the model name `google/gemini-4-hydrogen` in the request.
Migrating from Gemini 3.8 Flash to Hydrogen, at the default high setting, task completion speed increases 2.5 times, the Terminal-Bench 4.1 score rises from 18.2% to 57%, and fabricated claims decrease by more than 5 times.
On the Artificial Analysis Intelligence Index, Gemini 4 Hydrogen scores 51 at the default high thinking level and can reach 54 at the highest xhigh thinking setting.
Timeline (last 90 days)
Before the official architecture documentation is complete, developers can do an initial triage by task complexity: routine high-frequency tasks should favor Argon first, complex tasks requiring a stronger reasoning chain should turn to Carbon, and lightweight scenarios can be evaluated with Hydrogen.
A YouTube creator offered a rough classification framework: Argon corresponds to the Opus level, Carbon to the Fable or Astra level, and Hydrogen to the Sonnet or even Haiku level.
According to the analysis in the video, Argon is regarded as a frontier model available at any time, Carbon leans toward additional reasoning and higher-complexity tasks, and Hydrogen is positioned as more lightweight.
Google's models are named after the periodic table; the flagship is Carbon, and Hydrogen, the lightest element, corresponds to the speed and price positioning.
Developer Tony's test showed that Hydrogen took 1 hour 23 minutes and cost 4.27 dollars to complete a 3D scene of a ship passing through glaciers, while Carbon took 2 hours 39 minutes and cost 8.83 dollars.
By Artificial Analysis per-task cost, Sonic's highest tier is 10.2 dollars per task and Hydrogen's highest tier is 3.16 dollars per task; Tera is cheaper than Hydrogen but has a lower top score.
In the OS World 2.1 test, Hydrogen was slightly worse than Sonic.
In the MKA visual file test, Hydrogen scored about 53%, Sonic 52%, and Tera about 44%.
In the Automation Bench test, Hydrogen scored about 76%, Sonic 72%, and Tera 60%.
In the Agent's Last Exam test, Hydrogen scored close to 72%, Tera slightly below 70%, and Sonic about 60%.
40 more timeline events
All Facts (20)
To use Gemini 4 Hydrogen, you only need to specify the model name `google/gemini-4-hydrogen` in the request.
90%VerifiedMigrating from Gemini 3.8 Flash to Hydrogen, at the default high setting, task completion speed increases 2.5 times, the Terminal-Bench 4.1 score rises from 18.2% to 57%, and fabricated claims decrease by more than 5 times.
90%VerifiedOn the Artificial Analysis Intelligence Index, Gemini 4 Hydrogen scores 51 at the default high thinking level and can reach 54 at the highest xhigh thinking setting.
85%Partially VerifiedGoogle has consecutively released two Gemini 4 models: the lightweight Hydrogen and the flagship reasoning model Carbon.
90%Partially VerifiedGemini 4 Hydrogen has a 57% pass rate on Terminal-Bench 4.1.
75%Partially VerifiedHydrogen's output token pricing is $1.50 per million and input token pricing is $7.50 per million, held uniform within a 1-million-token context window, and the API offers a free tier.
70%Partially VerifiedHydrogen is priced at $1.5 input and $7.5 output; Carbon and Argon are $4 and $20 respectively after the promotional period (units unspecified).
70%Partially VerifiedHydrogen has a context window of 1 million tokens.
65%Partially VerifiedThe Gemini 4 series includes Hydrogen, Carbon and Argon, with the division of labor as follows: hard problems go to Carbon, complex reasoning uses Argon, and everyday execution goes to Hydrogen.
65%Partially VerifiedGoogle introduced an Orchestrator in Antigravity that can dynamically switch between models such as Carbon and Hydrogen within the same session.
75%Partially VerifiedHydrogen replaces Gemini 3.8 Flash and launched on day one in AI Studio, the API, the Gemini app, Antigravity (as the default agent model), and Google Search.
65%UnverifiedOn the same Terminal Bench, Claude Sonnet 5.5 scores 63.9, ahead of Hydrogen's 57%; on computer use tasks Sonnet wins 83.9 to 77.6.
60%UnverifiedHydrogen is derived from Carbon and is designed specifically for fast execution in agentic work.
60%UnverifiedHydrogen scored a perfect 43/43 on the standardized Plotly evaluation set, with accuracy on par with frontier models.
60%UnverifiedGemini 4 Hydrogen is currently open only to paying users, and users on free quotas or free plans may not see this model in Anti-Gravity.
60%UnverifiedOn November 12, Google released Hydrogen (internal codename), the model in the Gemini 4 family focused on speed and cost.
60%UnverifiedAccording to the analysis in the video, Argon is regarded as a frontier model available at any time, Carbon leans toward additional reasoning and higher-complexity tasks, and Hydrogen is positioned as more lightweight.
50%UnverifiedA YouTube creator offered a rough classification framework: Argon corresponds to the Opus level, Carbon to the Fable or Astra level, and Hydrogen to the Sonnet or even Haiku level.
50%UnverifiedBefore the official architecture documentation is complete, developers can do an initial triage by task complexity: routine high-frequency tasks should favor Argon first, complex tasks requiring a stronger reasoning chain should turn to Carbon, and lightweight scenarios can be evaluated with Hydrogen.
50%UnverifiedIn the Suesway code test, Hydrogen scored 13.5, Tera scored 8, and Sonic scored 3.
50%