Google Launches Gemini 4 Hydrogen, the Successor to 3.8 Flash: Faster, Ahead of Argon, but Twice as Expensive as the Current Flash

Following Argon and Carbon, which were reserved for demanding paid use cases, Google is rounding out its Gemini 4 family from the bottom up. Gemini 4 Hydrogen, announced on November 12, is the lineup’s fast model, designed for everyday coding and agents. It is also the one bringing Gemini 4 to everyone: the API’s free tier, the Gemini app, and Google Search.

What is available, and for whom

The model (gemini-4-hydrogen) is accessible immediately in the Gemini API and Google AI Studio, including the free tier (subject to rate limits), in Gemini for Enterprise, in the Gemini app’s model picker, and in Google Search. Free users of the app access it via automatic model selection. In Antigravity, Google’s agentic development environment, it is reserved for paid Google AI Pro and Ultra subscribers and becomes the agent’s default model.

Technically, Hydrogen shares its predecessors’ architecture: a unified 1-million-token context window and long-form generation supporting up to 1 million output tokens. It introduces a fourth thinking level, xhigh, above low, medium, and high, with the latter serving as the default setting (Carbon and Argon top out at high and start at medium). A synchronous call can return up to 65,536 tokens—twice as many as an unstreamed REST call on Carbon or Argon, according to the documentation.

Pricing: what the migration actually costs

Google positions Hydrogen as the recommended replacement for Gemini 3.8 Flash. Its pricing sheet makes it possible to quantify what this changes:

Take a typical agent task that reads 100,000 tokens and outputs 20,000 (thinking included). It costs $0.30 with Hydrogen, compared to $0.15 with Gemini 3.8 Flash at current rates, $0.40 with Argon through the end of November, and $0.80 with Argon starting in December. In other words, leaving Flash today doubles the per-token bill for seven weeks; starting in January, both models will cost the same per token, making Hydrogen the obvious choice if its performance holds up. Compared to Argon, it is 25% cheaper now and 62.5% cheaper in December. One unknown remains: the volume of thinking tokens consumed at the default high level, which directly inflates output costs.

Performance: between in-house figures and the first external ranking

Google puts forward flattering numbers: Hydrogen reportedly surpasses Argon on the Artificial Analysis Intelligence Index (51 vs. 50, and 54 at the xhigh level), completes tasks 2.5 times faster than 3.8 Flash, triples its score on Terminal-Bench 4.1 (57% vs. 18.2%), and invents five times fewer claims. These are figures published by Google.

A first external signal has emerged: on the Agent Arena leaderboard by arena.ai, viewed on November 13, Gemini 4 Hydrogen (High) ranks 5th with a 14.80% success rate, ahead of Claude Fable 5.1 (Max) and GPT-6 Astra (Max), and well ahead of Gemini 4 Argon (High) at 10.72%. Carbon, in 3rd place, performs better (16.90%), while Claude Fable 5.2 (Max) dominates at 30.48%. The gap with Argon is noteworthy for a model meant to be cheaper, though this leaderboard is based on user sessions and measures neither reliability over extended tasks nor Hydrogen’s actual cost per task, which Arena does not yet display.

At the same time, Google is launching Orchestrator in Antigravity, which automatically alternates between Carbon (planning, difficult problems) and Hydrogen (execution, testing, parallel sub-agents). On Terminal-Bench 4.0, this pair reportedly approaches the success rate of Carbon alone while consuming 24% less quota, again according to Google.