When you hand an AI agent a task spanning dozens of files, you quickly hit an awkward fork in the road. Either you run a flagship like Gemini 4 Carbon or Argon on every step—paying $4 per million input tokens and $20 for output starting December 1—while it chews through simple edits. Or you pick fast Gemini 3.8 Flash, which scores only 18.2% on Terminal-Bench 4.1 and makes things up. On November 12, 2026, Google DeepMind released Gemini 4 Hydrogen to solve that tradeoff, working both solo and paired with a flagship.
Division of labor: how Carbon plans and Hydrogen executes
The Gemini 4 lineup splits roles clearly. Larger Carbon and Argon handle deep reasoning, architecture, and hard bugs. Adapted from Carbon, Gemini 4 Hydrogen (models/gemini-4-hydrogen) replaces 3.8 Flash as the fast agentic executor.
To avoid burning expensive quota, Google Antigravity added an orchestrator mode for paid Pro and Ultra plans. It works like a site foreman paired with a fast crew. First, Carbon or Argon inspects the repository in read mode, checks links between modules, and drafts a plan. Then Hydrogen makes targeted edits without rewriting neighboring files, batches parallel file reads, and runs tests. If the executor hits a wall, control returns to the planner. In multi-agent builds across isolated workspaces, Carbon coordinates the job while Hydrogen copies write and verify the code.
According to benchmarks on the Google Developers Blog, the duo matches pure Carbon's quality on Terminal-Bench 4.0 while using 24% less quota and finishing 19% faster. In partner tests on graphs and SQL, Hydrogen alone ran 2x to 2.8x faster than 3.8 Flash and used 12–23% fewer tokens.
Speed, a 1M-token context window, and the model's real ceiling
At its default high setting, the model solves tasks 2.5x faster than Gemini 3.8 Flash, according to the Google DeepMind blog. On Terminal-Bench 4.1, accuracy jumped from 18.2% to 57%, with more than five times fewer fabricated claims. Agent safety checks logged 81% fewer unauthorized actions and 12x fewer attempts to cut corners. On the Artificial Analysis Intelligence Index, the newcomer scores 51 at high (vs. 50 for Argon) and reaches 54 at xhigh.
The shared context window holds 1 million tokens—1,048,576 text chunks across input and output combined—enough for a large codebase or a batch of videos. The model reads text, images, audio, video, and PDFs, and outputs text only (no image or audio generation). Synchronous output goes up to 64,000 tokens, while background long generation scales up to 1 million.
Reasoning depth can be set to low, medium, high, or xhigh. API pricing stays flat across the entire million-token window with no volume surcharge: $1.50 per million input tokens and $7.50 for output, including reasoning tokens. Carbon keeps an introductory rate of $2 and $10 through November 30 before moving to $4 and $20 on December 1.
Its Model Card also lists weak spots. Hallucinations, slowdowns, and timeouts still happen, and high or xhigh noticeably inflate output bills. Knowledge cuts off in March 2026, and false refusals on multimodal prompts are higher than 3.8 Flash, though medical HealthBench Pro refusals stay under 0.4%. Also, the headline speed and benchmark records still come from Google's internal measurements.
Where to try it free and how to pick the right mode
You can run Hydrogen for free in Google AI Studio, on the Gemini API free tier, in Android Studio, Google Search, and via auto-selection in the Gemini app. Exact per-minute and daily limits for the free tier aren't listed in the public AI Studio table. The Antigravity orchestrator is gated behind Google AI Pro and Ultra plans, while the 30%+ faster model in Google Sheets is available on Pro, Ultra, and business plans.
For one-off tasks like a script to catch silent automation failures, a single-page mockup, or a brief from a folder of PDFs and audio, run Hydrogen on low or medium. For tricky debugging and large spreadsheet reports, bump it to high or xhigh. When a project spans dozens of files with architectural risks, turn on the orchestrator with Carbon. Even then, generated interfaces remain demo mockups with no live CRM or email hooks, so a human still needs to check the final code and facts.

