Source: Google DeepMind's launch post on LinkedIn (2026-11-12) and the Google blog, Introducing Gemini 4 Hydrogen (Koray Kavukcuoglu, 2026-11-12). All benchmark numbers below are vendor-reported.
TL;DR
Hydrogen is the fast, cheap member of the Gemini 4 family, positioned for everyday coding and agentic execution while Carbon and Argon handle deep reasoning. The interesting part for engineers is not one more model but the explicit split: a planner-grade model and an executor-grade model in the same family, with Google's Antigravity orchestrator switching between them inside one session.
What shipped
Thinking levels: low, medium, high (default) and xhigh, set per request.
Pricing: $1.50 per million input tokens and $7.50 per million output tokens, flat across a 1M-token context.
Reported results: Artificial Analysis Index 51 at high (Argon is reported at 50) and 54 at xhigh; Terminal-Bench 4.1 57% versus 18.2% for the previous fast model; HLE-Diamond 58.8%; first on Vals Tax Agent Bench (52.3%) and tied first on Vals Legal Research (55.8%); Agents' Last Exam 71.6%.
Claimed improvements: cleaner code edits and updated safeguards.
Availability: rolling out in AI Studio and the Gemini API, Android Studio, Gemini Enterprise, Google Sheets, Search, the Gemini app, and Antigravity for paid subscribers. A model card is linked from the post.
Architecture pattern: planner and executor tiers
The launch frames a two-tier deployment. A reasoning model plans, decomposes and reviews; a fast model executes the many small steps (edits, tool calls, test runs). The orchestrator routes by step type. This is the same split many harnesses already build by hand, now offered inside one vendor's family with a shared tokenizer, tool format and safety layer, which removes some cross-model friction.
Thinking levels add a second dial inside the executor tier: the same model can spend more test-time compute on a hard step without switching models.
Tradeoffs
Benchmark parity with the reasoning tier is surprising. An executor scoring at or above Argon on a composite index suggests either the index rewards speed-friendly tasks or Hydrogen at xhigh is no longer cheap. Measure cost per solved task, not price per token.
Routing errors are silent. If the orchestrator sends a step that needed deep reasoning to the fast tier, the failure looks like a model mistake, not a routing one. Log which tier and level handled each step.
Flat long-context pricing makes it tempting to stuff 1M tokens per call; latency and attention quality at that length still need testing.
Single-vendor lock-in for the planner/executor pairing; a multi-vendor harness keeps the option to swap either tier.
Review checklist
Which steps in your agent loop are execution-grade, and have you measured them on the fast tier at each thinking level?
Is cost reported per completed task, including retries?
Does your harness record which model and thinking level produced each edit or tool call?
Are evaluations re-run on your own repos rather than trusting vendor benchmarks?
Related
R027 The tool-call index: the same model breaks tool calls at very different rates by provider, a reason to validate a new executor tier in your own harness.
R022 every harness component encodes an assumption about the model, so re-test it on each release.