Research review / models

Gemini 4 Hydrogen Is Google's Fast Agent Model, and Every Number in the Launch Is Google's Own

Google released Gemini 4 Hydrogen on November 12 as the quick execution model under Carbon and Argon. It costs $1.50 per million input tokens and $7.50 per million output tokens across a 1 million-token context. Google reports large gains over Gemini 3.8 Flash in speed, agent benchmarks and safety, plus an Antigravity feature that switches between Carbon and Hydrogen within one session.

By Sandeep Belgavi · Published 2026-11-13

Google released Gemini 4 Hydrogen on November 12, 2026. The launch post places it beside "our frontier models, Gemini 4 Carbon and Gemini 4 Argon", which "excel at deep, long-horizon reasoning", while Hydrogen "is built for rapid execution across diverse agentic workflows." It is "priced flat across its entire 1 million-token context window at $1.50 per million input tokens and $7.50 per million output tokens." Google says it is rolling out the same day across the Gemini API in AI Studio, Android Studio, Gemini for business, Google Antigravity, the Gemini app and Google Search.

What Google claims

  • Versus Gemini 3.8 Flash. Google says Hydrogen finishes tasks "2.5 times faster at its high default setting" and more than triples Terminal-Bench 4.1 scores, "57% vs. 18.2%". It also says Hydrogen cuts fabricated claims "by more than 5x".
  • Versus Argon. Google says it "surpasses Argon on the Artificial Analysis Intelligence Index at its default high thinking level (51 vs. 50, and reaching 54 at its highest xhigh thinking setting)".
  • Agent safety. The post reports overreach in agentic tasks down "by up to 81%", misaligned actions and task shortcutting down "by 12x", and that on impossible or blocked tasks it is "up to 2x less likely to bypass permission boundaries". Against indirect prompt injection it is "more than 2.3x more resilient on the Gray Swan IPI benchmark."
  • Orchestrator. A new Antigravity feature "dynamically switches between models like Carbon and Hydrogen within the same session." Google says that on Terminal-Bench 4.0 it "reached near-Carbon task success rates while, on average, consuming 24% less quota and completing 19% faster."

How to read these numbers

All of these figures come from Google's announcement. Several are relative ("up to", "more than 5x") without a published baseline in the post, and the honesty and overreach figures come from internal evaluations. Google points to a model card for details. Until independent groups run the same tasks, they are best treated as a direction of travel, not as measurements to plan against.

The part that matters most for developers is the shape of the lineup. Google now sells a cheaper execution model alongside its expensive reasoning models and builds tooling that hands work between them. That fits the pattern of the last few months: one model plans and checks, and another, cheaper one does the many small tool calls. The economics depend on how often the cheap model can finish without escalating, and Orchestrator's 24% quota saving is the first number Google has given for that trade-off.

The safety claims are also worth reading closely. "Less likely to bypass permission boundaries" means a model can still bypass them, only less often. An agent that runs on Hydrogen still needs permission rules enforced outside the model, because a lower rate of attempts is not zero.

What remains uncertain. We found no model identifier in the launch posts and have not tested the model. We do not know how Orchestrator decides when to switch models, how the "up to" figures are distributed across tasks, or how Hydrogen compares with other vendors' small agent models on identical harnesses.

Analysis: At Belgavi's Book Store, Tanishka could try a two-tier setup on a copy of last month's help-desk traffic: the cheap model drafts replies and runs order lookups for each CASE, and the reasoning model reviews only refunds above a set amount. The measurement that decides it is how many cases the cheap model closes correctly without escalation, scored against the human outcome, not the vendor's benchmark table.

Sources and evidence

Google released Gemini 4 Hydrogen on November 12 as a fast execution model under Carbon and Argon, at $1.50 and $7.50 per million input and output tokens across 1 million tokens of context. Speed, benchmark and safety gains over Gemini 3.8 Flash are Google's own figures. Orchestrator in Antigravity switches models mid-session. Test on your own tasks before relying on the numbers.

Views are my own and do not represent any employer, past or present.