Introducing Gemini 4 Hydrogen

Context

Gemini 4 Hydrogen is the fast, cheaper model in Google's Gemini 4 family, released November 12 beneath the two frontier models, Argon and Carbon, which Google positions for long, hard reasoning problems. Its model card says Hydrogen is adapted from Carbon. It is also the model that puts Gemini 4 in front of most people: it runs in the Gemini app and Google Search for all users, and unlike Argon and Carbon it has a free tier on the Gemini API.

For developers, the case rests on price against capability. Google charges $1.50 per million input and $7.50 per million output tokens, the same rate across the full 1-million-token . Argon and Carbon cost $2 and $10 through November 30 and then $4 and $20, according to Google's API pricing page, so Hydrogen is a quarter cheaper this month and well under half the price from December. On capability, Google says Hydrogen at its default "high" thinking level scores 51 on the Artificial Analysis Intelligence Index, an aggregate of many runs by an outside evaluator, against 50 for Argon. Artificial Analysis's own listing confirms the 54 Google reports for the new top "xhigh" setting. A one-point gap on one composite index supports rough parity with Argon, not a lead.

Google pitches Hydrogen as a standalone everyday model, and also as the executor in a pairing: it writes code, runs tests and checks changes while Carbon or Argon handle system design and subtle bugs. To automate that split, Google added Orchestrator to Antigravity, its agentic coding environment, which switches between Carbon and Hydrogen within one session. Google reports that on the Terminal-Bench 4.0 coding benchmark this came close to Carbon's success rate while using 24% less quota and finishing 19% faster on average. The announcement does not give the success rates themselves, so how close "near-Carbon" is cannot be checked from it.

For teams on Gemini 3.8 Flash, the previous fast model, Google claims Hydrogen finishes tasks 2.5 times faster at its default setting and raises the Terminal-Bench 4.1 score from 18.2% to 57%. Partners testing data-heavy enterprise workloads saw it run 2 to 2.8 times faster with 12% to 23% fewer tokens. The per-token price is the catch for the rest of this year: Hydrogen's rate equals what 3.8 Flash will cost from January 1, 2027, but 3.8 Flash lists at half that until then. Apart from the 54 that Artificial Analysis lists, every comparison here comes from Google or its partners, including the one-point edge over Argon, so Hydrogen's standing as a cheap stand-in for Argon on everyday coding still needs independent tests to confirm it.

Key takeaways · AI-distilled

Google pitches Hydrogen as the execution half of a multi-model setup. Its new Antigravity Orchestrator switches between Carbon and Hydrogen in one session, and Google says it reached near-Carbon Terminal-Bench 4.0 success while using 24% less quota and finishing 19% faster.
Against Gemini 3.8 Flash, Google reports Hydrogen finishes tasks 2.5x faster at its default setting, scores 57% vs. 18.2% on Terminal-Bench 4.1, and cuts fabricated claims by more than 5x. Partners saw 2x to 2.8x speedups with 12% to 23% fewer tokens.
Part of the speed comes from tool-call behavior: Google says Hydrogen batches file-read and search calls in parallel when exploring a large repository instead of inspecting files one by one, and adapts when a tool call fails or a search returns empty.
Hydrogen offers four thinking levels: low, medium, high by default, and a new xhigh tier. Google says the levels scale predictably so users pay only for the reasoning depth a task needs, and the model handles up to 1 million input and 1 million output tokens.
In safety results Google reported, Hydrogen cuts misaligned actions and task shortcutting by 12x versus 3.8 Flash, is up to 2x less likely to bypass permission boundaries on blocked tasks, and is over 2.3x more resilient on the Gray Swan indirect benchmark.

Terms in this piece · Glossary

token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
prompt injection — An attack that hides instructions in content an AI will read — a webpage, email, or document — tricking it into following the attacker instead of the user.