A cheaper Gemini with a frontier-level score

Google released Gemini 4 Hydrogen on 12 November, a model it places below the frontier pair it shipped in October and prices at $1.50 per million input tokens and $7.50 per million output tokens. Those rates are flat across the model’s full one-million-token context window, with no step increase for long prompts, and the model can return up to one million output tokens.

The release matters less for what it adds at the top of Google’s range than for what it moves down it. On the Artificial Analysis Intelligence Index, Google reports Hydrogen scoring 51 at its default “high” thinking level — one point above Gemini 4 Argon, one of the two frontier models it now sits beside. At a new “xhigh” setting Google puts the figure at 54. Both numbers appear on Google’s own announcement, written by Koray Kavukcuoglu, the company’s chief AI architect, and not on the independent leaderboard that maintains the index.

Illuminated fibre-optic strands fanning out against a dark background
Google reports the model running about 2.5 times faster than the version it replaces. Stock image. Suki Lee · pexels · Pexels License

What Google measured, and what it did not

The announcement is dense with scores, and nearly all of them are Google’s own. The company puts Hydrogen at 57% on Terminal-Bench 4.1 against 18.2% for Gemini 3.8 Flash, the model it effectively replaces; at 58.8% on HLE-Diamond; at 71.6% on Agents’ Last Exam v1.1; and at 52.3% on the Vals Tax Agent Bench, which it describes as first overall. On the Vals Legal Research Bench it reports 55.8% and a tie for first.

One figure is checkable against an outside source. Google says Hydrogen reaches 53.4% on the Multimodal Knowledge Agent benchmark, the knowledge-work test it built with Artificial Analysis and launched on 10 November. Artificial Analysis’s public MKA table is led by Anthropic’s Claude Fable 5.2 at 57%, with Gemini 4 Carbon second at 54%. Hydrogen’s number sits below both.

Speed is the clearer claim. Google says the model runs about 2.5 times faster than Gemini 3.8 Flash at the high setting, and that partners testing it measured between 2 and 2.8 times faster with 12% to 23% fewer tokens used. Inside Google Antigravity, the company’s agent environment, it says Hydrogen as orchestrator came close to Carbon’s task success on Terminal-Bench while consuming 24% less quota and finishing 19% faster on average.

An empty meeting room with mesh chairs around a long wooden table
Google's agent-safety figures for the model come from its own internal evaluations. Stock image. Max Vakhtbovych · pexels · Pexels License

The safety figures are internal

Google also published a block of agent-safety numbers, all from its own evaluations: up to an 81% reduction in overreach and up to 79% in under-service on agentic tasks, a twelvefold reduction in misaligned actions and task shortcutting across tool-use and coding evaluations, and a model up to twice as unlikely to push past a permission boundary when a task is blocked. It reports a 96.7% honesty rate on an internal agent honesty evaluation and 99.9% chain-of-thought legibility across more than 200,000 audited turns. One third-party figure appears: Google says the model is more than 2.3 times more resilient on the Gray Swan indirect prompt-injection benchmark.

Where it lands

Hydrogen is rolling out from 12 November through the Gemini API and Google AI Studio, the paid tier of Android Studio, and Antigravity for paying Google AI Pro and Ultra subscribers. Businesses get it through Gemini for business and in Google Sheets. Consumers get it in the Gemini app’s model picker, and free users reach it through automatic model selection. It also goes into Google Search for all users, which is the part of the launch with the largest audience and the fewest published numbers.

What to watch is whether Artificial Analysis’s independent run reproduces the index score Google quoted for it. The gap between a vendor’s figure and a leaderboard’s own measurement is where this release will be judged.