Skip to content
A Spellwright Studios publication
Spellwright News

Intelligence · AI · Google

Gemini 4 Hydrogen

Closed

Google released Gemini 4 Hydrogen on November 12, 2026, as a generally available Gemini 4 model built for fast, lower-cost coding and agent work, beside its frontier models Carbon and Argon. It is on the free and paid tiers of the Gemini API, and Google says it is rolling out in the Gemini app, Google Search, Gemini for business and Google Antigravity.

Benchmark scores

  • Agents' Last Exam v1.1, pass@1 · Reported by the model's maker, Nov 12, 2026Setting: Google says its results use the Gemini API at the highest thinking setting (xhigh for Hydrogen) unless noted; no numeric budget is stated.71.6%#2 of 3 results
  • OSWorld (best of 3 runs) 2.1 · Reported by the model's maker, Nov 12, 2026Setting: Google says its results use the Gemini API at the highest thinking setting (xhigh for Hydrogen) unless noted; no numeric budget is stated.77.6%#3 of 3 results
  • Vending-Bench 2 · Reported by the model's maker, Nov 12, 2026Setting: Google says its results use the Gemini API at the highest thinking setting (xhigh for Hydrogen) unless noted; no numeric budget is stated.$14549#2 of 3 results
  • Artificial Analysis Intelligence Index v5.0 · Reported by the model's maker, Nov 12, 2026Setting: xhigh thinking level, per Google; no numeric budget is stated.54 points#2 of 3 results
  • Artificial Analysis Intelligence Index v5.0 · Reported by the model's maker, Nov 12, 2026Setting: High thinking level (Hydrogen's default), per Google; no numeric budget is stated.51 points#3 of 3 results
  • AutomationBench-AA · Reported by the model's maker, Nov 12, 2026Setting: Google says its results use the Gemini API at the highest thinking setting (xhigh for Hydrogen) unless noted; no numeric budget is stated.76.1%#1 of 2 results
  • Chartography, pass@1 · Reported by the model's maker, Nov 12, 2026Setting: Google says its results use the Gemini API at the highest thinking setting (xhigh for Hydrogen) unless noted; no numeric budget is stated.66.6%#2 of 2 results
  • HLE-Diamond · Reported by the model's maker, Nov 12, 2026Setting: Google says its results use the Gemini API at the highest thinking setting (xhigh for Hydrogen) unless noted; no numeric budget is stated.58.8%#1 of 2 results
  • SWE-sweep, pass@1 · Reported by the model's maker, Nov 12, 2026Setting: Google says its results use the Gemini API at the highest thinking setting (xhigh for Hydrogen) unless noted; no numeric budget is stated.13.5%#1 of 2 results
  • Terminal-Bench-Science 0.1, pass@1 · Reported by the model's maker, Nov 12, 2026Setting: Google says its results use the Gemini API at the highest thinking setting (xhigh for Hydrogen) unless noted; no numeric budget is stated.51.4%#2 of 2 results
  • Multimodal Knowledge Agent (MKA), pass@1 · Reported by the model's maker, Nov 12, 2026Setting: Google says its results use the Gemini API at the highest thinking setting (xhigh for Hydrogen) unless noted; no numeric budget is stated.53.4%#1 of 1 results
  • Terminal-Bench 4.1, pass@1 · Reported by the model's maker, Nov 12, 2026Setting: High effort level, per Google's methodology; no numeric budget is stated.57.0%#1 of 1 results
  • Vals Legal Research Bench · Reported by the model's maker, Nov 12, 202655.8%#1 of 1 results
  • Vals Tax Agent Bench · Reported by the model's maker, Nov 12, 202652.3%#1 of 1 results

Access

Google lists gemini-4-hydrogen as a stable, generally available Gemini API model from November 12, 2026, on the Free and Paid tiers of the Gemini API and AI Studio, and in Gemini for business. Google Cloud's Gemini Enterprise Agent Platform lists it as GA from November 12 (global, us and eu). Google says it is rolling out globally in Google Search for all users, in the Gemini app model menu (free users reach it only through auto model selection), in Sheets' Gemini features, first for AI Pro and Ultra subscribers, in Android Studio's paid tier, and in Google Antigravity for paid AI Pro and Ultra subscribers. It is the default model for the Antigravity agent in Gemini Managed Agents. Thinking levels: low, medium, high (default) and a new xhigh. No API call, account entitlement, quota or regional deployment was tested. No deprecation or retirement date is stated.