Google releases Gemini 4 Hydrogen, claiming tasks run 2.5 times faster than 3.8 Flash

• Hydrogen finishes tasks 2.5 times faster than Gemini 3.8 Flash at its default setting, according to Google’s launch comparison.
• Early creative builds include animated spreadsheets and a shared voxel island, collected in minchoi’s thread.
• Carbon can coordinate Hydrogen workers for parallel coding and testing, according to Google’s launch details.
• Free Gemini app access is confirmed in Google’s reply; the rollout announcement also covers Search, developer tools and paid Antigravity access.

Google’s spreadsheet landscapes are a delightfully odd follow-up to the one-word “Gemini” teaser. A creator’s nineteen-cent volume slider animates a miniature Japanese room, while a Carbon-led Hydrogen swarm built a crimson moth in 45 minutes.

Hydrogen powered these ten projects, according to minchoi’s roundup. Costs, timings and prompt counts are those reported in the thread.

1. Google Sheets landscapes: Coastal trains, drifting clouds and falling leaves animate cell by cell. Google’s accompanying demo uses Python, native =LET(...) formulas and conditional formatting across thousands of shrunken cells.
2. Geminisland: A prompt adds a 3D voxel structure to a shared island that visitors can explore on foot.
3. Voxel koi garden: Twenty prompts inside the Gemini app produced a garden with a pagoda, cherry blossoms and lantern bridges.
4. Hanabi Switch: A volume slider brings a miniature Japanese room to life. Reported build cost: $0.19.
5. Interactive dining room: A 30-minute build with working lights and a refrigerator that opens.
6. Crimson moth: Carbon coordinated Hydrogen workers during a 45-minute build. Hydrogen also recorded the demonstration.
7. Desert ranch: Hydrogen finished the 3D scene in 56 seconds; 3.8 Flash reportedly took more than six minutes.
8. 3D Medic: A tool that inspects 3D files, identifies errors and patches them before export.
9. Island lagoon: Custom water, a pilotable sailboat and controls that move the scene from dawn to night.
10. Three.js futuristic city: Procedural towers, rain and flying traffic, with a fly-through interface after one follow-up prompt.

Google launched Hydrogen on November 12. Its announcement reports these improvements over 3.8 Flash:

• Terminal-Bench 4.1: 18.2% → 57%, a 38.8-percentage-point gain.
• Fabricated claims: More than a fivefold reduction.

The model exposes four reasoning settings:

• low
• medium
• high, the default
• xhigh, the new highest setting

Google reports an Artificial Analysis Intelligence Index score of 51 at high, versus Argon’s 50, rising to 54 at xhigh. Its announcement still positions Carbon and Argon for deep, long-horizon reasoning.

Antigravity’s new Orchestrator switches models during a single task. Google describes this handoff:

1. Carbon plans: Explore dependencies in a read-only phase, resolve ambiguities and define when replanning is necessary.
2. Hydrogen executes: Edit code, update packages and run tests.
3. Carbon revises: Inspect the state and repair the plan when Hydrogen hits friction, then resume execution.

Google reports near-Carbon task success with 24% lower quota consumption and 19% faster completion.

Parallel /swarm workflows use a different division of labor: Carbon coordinates while Hydrogen workers and verifiers build and test changes simultaneously in isolated workspaces.

Google’s developer announcement and Antigravity rollout specify access by surface:

• Developer tools: Gemini API, all tiers of AI Studio, and Android Studio.
• Antigravity: Standalone Hydrogen for paid Google AI Pro and Ultra subscribers.
• Orchestrator: Argon orchestration for paid Pro and Ultra subscribers; Carbon orchestration for Ultra subscribers.
• Search: Rolling out globally for all users.
• API pricing: $1.50 per million input tokens and $7.50 per million output tokens, flat across the one-million-token context window, according to the official launch post.

Free app access has a routing caveat: the launch post’s footnote says free users receive Hydrogen through automatic model selection when it is appropriate for their prompt.