Gemini 4 Hydrogen: Google’s Fast Mid-Tier Model and the January Pricing Cliff

Author

AI News Editorial

Published

2026-11-15 08:00

Google put a third Gemini 4 tier into its API and products in mid-November: Gemini 4 Hydrogen, model ID gemini-4-hydrogen. It sits below the frontier models Gemini 4 Carbon and Gemini 4 Argon. Google describes it as its “daily driver” for coding and everyday agent loops, and the model card states it is adapted from Gemini 4 Carbon. It is closed-weights and API-only — no self-hosting, no disclosed serving hardware — with a knowledge cutoff of March 2026.

The window and the controls. Hydrogen ships a 1,048,576-token unified window shared across input, thinking and output. Input accepts text, images, audio, video, PDFs, tables and charts; output is text only. Developers set thinking_level to low, medium, high or a new xhigh, and the default is now high rather than the medium used on 3.8 Flash. That default alone changes the bill of any long agent loop.

The speed claim is the headline. Google reports Hydrogen finishes tasks 2.5x faster than Gemini 3.8 Flash at default high thinking, with partner tests showing 2x to 2.8x faster runs and 12% to 23% fewer tokens. On Terminal-Bench 4.1 it reaches 57%, against 18.2% for 3.8 Flash, though it ranks only #5 on Agent Arena. It scores 51 on the Artificial Analysis Intelligence Index at high thinking and 54 at xhigh, edging past Argon’s 50.

The pricing calendar is the real story. Hydrogen costs $1.50 per 1M input tokens and $7.50 per 1M output, with output pricing including thinking tokens, and implicit caching cuts input to $0.075 for prompts of 4,096 tokens or more. Gemini 3.8 Flash costs $0.75/$3.75 — until 31 December 2026, when it rises to exactly $1.50/$7.50. Argon goes from $2/$10 to $4/$20 on 30 November. Hydrogen is not a cheaper option next to Flash; it is Flash’s price after the new year, arriving early. For teams doing long agentic coding, the deadline — not the benchmark — is what should drive the migration plan.

Two caveats worth stating. The model card reports 5x fewer fabricated claims than 3.8 Flash and 2.3x more resilience to indirect prompt injection on Gray Swan’s benchmark, but both are vendor numbers on vendor-selected suites. And in Google Antigravity the interesting pairing is Carbon orchestrating while Hydrogen edits files and runs tests, reaching near-Carbon task success on Terminal-Bench 4.1 with 24% less quota and 19% faster completion than Carbon alone.

Sources: Google DeepMind model card · MarkTechPost · Magica on Antigravity Orchestrator