跪拜 Guibai
← All articles
Agent

Gemini 4 Hydrogen: Full Rollout, but Flash's Bargain Era Is Over

By lucas_AI ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Agent pipelines that relied on 3.8 Flash's $3.75 output price will see that cost double overnight, and any workflow that assumed model names stay put is now silently pointing at a different behavior. This release forces a cost re-check and a compatibility review for anyone running high-volume or unattended AI workloads.

Summary

Gemini 4 Hydrogen is Google's new everyday model, silently replacing 3.8 Flash across Gemini 4. On November 12 it launched to every tier—including free users via automatic model selection—with no migration notice or grace period. Independent measurement from Artificial Analysis puts its intelligence index at 54 (xhigh reasoning), one point below Claude Sonnet 5.5's 55 and above Gemini 4 Argon's 50. CursorBench 4.0 scores it 54.3% on default settings, a sharp jump from 3.8 Flash's 39.6%.

The performance gains are real: official benchmarks claim 2.5x faster task completion, Terminal-Bench 4.1 improving from 18.2% to 57%, and fabricated statements down fivefold. But the trade-off is price: $1.50 per million input tokens and $7.50 per million output tokens (including reasoning) with no introductory promotion. That doubles the output cost compared to 3.8 Flash's promo of $3.75. For agent pipelines billed on output, the monthly cost will climb immediately.

Google also launched Orchestrator mode in Antigravity, where Carbon (planning) and Hydrogen (execution) hand off dynamically based on task needs, cutting quota consumption 24% and response time 19%. Cursor, Cognition's Devin, and Mercor integrated on day one. But the rollout has rough edges: the model swap was silent, and the community is split on self-routing reliability—some point out that models tend to overestimate their own abilities, so letting them choose their own sub-models is an unproven bet.

Takeaways
— Gemini 4 Hydrogen launched November 12, replacing 3.8 Flash, and reached all tiers—including free users—on the same day with no migration notice.
— Pricing is $1.50 per million input tokens and $7.50 per million output tokens (including reasoning), no introductory discount; output cost is double 3.8 Flash's promotional $3.75.
— Artificial Analysis gives Hydrogen an intelligence index of 54 at xhigh reasoning, one point below Claude Sonnet 5.5's 55 and above Gemini 4 Argon's 50.
— CursorBench 4.0 scores Hydrogen 54.3% on default settings, up from 3.8 Flash's 39.6%, ranking fifth behind Carbon High, Claude Fable 5.2, Opus 5.5, and Sonnet 5.5.
— Official benchmarks show task completion 2.5x faster, Terminal-Bench 4.1 from 18.2% to 57%, and fabricated statements down 5x, but Terminal-Bench 4.1 still has Sonnet leading 63.9% vs 57.0%.
— xhigh reasoning generated 304 million tokens in evaluation tests, over four times the 72.8 million median for all models.
— Orchestrator mode pairs Carbon and Hydrogen dynamically, cutting quota consumption 24% and improving response time 19%; Cursor, Cognition, and Mercor integrated on day one.
Conclusions

Google's decision to launch without a promotional price marks a shift: the workhorse model is no longer a loss leader, and high-volume agent users will feel the doubled output cost immediately; the era of cheap default inference for everyday AI work is over.

Hydrogen surpassing Argon on the intelligence index within four weeks shows Google's own tiering is unstable; a 'flagship' label now has a shorter shelf life, which complicates long-term model selection for developers who need consistency beyond a few weeks.

Self-routing in Orchestrator mode bets on models judging their own limits, but models tend to overestimate their abilities; until that's proven reliable, developers should expect unpredictable token burn and behavior, not a magic efficiency win.

The silent replacement of 3.8 Flash—with no migration notice—signals that model names are becoming unreliable as stable API endpoints; pipelines must pin versions or expect breakage.

Concepts & terms
Intelligence index
A composite metric from Artificial Analysis that averages nine independent benchmarks to rate a model's overall capability; higher is better. It's used to compare models across vendors on a single scale.
Orchestrator mode
A feature in Google Antigravity that lets two different models (Carbon for planning, Hydrogen for execution) hand off tasks dynamically during a single job, aiming to reduce token consumption and latency by matching model strength to task phase.
Self-routing
The design where the model itself decides which sub-model to call at each step of a task, rather than using a fixed pipeline. Critics worry that models overestimate their own abilities, making this loop unpredictable in cost and correctness.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗