跪拜 Guibai
← Back to the summary

Google Retires Pro/Flash, Ships Gemini 4 Hydrogen: 2.5x Faster, Four Thinking Gears, and a Hidden API Pitfall

Author: 一缕82年的清风
Positioning: Architect's hands-on analysis and production deployment guide
Overview: Google abandons old naming, launches Gemini 4 Hydrogen (Hydrogen) unexpectedly on November 12. This article tests and breaks down its 2.5x speedup, four thinking levels, API sampling pitfalls, and a dual-model cost-saving orchestration pattern.

Just after discussing the flagship Gemini 4 Carbon, while browsing AI Studio yesterday (November 12), the long-awaited fast partner went live——Gemini 4 Hydrogen (gemini-4-hydrogen).

Google DeepMind introduced this twin pair, declaring that the long-used Pro / Flash naming is fully retired. Last night I switched my local agent that runs unit tests and fixes lint errors to Hydrogen for a two-hour trial. The feeling was immediate: previously I always had to watch a lightweight model when running terminal commands to avoid accidents. Now this workhorse can actually handle things cleanly on its own.

01. Goodbye Pro and Flash: Why "Carbon" and "Hydrogen"?

Switching to chemical element names and mapping them to a pipeline makes the division of labor clearer:

02. Three things from real testing: The worker is no longer clueless

After running a few rounds with the integration script, three details stood out:

1. 2.5x faster, now viable in terminal workflows

It completes tasks about 2.5x faster than the previous generation (e.g., Gemini 3.8 Flash), and parameters appear almost instantly after hitting Enter. The contrast between speed and intelligence is notable: on Artificial Analysis Intelligence Index (v5.0), it scores 51 points in the default high mode, and reaches 54 points in xhigh mode; Terminal-Bench 4.1 scores 57.0%. Previously lightweight models often stumbled on complex terminal commands, but now the workhorse can handle a large share of daily development work by itself.

2. Four thinking levels plus a new API pitfall

Native support for low, medium, high (default), and xhigh thinking_level options. Set low for instant replies; switch to high or xhigh for stubborn async issues.

Pitfall from testing: passing custom temperature, top_p, or top_k to gemini-4-hydrogen gets ignored by the API! Its inference path and sampling convergence are entirely controlled by thinking_level; also, thinking tokens count toward the unified 1M (1 million) shared input/output context window.

3. Flat pricing across 1M context plus free tier

The entire 1M window uses a uniform flat price (no more doubling beyond 200k): $1.50 / million input tokens, $7.50 / million output tokens. Compared to Carbon's preview of $2/$10 and standard $4/$20, running high-frequency agent loops is far cheaper; and Google AI Studio plus Gemini API still retains the beloved Free Tier.

03. 15 lines of test code: thinking level and thought token monitoring

The official google-genai SDK can start high mode and print thought consumption with 15 lines:

from google import genai
from google.genai import types

client = genai.Client()
resp = client.models.generate_content(
    model="gemini-4-hydrogen",
    contents="Debug the root cause of pytest error 'Event loop is closed' and give a fix",
    config=types.GenerateContentConfig(
        thinking_config=types.ThinkingConfig(
            thinking_level="high", include_thoughts=True
        )
    ),
)
print(f"Thought tokens: {resp.usage_metadata.thoughts_token_count}")
print(resp.text)

It is recommended to log thoughts_token_count in your gateway so you can see at a glance which chain is spinning idle and burning money.

04. Working as a duo: How to configure the Orchestrator?

Don't use a cannon on a mosquito, and don't let a worker force its way through top-level design. Based on Google Antigravity's Orchestrator (dynamic orchestrator) concept, the recommended three-step division:

  1. Carbon sets the contract: call gemini-4-carbon to define the contract and split task boundaries;
  2. Hydrogen guns through code: use medium/high to concurrently edit files, run unit tests, fix lint errors;
  3. Escalate on dead ends: switch to xhigh or pass back to Carbon when hitting persistent dead ends.

With nearly pure Carbon-level pass rates, you save a lot of response waiting and token bills. While the free tier is still available, try switching branches tonight.

💡 Follow 【一缕82年的清风】 to understand the bottom layers of technology and ecosystem evolution
Feel free to discuss, like, and share in the comments

Comments

Top 1 from juejin.cn, machine-translated. The original thread is authoritative.

用户368345607470

The hydrogen model's real selling point isn't a single benchmark score, but cost-effectiveness at the execution layer. Its terminal benchmark of 57.0% far exceeds the Flash model's 18.2%, but when the thinking level is set to ultra-high, token consumption and latency also rise. At the $1.50 input / $7.50 output pricing, the cost advantage shrinks considerably for output-heavy tasks. When routing in tools like Cursor, use the low tier for simple tasks and switch to ultra-high only for complex ones.