Gemini 4 Hydrogen: everyday reasoning with four thinking levels and JSON output

Gemini 4 Hydrogen is Google's fast, low-cost model in the Gemini 4 family for everyday reasoning and multimodal input. Choose from four thinking levels, ground answers in Google Search, and return structured JSON that follows your schema. Text in, text out.

Gemini 4 Hydrogen is Google's fast, low-cost member of the Gemini 4 family, built as a daily driver for everyday reasoning and multimodal work. Send a text prompt — alone or with an image, a video (including YouTube links), audio, or a PDF — and it returns text. Its defining control is reasoning effort: four thinking levels, from low to xhigh, trade latency for depth. Hydrogen is the efficient tier; the Carbon and Argon siblings take the heaviest reasoning.

ModelGemini 4 HydrogenLow-cost reasoning LLM with adjustable thinking and JSON output.

What it does well

Reasoning that scales with effort

effort is the main lever. Four levels — low, medium, high and xhigh — set how hard Hydrogen thinks before it answers, and thinking is billed as output tokens, so you only pay for the depth a task needs. Keep it low for routine text, and raise it when a problem has many chained steps and a single correct answer. The puzzle below is a four-way constraint-logic deduction over potters, glazes and vessels, solved at high effort.

Four-constraint logic puzzlefor Gemini 4 Hydrogen
Four potters — Priya, Marco, Lena, and Tomas — work at a studio's four wheels, numbered 1 to 4 from left to right. Each used a different glaze (celadon, cobalt, amber, rose) and threw a different vessel (teapot, vase, bowl, mug). Clues: (1) Marco works at wheel 1. (2) Tomas sits immediately to the left of Priya. (3) Lena used the amber glaze. (4) The celadon glaze was used at wheel 4. (5) The rose glaze is somewhere to the right of the cobalt glaze. (6) The vase was thrown at wheel 3. (7) Priya threw the bowl. (8) The mug was not thrown at wheel 1. Work out the potter, glaze, and vessel at each of the four wheels, then give the answer in exactly this format: 'Marco threw the <vessel>; wheel 2 used the <glaze> glaze.'
Try it

At high effort, Hydrogen chained all eight clues to the one valid arrangement and returned the exact answer, "Marco threw the teapot; wheel 2 used the amber glaze", as schema-valid JSON on the first try, in a few seconds.

Structured JSON output

Set response_format to force machine-readable results. Pass {"type": "json_object"} for any valid JSON, or a json_schema that the reply has to match, which is ideal for pipelines that parse the answer directly. In the puzzle run, a one-field answer schema came back valid on the first attempt, with the format exactly as asked. Describe each field in the schema and Hydrogen fills it, so you can pass the output straight into the next step.

Multimodal input and grounded answers

Hydrogen reads more than text. Attach an image for OCR or chart reading, a video (including YouTube links) for summaries or timestamped questions, audio for transcription or Q&A, or a PDF for document extraction. You can also send extra files alongside the prompt through files. Turn on web_search to ground time-sensitive questions in live Google Search. A large context window lets you paste long documents, transcripts, or notes straight into the prompt.

Recommended settings

Match the thinking effort to the task, and state the output format you want — Hydrogen follows structure closely.

SettingValueWhen
effortlow (default)Routine drafting, rewriting, summaries, and simple Q&A
effortmediumLight reasoning and multi-part instructions
efforthighLogic, math, planning with constraints, and careful review
effortxhighThe hardest multi-step problems, when accuracy matters more than speed
response_formatjson_object or json_schemaWhen the output feeds another system
web_searchtrueWhen the answer depends on recent facts

Good to know

  • Hydrogen is the efficient tier. Google steers the deepest architectural judgment and the hardest problems to the Carbon and Argon models.
  • Very long or highly ambiguous tasks may need a higher effort level, which raises latency and output-token cost.
  • Only the prompt is required. Image, video, audio, PDF, and files are optional context you add as URLs when a question needs them; the video input also accepts YouTube links.
  • Thinking is billed as output tokens, so higher effort levels take longer. Step effort down when speed matters.

Use it from the API

Send the prompt, the effort level, and an optional response format in one call; the answer comes back directly.

APILogic puzzle as schema-valid JSONAPI reference
# pip install "segmind>=1.1.0"
# export SEGMIND_API_KEY="YOUR_API_KEY"
import segmind

result = segmind.run(
    "gemini-4-hydrogen",
    prompt="Four potters — Priya, Marco, Lena, and Tomas — work at a studio's four wheels, numbered 1 to 4 from left to right. Each used a different glaze (celadon, cobalt, amber, rose) and threw a different vessel (teapot, vase, bowl, mug). Clues: (1) Marco works at wheel 1. (2) Tomas sits immediately to the left of Priya. (3) Lena used the amber glaze. (4) The celadon glaze was used at wheel 4. (5) The rose glaze is somewhere to the right of the cobalt glaze. (6) The vase was thrown at wheel 3. (7) Priya threw the bowl. (8) The mug was not thrown at wheel 1. Work out the potter, glaze, and vessel at each of the four wheels, then give the answer in exactly this format: 'Marco threw the <vessel>; wheel 2 used the <glaze> glaze.'",
    effort="high",
    response_format={
        "type": "json_schema",
        "json_schema": {
            "name": "answer",
            "strict": True,
            "schema": {
                "type": "object",
                "properties": {
                    "answer": {
                        "type": "string",
                    },
                },
                "required": [
                    "answer",
                ],
                "additionalProperties": False,
            },
        },
    },
    web_search=False,
)
print(result.get("output"))
# 1. Submit the request; the reply carries its request_id
REQUEST_ID=$(curl -s -X POST "https://api.segmind.com/v2/gemini-4-hydrogen" \
  -H "x-api-key: YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "prompt": "Four potters — Priya, Marco, Lena, and Tomas — work at a studio'\''s four wheels, numbered 1 to 4 from left to right. Each used a different glaze (celadon, cobalt, amber, rose) and threw a different vessel (teapot, vase, bowl, mug). Clues: (1) Marco works at wheel 1. (2) Tomas sits immediately to the left of Priya. (3) Lena used the amber glaze. (4) The celadon glaze was used at wheel 4. (5) The rose glaze is somewhere to the right of the cobalt glaze. (6) The vase was thrown at wheel 3. (7) Priya threw the bowl. (8) The mug was not thrown at wheel 1. Work out the potter, glaze, and vessel at each of the four wheels, then give the answer in exactly this format: '\''Marco threw the <vessel>; wheel 2 used the <glaze> glaze.'\''",
  "effort": "high",
  "response_format": {
    "type": "json_schema",
    "json_schema": {
      "name": "answer",
      "strict": true,
      "schema": {
        "type": "object",
        "properties": {
          "answer": {
            "type": "string"
          }
        },
        "required": [
          "answer"
        ],
        "additionalProperties": false
      }
    }
  },
  "web_search": false
}' | jq -r .request_id)

# 2. Check its status until it is COMPLETED
curl -s "https://api.segmind.com/v2/requests/$REQUEST_ID/status" \
  -H "x-api-key: YOUR_API_KEY"

# 3. Fetch the result
curl -s "https://api.segmind.com/v2/requests/$REQUEST_ID" \
  -H "x-api-key: YOUR_API_KEY"

Resources

  • #text-generation
  • #google
  • #model-launch
  • #reasoning
  • #llm
AM

Abhilash Mirji

@abhilashmirji

Product engineer at Segmind, shipping new AI models faster than my cats can knock things off the desk. 11+ years turning ambiguous problems into dependable products. Off-call: drumming, scuba diving, and getting outvoted by four cats and two dogs.

View profile

Posts

More from Abhilash Mirji

Make your own with Segmind

Run every model and Pixelflow in this post, and build your own workflows.

Start creating