Document 2.5 Flash thinking in REST/curl examples

[maflcko]

Description of the feature request:

I was looking for documentation on how to use the new 2.5 Flash Thinking model with REST/curl. However, I couldn't find any examples or docs on how to toggle the thinking mode or set the thinking budget via curl.

I tried the following search: https://github.com/search?q=repo%3Agoogle-gemini%2Fcookbook+thinking+path%3Aquickstarts%2Frest+&type=code

What problem are you trying to solve with this feature?

No response

Any other information you'd like to share?

No response

[X901]

I ask the same thing here

but I don't get response

I didn't get any errors but it's not feel as it realy disabled as the OpenRouter model

const requestBody = {
            contents: [{
                parts: [
                    { text: preamble },
                    { text: prompt }
                ]
            }],
            generationConfig: {
                temperature: 0.3,   
                response_mime_type: "application/json",
                thinkingConfig: { thinkingBudget: 0 }
                ,response_schema: {
                    type: "object",
                    properties: {
                        translations: {
                            type: "array",
                            items: { type: "string" }
                        }
                    },
                    required: ["array"]
                }
            }
        };

[andycandy]

Here this is how you do it:

prompt = """
    You are playing the 20 question game. You know that what you are looking for
    is a aquatic mammal that doesn't live in the sea, and that's smaller than a
    cat. What could that be and how could you make sure?
"""

MODEL_ID = "gemini-2.5-flash-preview-04-17"
url = f"https://generativelanguage.googleapis.com/v1beta/models/{MODEL_ID}:generateContent?key={GOOGLE_API_KEY}"  

headers = {
    'Content-Type': 'application/json',
}
data = {
    "contents": [{
            "parts":[
                {"text": prompt}
            ]
        }],
    "generationConfig": {
        "thinkingConfig": {
            "thinkingBudget": 0
        }
    }
}
response = requests.post(
    url,
    headers=headers,
    data=json.dumps(data)
)

Output:

{
    "candidates": [
        {
            "content": {
                "parts": [
                    {
                        "text": "Okay, playing 20 Questions with those constraints! Here's what fits the criteria and how to refine:\n\n**What it could be:**\n\nGiven the constraints \u2013 an aquatic mammal that *doesn't* live in the sea and is smaller than a cat \u2013 the most likely candidate is a. .................."
                    }
                ],
                "role": "model"
            },
            "finishReason": "STOP",
            "index": 0
        }
    ],
    "usageMetadata": {
        "promptTokenCount": 58,
        "candidatesTokenCount": 609,
        "totalTokenCount": 667,
        "promptTokensDetails": [
            {
                "modality": "TEXT",
                "tokenCount": 58
            }
        ]
    },
    "modelVersion": "gemini-2.5-flash-preview-04-17"
}

Here you can see it didn't use any thinking token.

Moreover if you rerun the same code but without limiting the token you get this output:

Code:

#before and after stays the same
data = {
    "contents": [{
            "parts":[
                {"text": prompt}
            ]
        }],
}

Output:

{
    "candidates": [
        {
            "content": {
                "parts": [
                    {
                        "text": "Okay, let's break down the clues:\n\n1.  **Aquatic Mammal:** Lives in or spends significant time in water, and is a mammal ..............."
                    }
                ],
                "role": "model"
            },
            "finishReason": "STOP",
            "index": 0
        }
    ],
    "usageMetadata": {
        "promptTokenCount": 58,
        "candidatesTokenCount": 3622,
        "totalTokenCount": 3680,
        "promptTokensDetails": [
            {
                "modality": "TEXT",
                "tokenCount": 58
            }
        ],
        "thoughtsTokenCount": 2696
    },
    "modelVersion": "gemini-2.5-flash-preview-04-17"
}

[andycandy]

@Giom-V Regarding this query should I make a new Get_started_with_thinking_REST notebook to address this?

[maflcko]

The docs are also here: https://ai.google.dev/gemini-api/docs/thinking#set-budget

So I'll close this for now, because my immediate question has been answered. The notebook can still be added, if needed, or course.

[X901]

I tried it, it's too slow as if thinking is still enable when I use the Google AI Studio

but in OpenRouter it faster like Flash 2.0

why performance are different ?

[maflcko]

It could be due to https://twitter.com/OfficialLoganK/status/1912986097765789782, but I don't have any insight here.

In some rare cases, the model still thinks a little even with thinking budget = 0, we are hoping to fix this before we make this model stable and you won't be billed for thinking. The thinking budget = 0 is what triggers the billing switch.