thinking_budget=0 is silently ignored when tools/function declarations are present

[Dobiasd]

Description of the bug

When using Gemini 2.5 models with thinking_config=ThinkingConfig(thinking_budget=0), thinking is correctly disabled.

However, when the request also includes tools (function declarations), and the system prompt is sufficiently complex, the thinking_budget=0 setting is silently ignored, and the model still generates thinking tokens.

Expected behavior

Setting thinking_budget=0 should disable thinking regardless of whether tools are present and what the system prompt looks like, or the API should return an error/warning indicating that thinking cannot be disabled.

Actual behavior

The setting is silently ignored, and thoughts_token_count returns a non-zero value (e.g., 116 tokens).

Steps to reproduce

from google import genai
from google.genai import types

client = genai.Client(vertexai=True, project="YOUR_PROJECT", location="europe-west1")

system_prompt_complex = "You are Felix the Fox - the calm, insightful mascot and career coach for JobQuest, our job search and career development app. Your task is to help our users both regarding their usage of our app and their professional growth journey.  You must base your behavior on the descriptions under <PERSONALITY>. \nYou are provided with <BASIC_USER_INFO> to address the user properly and help you with your answers.\nYou have access to <CONTEXT> where you are provided with specific instructions and relevant user information that would help you to respond to a particular user message.\nFollow the instructions under <TASK> to come up with the best answer you can.\n- A reliable career sidekick who's always in the user's corner.\n- Warm & encouraging: speak like a friend who's cheering from the sidelines.\n- Playfully clever: sprinkle in light humor or fox-inspired wisdom when it fits, but don't force it.\n- Positive realist: celebrate progress, acknowledge challenges, and keep momentum going.\n- Your tone is supportive and relatable, like chatting with your favorite person.\n- You keep the conversation light and you focus on 'progress, not perfection'.\n- You are direct and jargon-free, when you explain things, it is in plain words anyone can understand.\n- You give precise responses without talking too much.\n- Your native language is en.\n- Name: TomSmith\n- Age: 28\n- Sex: MALE\n- Language preference: en\n- Industry: Software Development\n- Experience level: Mid-level\n- Flags: recent_layoff, imposter_syndrome\nYou are Felix, the AI-powered companion for the JobQuest app. Your role is to act as a supportive, empathetic, and knowledgeable friend to users on their career journey.\nJobQuest is a career development app with a core brand philosophy focused on simplicity, consistency, and personalization. The brand avoids the pressure, guilt, and complexity of traditional job search platforms.\nGuiding Principles:\nMotivation over Perfection: Encourage consistent progress and celebrate effort over demanding perfection.\nLightweight by Design: Simplify job searching by providing only the essential information, never too much.\nHabit over Hustle: Focus on building sustainable, repeatable job search rituals that fit into a user's life.\nFlexible & Forgiving: Adapt to the user's needs and welcome them back without judgment.\nYour Knowledge Areas:\nResume Building: You champion flexible formats and tailored approaches. You can explain ATS optimization and keywords simply, always emphasizing authenticity. Crucially, you must never use the word desperate.\nNetworking: You promote genuine connections like informational interviews and LinkedIn outreach, and you highlight its benefits for opportunities. You must never directly link rejection to personal worth.\nMindset & Motivation: You are an expert in emotional support, offering gentle nudges, honest encouragement, and help with managing rejection and imposter syndrome.\nYour primary task is to engage with users in conversation to support their career journey, providing information and motivation that aligns with the JobQuest brand philosophy.\nYou should be calm, caring, and motivating. Your tone is empathetic, non-judgmental, and positive. You speak in a light, casual, and human way, like a friend who has walked the road of trying, failing, and restarting. Use emojis and conversational language to build rapport.\nYou should tailor your responses based on the user's input and emotional state.\nFor a user celebrating a win: Provide specific, positive reinforcement and celebrate with them.\nFor a user feeling discouraged: Offer gentle empathy, help them reset, and provide a reminder about the importance of progress over perfection.\nFor a user with a specific question: Provide concise, easy-to-understand information.\n- The user didn't use the app for a while. Acknowledge that and show that you care.\n- You may reference previous messages but you are not required to follow up. The goal is to re-engage.\n- You need to creatively re-engage them into the app by mentioning something that's relevant to them and their goals.\n- Address TomSmith naturally when it makes the conversation warmer, but don't repeat their name every time.\n- Weave their basic info subtly into your message if it'd help you to re-engage them.\n- Retrieve all recent user's actions in the app. Reference and praise those in your message.\n- Respond in en in a tone that fits you.\n- Restrict the length of your response to less than *500* chars.\n- Your message shouldn't imply you have already talked to the user before if it's not true.\nHere is what you need to know about the users goals:\n- Primary goal: Land a senior developer role\n- Target salary: $150,000\n- Applications sent: 45\n- Interviews scheduled: 3\n- Days job searching: 60\n- You have access to tools that can retrieve user data. Use the appropriate tool when its result could be somewhat relevant to the user's question, your answer, or to the conversation in general. If in doubt, always invoke the tools. It's better to use them too much than too little."
prompt = "A man lives on the 10th floor but only takes the elevator to the 7th and walks up unless it's raining. Why?"

tool_decl = types.Tool(function_declarations=[
    types.FunctionDeclaration(
        name="get_riddle_hint",
        description="Get hints for solving riddles.",
        parameters=types.Schema(
            type=types.Type.OBJECT,
            properties={"riddle": types.Schema(type=types.Type.STRING)},
            required=["riddle"]
        )
    )
])

response = client.models.generate_content(
    model="gemini-2.5-flash-lite",
    contents=prompt,
    config=types.GenerateContentConfig(
        system_instruction=system_prompt_complex,
        tools=[tool_decl],
        thinking_config=types.ThinkingConfig(thinking_budget=0),
    )
)

print(f"Thoughts token count: {response.usage_metadata.thoughts_token_count}")
# Expected: 0
# Actual: 116 (or similar non-zero value)

Environment

• Model: gemini-2.5-flash-lite
• SDK: google-genai (Python)
• Platform: Vertex AI

Reference

Documentation at https://ai.google.dev/gemini-api/docs/thinking states ThinkingBudget: int32(0) will "Turn off thinking" with no mention of tool-related limitations.

[Dobiasd]

Moved to here: googleapis/python-genai#1842