Gemini 2.5 Pro intermittently emits reasoning-like content as normal text with include_thoughts=False #2925

Summary

When using gemini-2.5-pro through the Gemini Developer API with the Python GenAI SDK, reasoning-like content is intermittently emitted through the ordinary text path even though ThinkingConfig(include_thoughts=False) is set.

The response may begin with a preamble shaped like this synthetic, sanitized example:

thought. Here's a thinking process for how to respond...
[intended final response]

The application separates every response part for which bool(part.thought) is true before exposing normal text. The preamble nevertheless reaches the ordinary text path, so it is not structurally identifiable as a thought part at the point where the SDK response is consumed.

This appears related to #2121, but the observed environment uses Gemini 2.5 Pro, the Gemini Developer API, a newer SDK version, and explicitly sets include_thoughts=False.

Environment details

• Programming language: Python
• OS: Linux
• Language runtime version: 3.14.3
• Package version: google-genai==2.18.0
• Model: gemini-2.5-pro
• Backend: Gemini Developer API
• Mode: async generate_content_stream()

Sanitized request shape

from google import genai
from google.genai import types

client = genai.Client(api_key=API_KEY)

stream = await client.aio.models.generate_content_stream(
    model="gemini-2.5-pro",
    contents="Respond with only the final answer.",
    config=types.GenerateContentConfig(
        thinking_config=types.ThinkingConfig(
            thinking_budget=128,
            include_thoughts=False,
        ),
    ),
)

async for chunk in stream:
    for part in chunk.candidates[0].content.parts:
        if getattr(part, "thought", False):
            continue
        if part.text:
            print(part.text)

The behavior is intermittent, so this simplified prompt does not reproduce it deterministically. It shows the request and response-consumption shape without including any private prompt or application data.

Expected behavior

With include_thoughts=False, reasoning summaries should not appear in the ordinary visible text stream. If reasoning content is returned, applications should have a reliable structured field such as part.thought=True with which to separate it from the final response.

Actual behavior

Reasoning-like prose can appear before the intended final response through the same ordinary text path as the answer. Filtering all parts marked with part.thought=True therefore does not prevent the preamble from being exposed.

For affected responses, application-level usage metadata does not report thought tokens. Raw pre-SDK SSE payload capture is not currently available, so it is not yet clear whether the marker is absent in the API response or lost during SDK transformation.

Questions

1. Is reasoning content ever expected to appear in ordinary part.text with part.thought unset when include_thoughts=False?
2. Is this a known issue for gemini-2.5-pro on the Gemini Developer API?
3. Is there a reliable structured signal clients can use to distinguish this content from the final answer?
4. Is additional raw-response instrumentation recommended to determine whether this originates in the backend or the SDK transformation layer?

Related issues

• Gemini returns THOUGHT: content in text parts without part.thought=True #2121
• include_thoughts=False does not suppress thought parts for image gen models on Vertex AI #2239