gemini-3.5-flash-lite rejects zero budget at thinking mode #2936

Activity

1. phenobarbital

   added

   type: bugError or flaw in code with unintended results or allowing sub-optimal usage patterns.Error or flaw in code with unintended results or allowing sub-optimal usage patterns.

   priority: p2Moderately-important priority. Fix may not be included in next release.Moderately-important priority. Fix may not be included in next release.

   on Sep 3, 2026
2. Venkaiahbabuneelam

   self-assigned this

   on Sep 3, 2026
3. Venkaiahbabuneelam commented on Sep 3, 2026

   

   Venkaiahbabuneelam

   on Sep 3, 2026

   More actions

   Hi @phenobarbital, Thanks for reaching out us!
   To help us understand the issue, could you please share a code snippet of what you've tried so far?
   Thanks
4. Venkaiahbabuneelam

   added

   status:awaiting user response

   on Sep 3, 2026
5. phenobarbital commented on Sep 4, 2026

   

   phenobarbital

   on Sep 4, 2026

   Last edited by phenobarbital

   Author

   More actions

   @Venkaiahbabuneelam yes, sure!

   I tested directly using this snippet:

   import asyncio
   import json
   import os
   import aiohttp
   from google import genai
   from google.genai import types
   from parrot.conf import config
   
   API_KEY = config.get("GOOGLE_API_KEY") or os.getenv("GOOGLE_API_KEY") or os.getenv("GEMINI_API_KEY")
   
   async def test_raw_rest(session, model, thinking_budget, label):
       url = f"https://generativelanguage.googleapis.com/v1beta/models/{model}:generateContent?key={API_KEY}"
       generation_config = {}
       if thinking_budget is not None:
           generation_config["thinkingConfig"] = {"thinkingBudget": thinking_budget}
       
       payload = {
           "contents": [{"parts": [{"text": "Say hello in 5 words."}]}],
       }
       if generation_config:
           payload["generationConfig"] = generation_config
   
       try:
           async with session.post(url, json=payload) as resp:
               status = resp.status
               text = await resp.text()
               try:
                   data = json.loads(text)
                   if status == 200:
                       cand = data.get("candidates", [{}])[0]
                       content = cand.get("content", {}).get("parts", [{}])[0].get("text", "")
                       print(f"[REST] {model} | {label} -> HTTP {status} OK: {content.strip()[:60]}")
                   else:
                       err = data.get("error", {}).get("message", text[:100])
                       print(f"[REST] {model} | {label} -> HTTP {status} ERROR: {err}")
               except Exception:
                   print(f"[REST] {model} | {label} -> HTTP {status} Raw: {text[:100]}")
               return status
       except Exception as e:
           print(f"[REST] {model} | {label} -> EXCEPTION: {e}")
           return None
   
   async def test_sdk(client, model, thinking_budget, thinking_level, label):
       th_cfg = None
       if thinking_budget is not None or thinking_level is not None:
           kwargs = {}
           if thinking_budget is not None:
               kwargs["thinking_budget"] = thinking_budget
           if thinking_level is not None:
               kwargs["thinking_level"] = thinking_level
           th_cfg = types.ThinkingConfig(**kwargs)
   
       gen_cfg = types.GenerateContentConfig(thinking_config=th_cfg) if th_cfg else None
   
       try:
           resp = await client.aio.models.generate_content(
               model=model,
               contents="Say hello in 5 words.",
               config=gen_cfg,
           )
           print(f"[SDK]  {model} | {label} -> SUCCESS: {resp.text.strip()[:60] if resp.text else 'EMPTY'}")
           return 200
       except Exception as e:
           print(f"[SDK]  {model} | {label} -> EXCEPTION: {type(e).__name__}: {e}")
           return str(e)
   
   async def main():
       print("=" * 60)
       print("TESTING GEMINI-3.5-FLASH-LITE & CONTROL MODELS")
       print(f"SDK version: {genai.__version__}")
       print("=" * 60)
   
       models_to_test = [
           "gemini-3.5-flash-lite",
           "gemini-2.5-flash-lite",
       ]
   
       async with aiohttp.ClientSession() as session:
           print("\n--- 1. RAW REST (v1beta) ---")
           for m in models_to_test:
               for tb, label in [
                   (None, "omitted"),
                   (0, "thinkingBudget: 0"),
                   (-1, "thinkingBudget: -1"),
                   (512, "thinkingBudget: 512"),
               ]:
                   await test_raw_rest(session, m, tb, label)
   
       print("\n--- 2. GOOGLE-GENAI SDK ---")
       client = genai.Client(api_key=API_KEY)
       for m in models_to_test:
           for tb, tl, label in [
               (None, None, "omitted"),
               (0, None, "thinking_budget=0"),
               (-1, None, "thinking_budget=-1"),
               (512, None, "thinking_budget=512"),
               (None, "MINIMAL", "thinking_level='MINIMAL'"),
               (None, "LOW", "thinking_level='LOW'"),
           ]:
               await test_sdk(client, m, tb, tl, label)
   
   if __name__ == "__main__":
       asyncio.run(main())
   

   When is called with plain aiohttp calls I got this:

   --- 1. RAW REST (v1beta) ---
   [REST] gemini-3.5-flash-lite | omitted -> HTTP 200 OK: Hello, how are you today?
   [REST] gemini-3.5-flash-lite | thinkingBudget: 0 -> HTTP 400 ERROR: Request contains an invalid argument.
   [REST] gemini-3.5-flash-lite | thinkingBudget: -1 -> HTTP 200 OK: Hi, nice to meet you.
   [REST] gemini-3.5-flash-lite | thinkingBudget: 512 -> HTTP 200 OK: Hello there, how are you?
   [REST] gemini-2.5-flash-lite | omitted -> HTTP 200 OK: Hello there, how are you?
   [REST] gemini-2.5-flash-lite | thinkingBudget: 0 -> HTTP 200 OK: Hello, it's great to see you.
   [REST] gemini-2.5-flash-lite | thinkingBudget: -1 -> HTTP 200 OK: Hello! Wishing you a nice day.
   [REST] gemini-2.5-flash-lite | thinkingBudget: 512 -> HTTP 200 OK: Hello, how are you doing?
   
   

   but with google-genai (version: 2.18.1) I got:

   [SDK]  gemini-3.5-flash-lite | omitted -> SUCCESS: Hello there, how are you?
   [INFO] 2026-09-04 12:34:58,055 google_genai.models(models.py:8757) :: AFC is enabled with max remote calls: 10.
   [SDK]  gemini-3.5-flash-lite | thinking_budget=0 -> EXCEPTION: ClientError: 400 INVALID_ARGUMENT. {'error': {'code': 400, 'message': 'Request contains an invalid argument.', 'status': 'INVALID_ARGUMENT'}}
   

   the unrecognized thinkingBudget=0 only affects gemini-3.5-flash-lite, any other flash models (e.g. 3.8) or older (like 3.1 or 2.5) are not affected by this issue.

   I was using thinkingBudget=0 on any model but gemini-3.5-flash-lite.

   as conclusion:

   Key Inconsistencies for the Incident Report

   1. SDK Docstring vs Reality:
      The google-genai SDK docstring (types.py:13606) states:
      │ "Starting from Gemini 3.5 models, the old thinking_budget will no longer be supported and will result in a user error if set. Instead, users should use the thinking_level parameter to control the maximum thinking depth."
      However, thinking_budget is not completely unsupported on 3.5: -1 and 512 succeed without errors. Only 0 is rejected.
   2. No Mechanism to Disable Thinking:
      types.ThinkingLevel only defines MINIMAL, LOW, MEDIUM, and HIGH. It does not provide an OFF or NONE level. Since thinking_budget=0 raises a 400 error, developers migrating from 2.5 have no way to turn thinking off.
   3. Opaque Error Message:
      Pro models explicitly explain invalid budget configurations (e.g. "Budget 0 is invalid. This model only works in thinking mode"). gemini-3.5-flash-lite only outputs the generic message: "Request contains an invalid argument.")

   👍React with 👍1Reacted by Venkaiahbabuneelam
6. Venkaiahbabuneelam

   removed

   status:awaiting user response

   on Sep 7, 2026
7. kkorpal commented on Sep 8, 2026

   

   kkorpal

   on Sep 8, 2026

   More actions

   Hi @phenobarbital, thanks for the fantastic report and test script! We ran it and perfectly reproduced the 400 error with thinkingBudget=0 on gemini-3.5-flash-lite, as well as the missing OFF state in the new thinking levels. I've escalated this straight to the engineering and API teams to get both issues sorted, thanks again for flagging this.