Gemini 3 Flash Preview is a total mess right now. Fix the API structure

Gemini API
, , , ,
1

Hello again.

I’m currently on paid tier 1 using the Gemini 3 Flash Preview, and honestly, it’s been a nightmare. I don’t know what you guys did with the response structure, but it’s completely inconsistent and honestly feels like a scam with how it’s burning my tokens.

First off, why can’t we disable “Thinking” anymore?

I’ve already complained about this issue a long time ago, but I’d like to remind you again. In 2.x, we had the option. But now? Even on “Minimal,” it still triggers whenever it feels like it. Every time it does that, I’m paying for tokens I never wanted. Give us the “Off” switch back for the stable release. Period.

The JSON structure is a joke.

I’ve had to rewrite my parsing logic three times today because the parts array keeps changing. Sometimes parts[0] is “thought” and the actual reply is at the end. But then, out of nowhere, the API sends a useless \n (blank newline) as the last part. So if I try to filter out the “thought” text, my bot ends up showing a blank screen because you guys decided to throw a random empty part at the end. It’s inconsistent as hell.

JSON structure from Gemini API Logs and Datasets:

[Output - When the thinking is exposed]
{
  "candidates": [
    {
      "content": {
        "parts": [
          {
            "text": "thought\n"
          },
          {
            "text": "Model's Thinking Output"
          },
          {
            "text": "Acutal Model's Output"
          }
        ],
        "role": "model"
      },
      "finishReason": "STOP"
    }
  ],
(code omitted below...)

[Output - Suspected as a bug]
{
  "candidates": [
    {
      "content": {
        "parts": [
          {
            "text": "Acutal Model's Output"
          },
          {
            "text": "\n" <- wtf is this??
          }
        ],
        "role": "model"
      },
      "finishReason": "STOP"
    }
  ],
(code omitted below...)

And what’s with the random underscores?

Very intermittently, I’m seeing a _ at the end of bot’s responses for no reason. It looks unprofessional and broken. (e.g., “Is there anything also I can do for you today? :blush:_”, “Have a good night!_“)

I’m literally paying for this, and I’m wasting money on hidden tokens and debugging your garbage output. Stop over-complicating the response schema and just give us a clean, consistent string we can actually use.

Fix the “Thinking” toggle and clean up the parts array. This is unusable in its current state.

2 Likes