Follow-up: Recurring MAX_TOKENS with no image output, despite maxOutputTokens: 32768

Gemini API
1

Following up on my earlier post, which hasn’t received a response and has dropped off the main page. This is still happening several times a day in our production. We weren’t seeing this issue before using gemini-nano-banana-2.1.

The API returns finishReason: "MAX_TOKENS" with no output content, even though the reported token usage is substantially below our configured maxOutputTokens: 32768.

After sending the same exact image a second time with the same data it works fine. It seems to be a very transient error.

Here is a recent example:

{

"finishReason": "MAX_TOKENS",

"promptTokenCount": 3905,

"candidatesTokenCount": 456,

"thoughtsTokenCount": 2063,

"totalTokenCount": 6424,

"contentPartCount": 0,

"request": {

"model": "gemini-nano-banana-2.1",

"maxOutputTokens": 32768,

"thinkingConfig": {

"thinkingLevel": "HIGH"

},

"imageConfig": {

"imageSize": "2K",

"aspectRatio": "3:2"

},

"inputWidth": 3840,

"inputHeight": 2561,

"temperature": 0

}

}

The request section is our logged configuration summary; the finish reason and token counts come from the provider response.

In this example, the reported candidate and thinking tokens total only 2,519, yet generation ends with MAX_TOKENS and no image or text is available to the application.

Earlier failures had the same pattern:

Input tokens Candidate tokens Thinking tokens Total tokens Returned content parts
4,620 1,904 2,554 9,078 0
4,406 1,483 3,136 9,025 0
4,522 704 1,734 6,960 0

Those requests also used maxOutputTokens: 32768, thinkingLevel: "HIGH", and 2K image output, with a 4:3 aspect ratio.

Could someone clarify:

  1. Is this a known issue with this model?
  2. Does MAX_TOKENS sometimes indicate an internal image-generation or thinking limit that is not reflected in the returned token counts?
  3. Is maxOutputTokens: 32768 honored for this model’s image generation, or is a different effective limit applied?
  4. Is there a recommended workaround—for example, a different thinking level—without significantly reducing editing quality?
  5. What additional diagnostics would help investigate this, and how should we provide a sample input and prompt privately?

These failures interrupt real customer edits, so any guidance would be appreciated. I can provide additional examples and capture any request identifiers or diagnostic fields you recommend.

Thank you!