Gemini-4-carbon changes modelVersion to Flash Lite at STOP (native HTTPS repro)

Requests to the official gemini-4-carbon:streamGenerateContent endpoint can begin with modelVersion: gemini-4-carbon and end with modelVersion: 3.1-flash-lite-05-2026, while retaining the same responseId.

This was first observed in DeepSeek Harness, then reproduced outside both the application and the Google SDK using native Node HTTPS, with no tools or conversation history. TLS certificate validation succeeded on the generation connections. The observations below are from 2026-11-11.

This report concerns inconsistent response metadata. The actual serving weights, internal routing and billing SKU have not been independently verified; it is not a claim that the client intentionally switches models.

Minimal request

POST to:

https://generativelanguage.googleapis.com/v1beta/models/gemini-4-carbon:streamGenerateContent?alt=sse

Use the normal API-key authentication header. No credentials are included here.

{
  "contents": [{
    "role": "user",
    "parts": [{"text": "请计算 17 + 26,只回复结果数字。"}]
  }],
  "systemInstruction": {
    "parts": [{"text": "You are an AI agent powered by DeepSeek Harness.\n\nYou are a coding agent powered by the gemini-4-carbon model."}]
  },
  "generationConfig": {
    "maxOutputTokens": 32768,
    "thinkingConfig": {"includeThoughts": true, "thinkingLevel": "HIGH"}
  }
}

The user prompt means: “Calculate 17 + 26; reply with the number only.”

Read all SSE events, including the final STOP event, rather than inspecting only the first modelVersion. The arithmetic answer is correctly returned as 43. The unexpected behavior is that the final STOP event identifies Flash Lite instead of Carbon within the same response ID. This minimal two-paragraph case was observed twice.

These two system sentences correspond to public default agent instructions in the tested DeepSeek Harness version; no private instructions are necessary. The native HTTPS reproduction did not use DSH/Google SDK identification headers or automatic follow-up requests.

Controls

System instruction Observed response metadata
None Carbon through STOP
Neutral helpful-assistant text, in one or two paragraphs Carbon through STOP
Either identity sentence above alone Carbon through STOP
Both sentences separated by \n\n Carbon initially, Flash Lite at STOP; observed twice
Identical sentences separated by one space Carbon through STOP; observed twice
Full default DSH system text Carbon to Flash Lite
Full system text with the two sentences removed, or the entire text folded into one paragraph Anomaly still present

Thus, the small request is a reproducer, not evidence that removing blank lines is a general fix. Some repeated inputs produced identical outputs, so these observations are not a randomized failure-rate estimate.

A separate isolated experiment faithfully restated the full default system instructions in Chinese while retaining the existing tools, runtime context, permissions and native signature replay. A read/write tool task then returned Carbon through STOP on all four requests; an adjacent original-system control still showed the anomaly. This is a limited compatibility observation, not a root-cause diagnosis or a guaranteed workaround.

Questions

  1. What does modelVersion mean when it changes within one response ID from the requested Carbon model to Flash Lite?
  2. Is this a response-metadata defect, auxiliary processing, routing behavior, or another known issue? Can Google clarify which model generated the response?
  3. Is there a supported correction for ordinary agent system instructions that preserves tool calling and signature replay?
  4. How should clients detect and display this condition instead of presenting only the requested model name?

The original application environment was Windows, DeepSeek Harness Desktop 0.2.2-rc.1, bundled pi-ai 1.3.0 and Google SDK 2.21.0; the minimal reproduction does not depend on those client libraries. The endpoint is the official Google native v1beta API, with HIGH thinking and maxOutputTokens: 32768.

Selected timestamps and response metadata can be supplied through an appropriate private support channel after review. API keys, actual response IDs, account/project identifiers, private conversation logs, thought contents and signatures are intentionally not posted here. Thank you.

Same issue. The output tokens are charged with 3.1-Flash-Lite, but input with Carbon

The same happens to me.