Corruption in response and stored response

API Bugs
1

Has web_search_preview been graduated to web_search in Responses API and has anyone else encountered this issue:

  • Model: gpt-5-mini-2025-08-07
  • API: Responses API (non-streaming)
  • Tool: web_search_preview enabled
  • SDK: Verified corruption exists in raw HTTP response (not SDK-related)

Response ID with corruption

resp_0b70cab40d5b69a4006976b59420c08192911f4b23921bed3d
You can retrieve this response directly - the corruption is in your stored data.

Examples of corrupted output (17 instances in one response)
keyscrosoft → “keys. Microsoft”

  1. legalSuch → “legal. Such”

  2. outshe → “outside, the”

  3. recovntire → “recovery key unlocks the entire”

  4. Investigatps → “Investigators”

  5. assumffers → “assume it offers”

  6. operationssessments → “operational assessments”

  7. cooperationnd → “cooperation and”

  8. storeuences → “stored and the consequences”

  9. provioves → “provider-blind improves”

  10. mortability → “portability”

  11. storaccount → “stored in your Microsoft account”

  12. tponse → “response”

  13. musfinitive → “must avoid definitive”

  14. ysis → “Analysis” (truncated at the start)

    Steps to reproduce

    1. Call Responses API with gpt-5-mini and web_search_preview enabled
    2. Request long-form content generation (~15-20k characters)
    3. Approximately 3-5% of responses contain merged/truncated words

    Verification

    Retrieved the response via raw HTTP GET (no SDK):
    GET https://api.openai.com/v1/responses/resp_0b70cab40d5b69a4006976b59420c08192911f4b23921bed3d
    Corruption is present in the stored response - this is server-side, not client-side.

    Related

    • GitHub Issue #2458: “OpenAI API responses.parse with web_search_preview tool returns corrupted JSON with control characters”
    • Affects multiple models (gpt-4.1, gpt-5-mini) when web_search_preview is used

    Impact

    Production content is being published with corrupted text, requiring manual review and correction of all outputs.