Java SDK: prompt_cache_retention uses "in-memory" instead of "in_memory" (API mismatch)

[360iresearch]

I don’t have the ability to file issues directly on your behalf, but I can give you a clean, ready-to-submit report that will get traction with the OpenAI SDK team.

Here’s a concise bug report you can paste into GitHub:

⸻

📌 Title

Java SDK: prompt_cache_retention uses "in-memory" instead of "in_memory" (API mismatch)

⸻

🐛 Description

The latest openai-java SDK exposes prompt_cache_retention with enum value:

"in-memory"

However, the OpenAI API expects:

"in_memory"

This causes a mismatch between SDK serialization and API contract.

⸻

✅ Expected behavior

SDK should send:

"prompt_cache_retention": "in_memory"

⸻

❌ Actual behavior

SDK sends:

"prompt_cache_retention": "in-memory"

Which may:

• Be rejected by the API, or
• Be ignored (falling back to default behavior)

⸻

🔁 Reproduction

ResponseCreateParams params = ResponseCreateParams.builder()
.model(ChatModel.GPT_5_2)
.input("Test prompt")
.promptCacheRetention("in-memory") // only allowed by SDK
.build();

⸻

🔎 Impact

• Prevents correct usage of extended prompt caching (in_memory vs 24h)
• Forces users to bypass SDK or use raw HTTP
• Breaks alignment with documented API behavior

⸻

💡 Suggested fix

• Update enum/string value to "in_memory"
• OR map "in-memory" → "in_memory" during serialization
• Ensure consistency across all SDKs

⸻

📚 References

• API docs: prompt caching uses "in_memory"
• SDK currently enforces "in-memory"

⸻

🛠 Workarounds

• Use raw HTTP request
• Inject raw string if SDK allows
• Use providerData / passthrough fields

⸻

Where to submit

Go here:
👉 https://github.com/openai/openai-java/issues

Click “New Issue” and paste the above.

⸻

If you want, I can tighten this further (e.g., include SDK version, exact class path, or failing HTTP trace) to increase the chance it gets fixed quickly.

[awa-xima]

See also openai/openai-python#2883 This seems to have happened to multiple SDKs.

Can confirm this happens on the latest SDK version 4.32.0

Also, this happens not only for the request parameters, but also for the response received from the API. The response also has a prompt_cache_retention field. As soon as you either call #validate() / #isValid on the response or access the prompt_cache_retention, you get an error

com.openai.errors.OpenAIInvalidDataException: Unknown PromptCacheRetention: in_memory

Curiously, we always call #validate() on the response and this used to work. I just realized our OpenAI integration is currently broken due to this issue -- not sure what changed?

Finally, the API documentation also says in-memory, so that may need to be changed as well.