Hi everyone,
I’ve been exploring the gpt-4o-mini-search-preview model recently, and I’m really impressed with its performance. However, I noticed that it currently doesn’t support streaming responses.
Does anyone know if streaming support is planned for this model, and if so, when it might be available?
Would love to hear from anyone who has insights or updates from the OpenAI team. Thanks in advance!
Now?
This is a Chat-Completions-only model. On Responses, you can use search as a tool (also less expensive as they are at the option of the AI).
I only had to remove top_p from my streaming benchmark.
Stream:gpt-4o-mini-search-preview
...................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................
ChatCompletionChunk(id=‘chatcmpl-1234’, choices=, created=1742547098, model=‘gpt-4o-mini-search-preview-2025-03-11’, object=‘chat.completion.chunk’, service_tier=‘default’, system_fingerprint=‘’, usage=CompletionUsage(completion_tokens=1381, prompt_tokens=1852, total_tokens=3233, completion_tokens_details=CompletionTokensDetails(accepted_prediction_tokens=0, audio_tokens=0, reasoning_tokens=0, rejected_prediction_tokens=0), prompt_tokens_details=PromptTokensDetails(audio_tokens=0, cached_tokens=0)))
| Stat | Average | Cold | Minimum | Maximum |
|---|---|---|---|---|
| stream rate | Avg: 90.350 | Cold: 92.4 | Min: 88.3 | Max: 92.4 |
| latency (s) | Avg: 2.577 | Cold: 3.1215 | Min: 2.0329 | Max: 3.1215 |
| total response (s) | Avg: 18.071 | Cold: 18.4752 | Min: 17.6661 | Max: 18.4752 |
| total rate | Avg: 77.489 | Cold: 76.806 | Min: 76.806 | Max: 78.172 |
| response tokens | Avg: 1400.000 | Cold: 1419 | Min: 1381 | Max: 1419 |
To report:
max_tokens=256 was ignored. That’s len(dots) == 1123 showing here, the chunk count (and more tokens in them to total 1381).
The internet retrieval on this model is not a tool under main AI control; you get billed for “web search tool calls” anyway, $0.03 extra for poems.
No cache, despite the same 1800 tokens input to activate it. Search injected into system likely breaks it.
oh sorry i’m ganna try without top_p now thank’s for your info