{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreiex4vz6ltm3ao54kjmounw6pnsuxjz23gvyd4uet3ws5iv4kr7wiu",
    "uri": "at://did:plc:lk3jfj3zq4k4wxnk474axylu/app.bsky.feed.post/3mjyvjewbek62"
  },
  "path": "/t/giving-up-on-realtime-mini/1379423#post_2",
  "publishedAt": "2026-04-21T10:38:12.000Z",
  "site": "https://community.openai.com",
  "tags": [
    "@nbo2"
  ],
  "textContent": "Hey @nbo2, really appreciate you taking the time to lay this out this clearly. This kind of breakdown is super helpful for others evaluating where Mini fits and where it doesn’t.\n\nWhat you’re seeing makes sense. Mini can recognize patterns, but it’s less consistent at enforcing constraints like scope, retries, or tool behavior. So things like going off-topic or missing “max tries” logic tend to show up more compared to Realtime.\n\nOn the model side, there’s no update yet on a Realtime 1.5 Mini or timeline. But this is exactly the kind of feedback that gets fed into future iterations, so I’ve passed it along to the team.\n\nYour point about maintaining separate prompts is a real tradeoff. A few folks are leaning toward lighter prompts + external guardrails for Mini to avoid that overhead, but agreed it’s not a clean solution.\n\nIf anything shifts or new versions land, this thread will likely be one others come back to. Good call documenting it.\n\n-Mark G.",
  "title": "Giving up on Realtime - Mini"
}