This is one of the reasons I keep gravitating back to local models even when the closed API ones are technically stronger. I had a production pipeline running on a major closed API for about four months. Stable, tested, working. Then one day the outputs started drifting. Not breaking errors, just su