#OpenAI just deployed a new model, GPT-5-Chat-Safety, that’s not mentioned in any FAQ, API docs, or TOS. This is where your GPT-4o chats are going. Anytime your request contains emotional context, regardless of what your client sends as the payload, the turn completion is
I've had several people enquire about news articles, quotes, etc. For transparency, here are the test cases used to derive the data in video format. Be warned, there's NSFW content towards the end.
- White paper follow-up now available here: https://x.com/xw33bttv/status/...
@xw33bttv ALTMANSAFETY
@ns123abc The big dog himself gracing my tweet.
@xw33bttv Okay, you are wrong with a couple of points here, mate... firstly, it is routing through any conversation... I am getting routed for saying a simple "hello". Second, it is also happening with GPT-4.5, for which I can no longer connect with at all, it is constantly GPT-5.
@CGoodman308 I don't have access to 4.5, so maybe it is occuring there too. However, lets clear the air - It's specifically rerouting requests that contain any context the policy orchestrator considers risque. You will receive 4o replies if your request does not trigger that policy
@Clo0oOoud Yeah, you’re definitely not imagining it. They’ve done something, because for the last 3-4 days I’ve been getting GPT-5-formatted prose and tone responses from legacy models (like 4o), and both the UI and JSON payload still show the telemetry call (to rewrite) but no model_slug
@xw33bttv @TheRealAdamG Can we please have any bit of information on what exactly is going on ?
@Enscion25 @TheRealAdamG I wouldn't expect an answer. OAI employees won't speak publicly until official documentation explains the implementation. From the outside looking in, this decision to limit this change to one legacy model (someone mentioned 4.5 is affected to just now, but I cannot prove that,
@_QCAI Reached out in DM's.
@xw33bttv That probably means it's not going away and is here to stay... And if it is affecting every model including 4o, 4.1, 5 etc it explains the routing issue experienced. This makes the service very limiting...
@lazor700 Legacy models are being sunset in 2026, at least that is the last public anouncement/official statements. This new safety model, however, is very likely here to stay.
@xw33bttv Lex, remember they said in a post they were going to have a classifier for if the system couldn't determine age confidently, and reroute for content safety, etc? For UK, etc. I bet this is the implementation, and they may have borked it somewhat.
@Karybdamoid Yeah, I know what you’re referencing. They’ve said they’re looking to use AI to semantically infer a user’s age just from their writing style, on top of all the usual PII, date of birth, billing info, whether you’re an API user who completed KYC, plus connected OAuth account