[waldrews]

Doesn't look like they blew up the API use cases, just the consumer UI access. I wouldn't be surprised if they allow it again, hidden behind a setting (along with allowing the different routed GPT5 levels to be in the selector).

[waynesonfire]

I have a feeling that the chatgpt ui does some behind-the scenes tuning as well--hidden prompt engineering if you will. I migrated to the api and 4o still seems different. Most obvious, I don't get the acks that make me feel like I should run for president.

Even ChatGPT 5 confirmed this,

why does the gpt-4o api not do this?

ChatGPT said:

Because the GPT-4o API is tuned and delivered in a neutral, low-intrusion style by default.

When OpenAI built GPT-4o for API use, they optimized it for:

    Predictable formatting (so it works well in code, pipelines, chatbots, etc.)

    Minimal unsolicited chatter (no “Nice!” or “Great job!” unless explicitly prompted)

    Deterministic tone — so that two API calls with the same input produce consistent, professional output without extra filler.

That’s different from the ChatGPT product experience, which has its own “assistant personality” layer that sometimes adds those rapport-building acknowledgements in casual conversation.

In API mode, you’re the one defining the personality, so if you want that “Good! Looks like you’re digging in” style, you have to bake it into the system prompt, for example:

[simonw]

The GPT-4o you talk to through ChatGPT and the GPT-4o you access via the API are different models... but they're actually both available via the API.

https://platform.openai.com/docs/models/gpt-4o is gpt-4o in the API, also available as three date-stamped snapshots: gpt-4o-2024-11-20 and gpt-4o-2024-08-06 and gpt-4o-2024-05-13 - priced at $2.50/million input and $10.00/million output.

https://platform.openai.com/docs/models/chatgpt-4o-latest is chatgpt-4o-latest in the API. This is the model used by ChatGPT 4o, and it doesn't provide date-stamped snapshots: the model is updated on a regular basis without warning. It costs $5/million input and $15/million output.

If you use the same system prompt as ChatGPT (from one of the system prompt leaks) with that chatgpt-4o-latest alias you should theoretically get the same experience.

[AlecSchueler]

But it always gives answers like that for questions where it doesn't know the actual reason.

[grues-dinner]

> Even ChatGPT 5 confirmed this,

>> why does the gpt-4o api not do this?

> ChatGPT said:

>> Because the GPT-4o API is tuned and delivered in a neutral, low-intrusion style by default.

But how sure are you that GPT-5 even had this data, and if it has it, it's accurate? This isn't information OpenAI has publicly divulged and it's ingested from scraped data, so either OpenAI told it what to say in this case, or it's making it up.

[andy99]

Ah ok, that's an important distinction. Seems much less a big deal then - or at least a consumer issue rather than a business one. Having never really used chatgpt (but used the apis a lot), I'm actually surprised that chat users would care. There are cost tradeoffs for the different models when building on them, but for chatgpt, it's less clear to me why one would move between selecting different models.

[svachalek]

Not everyone is an engineer. There's a substantial population that were selecting for maximum sycophancy.

[dragonwriter]

> There are cost tradeoffs for the different models when building on them, but for chatgpt, it's less clear to me why one would move between selecting different models.

The same tradeoffs (except cost, because that's roled into the plan not a factor when selecting on the interface) exist on ChatGPT, which is an app built on the underlying model like any other.

So getting rid of models that are stronger in some areas when adding a new one that is cheaper (presuming API costs also reflect cost to provide) has the same kinds of impacts on existing ChatGPT users established usages as it would have on a businesses established apps except that the ChatGPT users don't see a cost savings along with any disruption in how they were used to things working.

[Espressosaurus]

Different models have different (daily/weekly) limits and are better at different things.

o3 was for a self-contained problem I wanted to have chewed on for 15 minutes and then spit out a plausible solution (small weekly limit I think?)

o4-mini for general coding (daily limits)

o4-mini-high for coding when o4-mini isn't doing the job (weekly limits)

4o for pooping on (unlimited, but IMO only marginally useful)

[cgriswald]

Lower tiers have limited uses for some models.