OpenAI's GPT-5 Bait and Switch | Hacker News Reader

OpenAI's GPT-5 Bait and Switch

vincentschmalbach.com

38 pointsvincent_s20 comments

"Apple Intelligence" meant whatever Apple marketing wants it to be.

Now "GPT 5" is also a marketing term. You are paying for the abstracted experience and brand. Not a specific model or defined capabilities anymore.

> "Apple Intelligence" meant whatever Apple marketing wants it to be. Now "GPT 5" is also a marketing term. You are paying for the abstracted experience and brand. Not a specific model or defined capabilities anymore.

Seems like an apples to oranges comparison (pun not intended). The term Apple Intelligence was introduced for the new service/product. GPT<n> has a history of standing for something (a specific model).

As a user I like to pick models. Maybe I am using them wrong but I actually find reasoning/chain-of-thought models worse at some things, even things they are supposedly better at like coding. They are also buggier in my experience and an order of magnitude slower.

This also doesn't bode well for AI progress. It feels like the kind of thing you release if you want to release something but haven't made much progress on improving the fundamental language model. Even chain of thought itself, while interesting, is a bit disappointing compared to language model improvements. It is more comparable to automated prompt engineering than a fundamental improvement of the model to represent knowledge and problem solving. This development seems to be doubling down on this direction.

> Maybe I am using them wrong but I actually find reasoning/chain-of-thought models worse at some things

You are not using them wrong. Sonnet-3.5, famously not a reasoning model, is still easily the best generalist model that isn't enormously expensive and slow (o1). R1 is at times better but also much less consistent, and almost unusably slow.

This is based on a wide-range of tasks, both coding, non-coding objective, as well as clearly subjective non-STEM, often run in parallel across 4+ models for comparison purposes. Both standard chat completion as well as agentic.

o3-mini has been a big disappointment, sometimes being good at one-shot coding but being much worse than Sonnet at everything else. It feels very rushed as a response to Deepseek.

Any kind of benchmark that had ever been quoted by an LLM provider in their own reports is completely worthless.

I think the idea is that since o1 is after all demonstrably SOTA on the large majority of real-world tasks, that is the way to go, because they feel like they can make it cheaper/smaller/faster while keeping that performance.

Whether that actually holds remains to be seen.

This is really concerning. Why should i pay for the pro subscription if i have no real control over when i get the most advanced reasoning capabilities. Usually i defend OpenAI but in this case i really hope their competitors catch up fast and don't follow the same bullshit path like Sam does.

> OpenAI could be setting the stage for a gradual reduction in service while maintaining premium prices.

This is rich. OpenAI is basically setting money on fire. They've always been very generous with the free tier and and going to offer unlimited gpt-5 with "standard intelligence" to users. They even allow users some access without an account.

For all its flaws, this is not one of them. It discredits the authors argument because it just comes off as OpenAI hater.

Regarding picking a model, sure you think you want control, and maybe you do. But likely the model will be better at picking the correct model for you. You see this with prompt engineering being less important with every generation of model. Not to mention 95% of users have no idea what model they want.

I'm wondering: what will happen with the API? While they can give "normal" users the middle finger with shrinkflating model capabilities and opacity; I don't think businesses will be too happy sending prompts to the "super-intelligent" model that suddently is performing way worse despite the client not changing their prompts.

Or they'll keep that "weird" (not so hard to understand IMO) naming scheme for the API, which means any actual "power user" might choose instead to pay API credits and run their own Open WebUI (or any other ChatGPT-Like UI).

> OpenAI's Developer Experience Community Lead Edwin Arbus clarified in the forum: "The API will support the o3 model for developers to use [...] For the purposes of the developer platform: the broader GPT-5 model system (and o3 as a part of it) will maintain the granularity that developers need."

https://community.openai.com/t/openai-roadmap-and-characters...

The blackboxing makes it tougher to compare with anything, and means they can say you're holding it wrong if the router brings you a second-rate response.

So GPT-5 will be about unifying GPT-4o, GPT-4o-mini, GPT-o3 and GPT-o3-mini into one single service which behaves similar in how OpenRouter functions?

> The free tier of ChatGPT will get unlimited chat access to GPT-5 at the standard intelligence setting

Fantastic! But what if I am not satisfied with the result chatGPT has provided? Can I make it understand that the model they routed my prompt to wasn't enough?

Soooo, starting with GPT-5, they're making it more opaque by no longer letting users choose what model they use?

Also, in the process of eliminating model selection, I wonder what kind of pricing will be associated with the subscription tiers for "intelligence levels".

They are pretty clear about the pricing: $0 for the lowest tier, $20 for Plus, $200 for "Pro". They will probably keep that, maybe add some more tiers above that.

So we will soon see according to this price progression $2000 for the most advanced intelligence level starting with $0 for the working class tier.

When did anyone with credibility say that?

the model names are confusing and meaningless to me.

Have we picked all the low hanging fruit of LLMs?

Recent advancements have been mostly about scaling test-time compute, which is much more expensive for companies than the previous paradigm. It makes sense that they'd re-evaluate how to price their services.

Anthropic is ridiculously expensive compared to (not at all)OpenAI. I pay $50 a month to a OpenAI while I spent $50 in a day with Anthropic during a coding session. Was Claude better than Chatgpt 4 at certain tasks? Sure, it was also worse on some.

I've been saying (not at all)Open AI are subsidising the costs of running the models.The moment we get really dependent on them, they'll hike the prices.!

It's not when you compare it to the right thing.

> Was Claude better than Chatgpt 4 at certain tasks? Sure, it was also worse on some.

For simple question/answer things, sure, but then you might as well use Gemini. For price-conscious personal use, its range of models are easily the best option, especially given the generous free tier.

4o-08 is much worse at following instructions and especially using tools intelligently than Sonnet3.5, it's not close. If for you the money isn't worth it, great, but there's a good reason why pretty much every single LLM-based tool is built on top of Sonnet instead of gpt-4o despite it being more expensive - clearly it's worth it. Those tools would love nothing more than to save costs by going with cheaper models.

Prices are only going to go down, Gemini Flash is dirt cheap as compared to GPT 4o mini, and soon Open AI will also hopefully bring prices further down.

They will keep prices high for tech like O3 which is leading the market but probably it will also come down as competition arises and their infra also gets optimised.

Reply on news.ycombinator.com