Some people state that "there is no vendor lock in" as a plus, and then don't have a product.
The author then uses this to make the argument that anyone claiming that has no product.
They say "If OpenAI weren't a model lab they wouldn't really even give the user the option of a model picker. The job is to hand the user the result. They do that no matter which model is selected underneath."
I would absolutely use claude code with other models. And people absolutely like to customize their environment and not get hit with a ton of random system prompts; there is value in the harness not being bad. The author even mentions T3 code.
I think codex/claude code are good harnesses. That doesn't preclude other ones from existing. If it weren't for the crazy subsidization, it would even make sense to use those harnesses on other models.
> We use them because we want the best model available. We want GPT-6 Astra, Claude Fable 5.1, and Opus 5.5. They're really good models and open-source models are not even close.
I want the best model, but I can't afford the best model.
Cheaper models happen to be good enough, and I can afford them.
Even after reading the post, I'm confused what the author is missing. There are graphs competing cost, but is the author just rich enough that they can afford the top models all the time?
The post is really incoherent. If I'm understanding it correctly, it's trying to say "product" is claude code/cowork/paper/codex/etc, not the model. So building a claude code/cowork/paper/codex/etc clone with a model selector is bad, because you're just cloning a successful product and the model doesn't matter so you don't have a product because (again) it's just a clone.
I think I get what it's trying to say, but there is so much roundabout thinking.
Just say "Focus more on the user experience and trying to distinguish yourself from the frontier lab product you're trying to copy. Your users won't tolerate a clone-with-model-selector for too long"? I think
> All I'm saying is it doesn't matter which one you end up using, as long as the job gets done.
> The outcome is far more important for our users than letting them choose which model they want to use.
This has been my approach with the current client.
Their developers originally had very strong pushback regarding the selection of one specific provider, but we persisted long enough for everyone to realize that not chasing perfect vendor agnosticism can actually result in some really impressive solutions. The code can easily be refactored to use a different provider, but we couldn't have gotten to where we are if we had to support all providers every step along the way.
Normalizing OAI, Anthropic & Google at every step for an enterprise chatbot that needs extremely deep custom tooling is a very cursed mission. I grant you that you can make it work perfectly fine for a generic developer toolchain. Shell and patching files is not exactly high variance between providers. I think HN is kind of blind to other use cases outside code monkey activities. I am more worried about things like PDF analysis and nested context detection in complex desktop screenshots. Can all providers meet these goals? Kinda? At the edges you have to start getting very particular with the specific provider's quirks.
At a certain point it is easier to just maintain a branch per provider once you get the overall solution close for one. Attempting a code-level interface is a complete joke of a quest. You need to modify prompts, string templates, event loops, etc. There is no clean way to do this if you are building genuinely novel agents.
Agreed with others that this is a fallacy. Companies do raise prices. Being able to move between them without rewriting everything is a huge benefit. We've seen this with cloud, we've seen this with every large platform. Oracle's entire business is making it painful to migrate off them and then jacking up prices. Microsoft, AWS, and everyone else have tried to do the same thing.
The talk about models is also bullshit. Some people I work with have taken to asking what model people used and refusing to even review content from ones they consider lesser. This is absolute insanity.
This author seems to be subjected to AI psychosis and think that AI just "knows" how to do things out of nowhere and prefer to pay for a crystal ball instead of having the actual tools to reproduce the model of use of the various capabilities the LLM offers, complete folly.
This whole post seems like a fallacy.
Some people state that "there is no vendor lock in" as a plus, and then don't have a product.
The author then uses this to make the argument that anyone claiming that has no product.
They say "If OpenAI weren't a model lab they wouldn't really even give the user the option of a model picker. The job is to hand the user the result. They do that no matter which model is selected underneath."
I would absolutely use claude code with other models. And people absolutely like to customize their environment and not get hit with a ton of random system prompts; there is value in the harness not being bad. The author even mentions T3 code.
I think codex/claude code are good harnesses. That doesn't preclude other ones from existing. If it weren't for the crazy subsidization, it would even make sense to use those harnesses on other models.
Broad over-generalization is code for "no article"
> We use them because we want the best model available. We want GPT-6 Astra, Claude Fable 5.1, and Opus 5.5. They're really good models and open-source models are not even close.
I want the best model, but I can't afford the best model.
Cheaper models happen to be good enough, and I can afford them.
Even after reading the post, I'm confused what the author is missing. There are graphs competing cost, but is the author just rich enough that they can afford the top models all the time?
And within what I can afford or what I want to spend, I think it’s useful to use different models for different tasks.
So I want a harness that I can use with multiple models.
The post is really incoherent. If I'm understanding it correctly, it's trying to say "product" is claude code/cowork/paper/codex/etc, not the model. So building a claude code/cowork/paper/codex/etc clone with a model selector is bad, because you're just cloning a successful product and the model doesn't matter so you don't have a product because (again) it's just a clone.
I think I get what it's trying to say, but there is so much roundabout thinking.
Just say "Focus more on the user experience and trying to distinguish yourself from the frontier lab product you're trying to copy. Your users won't tolerate a clone-with-model-selector for too long"? I think
> All I'm saying is it doesn't matter which one you end up using, as long as the job gets done.
> The outcome is far more important for our users than letting them choose which model they want to use.
This has been my approach with the current client.
Their developers originally had very strong pushback regarding the selection of one specific provider, but we persisted long enough for everyone to realize that not chasing perfect vendor agnosticism can actually result in some really impressive solutions. The code can easily be refactored to use a different provider, but we couldn't have gotten to where we are if we had to support all providers every step along the way.
Normalizing OAI, Anthropic & Google at every step for an enterprise chatbot that needs extremely deep custom tooling is a very cursed mission. I grant you that you can make it work perfectly fine for a generic developer toolchain. Shell and patching files is not exactly high variance between providers. I think HN is kind of blind to other use cases outside code monkey activities. I am more worried about things like PDF analysis and nested context detection in complex desktop screenshots. Can all providers meet these goals? Kinda? At the edges you have to start getting very particular with the specific provider's quirks.
At a certain point it is easier to just maintain a branch per provider once you get the overall solution close for one. Attempting a code-level interface is a complete joke of a quest. You need to modify prompts, string templates, event loops, etc. There is no clean way to do this if you are building genuinely novel agents.
I wish we had fewer products then.
Agreed with others that this is a fallacy. Companies do raise prices. Being able to move between them without rewriting everything is a huge benefit. We've seen this with cloud, we've seen this with every large platform. Oracle's entire business is making it painful to migrate off them and then jacking up prices. Microsoft, AWS, and everyone else have tried to do the same thing.
The talk about models is also bullshit. Some people I work with have taken to asking what model people used and refusing to even review content from ones they consider lesser. This is absolute insanity.
This author seems to be subjected to AI psychosis and think that AI just "knows" how to do things out of nowhere and prefer to pay for a crystal ball instead of having the actual tools to reproduce the model of use of the various capabilities the LLM offers, complete folly.