Best way to get "model agnostic" coding
Discussion
I'm not a developer, but I'm really getting into 'vibe' coding. I have been using Codex with ChatGPT-Pro and their Astra model with good results.
Whilst that has worked well, I've reached an issue. I feel something serious has changed on my account recently, as it is now burning through my allowance super-quickly. I've tried the usual stuff (not using the high-speed settings, not using ultra, pausing scheduled tasks, lowering the max thinking effort when appropriate), but that didn't help.
I even upgraded from Pro-200 to Pro-500 and I'm still getting the same issue. So I raised a ticket with OpenAI "support" but they can't seem to organise a proper audit on my account to see what the problem is (I've read other people had similar problems which was caused by not being allocated the correct allowance for their subscription). Instead they keep replying with AI-generated-slop from their FAQs, which is very annoying, particularly as I am supposedly speaking with a real person.
So I am wondering how to fix this. And I'm wondering if this is a good opportunity to move to a solution which is "model agnostic", so I can switch to Claude, or Kimi-k3 or whoever else launches a good model, without feeling trapped in the existing structure of OpenAI, which is where all my current history is stored.
As a first step, I installed Hermes, and connected it to my Chatgpt subscription. And whilst it works, it feels much slower when coding to achieve a similar result to Codex. I don't know if that is just a function of Hermes, or whether I am using the wrong settings. But I can also see there are other coding harnesses specific to the task, like OpenCode, and I'm wondering if that might work better?
So my question is has anyone faced a similar challenging of upgrading their coding harness to be model agnostic, and whether they have any tips/advice?
Whilst that has worked well, I've reached an issue. I feel something serious has changed on my account recently, as it is now burning through my allowance super-quickly. I've tried the usual stuff (not using the high-speed settings, not using ultra, pausing scheduled tasks, lowering the max thinking effort when appropriate), but that didn't help.
I even upgraded from Pro-200 to Pro-500 and I'm still getting the same issue. So I raised a ticket with OpenAI "support" but they can't seem to organise a proper audit on my account to see what the problem is (I've read other people had similar problems which was caused by not being allocated the correct allowance for their subscription). Instead they keep replying with AI-generated-slop from their FAQs, which is very annoying, particularly as I am supposedly speaking with a real person.
So I am wondering how to fix this. And I'm wondering if this is a good opportunity to move to a solution which is "model agnostic", so I can switch to Claude, or Kimi-k3 or whoever else launches a good model, without feeling trapped in the existing structure of OpenAI, which is where all my current history is stored.
As a first step, I installed Hermes, and connected it to my Chatgpt subscription. And whilst it works, it feels much slower when coding to achieve a similar result to Codex. I don't know if that is just a function of Hermes, or whether I am using the wrong settings. But I can also see there are other coding harnesses specific to the task, like OpenCode, and I'm wondering if that might work better?
So my question is has anyone faced a similar challenging of upgrading their coding harness to be model agnostic, and whether they have any tips/advice?
alock said:
All AI companies need to grow income by about 20 times just to break even, without increasing costs.
You've been enticed in with massively subsidised prices, it's now time to start paying.
Ha, that first part is definitely true. Some days last week I was using 1-2 billion tokens a day (using some usage resets I had saved up). The cost of buying those tokens via API would be huge. You've been enticed in with massively subsidised prices, it's now time to start paying.
However, rather than "paying up", I prefer to find a way to use a different harness, so I can surf the subscriptions from different providers, without affecting my work-flow
paulrockliffe said:
... Usually the problem is a sprawling codebase so the model has to do loads more work to complete new requests.
I suspect this is part of the issue. I'm not a developer, so I can't structure what it does (in the way a professional dev might do). And it does have a tendency to patch 'issues' by adding complexity which does increase the scale of the codebase, probably often unnecessarily.But something has changed in the last 3-4 days when the usage has rocketed up, even though I'm doing less work, hence why I'm now thinking about switching to something like OpenCode etc.
All the harnesses are different. Some provide you a nice tidy experience but lock you in, others are more generic, some are agnostic and don't tie you in tightly at all.
Not sure any are wildly efficient. Obviously things are better when they can support specific vendor features like their caching.
Some popular ones are degenerating into horrible vibe coded s
t under the skin. Genuinely terrible adverts for unsupervised agents applying updates.
One thing to be wary of is that there can be a lot going on under the surface with prompts and APIs so results can vary a lot by what model you're running against what harness/tool, and also between versions of the tool.
I've just been having the joy of trying to evaluate a shiny new model and it turned out most of the issues were down to everything around it, not the model itself so gave a totally false impression.
Not sure any are wildly efficient. Obviously things are better when they can support specific vendor features like their caching.
Some popular ones are degenerating into horrible vibe coded s
t under the skin. Genuinely terrible adverts for unsupervised agents applying updates.One thing to be wary of is that there can be a lot going on under the surface with prompts and APIs so results can vary a lot by what model you're running against what harness/tool, and also between versions of the tool.
I've just been having the joy of trying to evaluate a shiny new model and it turned out most of the issues were down to everything around it, not the model itself so gave a totally false impression.
https://openrouter.ai lets you access various models from the big suppliers.
Gassing Station | Computers, Gadgets & Stuff | Top of Page | What's New | My Stuff