Why I prefer using LLM API aggregators over subscription services
Note: I have no affiliation with Cortecs or open-webui. This is just my opinion.
An LLM API aggregator is a platform that provides a single, unified API endpoint to access multiple large language models from different providers (like from OpenAI, Anthropic, Google, Mistral, Alibaba, Z.ai ...). This lets you switch between models or route requests without managing separate accounts, keys, or SDKs for each one. Services like OpenRouter, Cortecs and so on...
Before I switched to Cortecs, I used numerous subscription services who charged anywhere from 8-25 eur.
I just finished my first month at Cortecs. Do you know how much I spent in that month? Exactly 0.40 eur. And that's only because at the start I wasted a lot of tokens experimenting and figuring stuff. It would probably be closer to 0.30 eur or even lower. Now yes yes ... I have to "self-host" the front-end to access it and perhaps I wasn't as active as I had some other things to attend to... But it will still be not even close to what I used to spend and over the long-term I'm guaranteed to save quite a lot of money.
Do you know why ? Because I am more aware of each prompts cost and choose to reduce my usage. People spend more with a card than with cash. That's because they see the money that they spend directly. It's the same with LLM usage. If you do a prompt on an expensive model with a filled up context window you will know it costs a shit ton because you will go to Cortecs console and see the total cost go up. THE COST IS NOT ABSTRACTED AWAY FROM YOU!
These LLM services, on the other hand, hide that from you ! And you're more likely to use because you're paying for it. Even if you don't use it at all this month you will pay the same. API's are pay as you go !
This motivates me to not over-rely on them, to use cheaper, smaller, open-weight models, to send less useless context and to use my brain instead of asking stupid ass questions! In other words, it pushes me to use them RESPONSIBLY.
Now of course there's are downsides and I will not pretend it's a perfect solution for everyone.
- Requires more technical know-how to set up. You do need to run your own front-end. But it's not that difficult. All I had to do was pull a docker image, run the container and then write my API key from Cortecs. The docs made it easy for the most part. I run it on my regular PC and if I need LLM's on my mobile phone when PC is off I just use Mistral. You can self-host it on a server, obviously.
- Might be bad if you're abusing subscription prices and using 100$ worth of tokens on 20$ / month plan. It's not sustainable / responsible and I would recommend not doing that but hey what do I know.
- You might not have some of the fancy tools that these services might offer. Personally, I never used them.
Yesterday I had to do some things on my Linux machine. Naturally I used the help of LLM's for some things to clarify and so on...I solved the thing I was working on pretty smoothly, it wasn't that difficult. Cortecs shows I did 31 requests. Total tokens : 9.5K. Total Costs: 0.00 eur. I didn't even spend a cent!
How?
I used small open-weight models, kept conversations short, only sent relevant context, didn't over-rely on them and asked to keep the answers "short, to the point". The thing is, I don't even feel like "I'm missing something". These models like Qwen 3.8 flash-next, Gemma 4 26B, Deepseek v4 flash are seriously good!
I cringe hard thinking how a year or two back I used Claude Sonnet for trivial stuff that even a 9B parameter model can answer. I still cringe at some of the agentic coding workflows and vibe coding nonsense that I see online.
Using small, open-weight cheap models is usually not possible if you're on a ChatGPT subscription. And even if you could, why would you? You will be paying the same anyways. Do you know the sizes of the models that you use there? What data center is the model hosted on ? Is it powered by renewables or gas turbines? Do they write how many tokens your prompt burns through? No, why would you care about such things? That's the problem.
It's hard to use LLM's responsibly when the service discourages it.
Anyways, that's all. Just putting it out there for folks that have lowish usage but spend 20$ on subscriptions and who want to use LLM's more responsibly.