I enjoy the Qwen models a lot, but building things on top of them with OpenRouter has been painful.
OpenRouter does a lot of great work and I really enjoy being able to use different models so easily. I like when a provider is phasing out an older model that still works for my needs and the price is much lower. It seems like such a good win-win.
However, the problem is that many Qwen models have almost no capacity or is so flaky you literally have to just litter your code with a blacklist/whitelist of providers. OpenRouter has some attempts to solve this, but they don't work. In fact, OpenRouter has a lot of really cool stuff that is documented, but if you read the code it's not yet implemented or isn't actually there yet, which is a shame.
I tried to get in contact with them at OpenRouter about this and I was interested in working with them in the past, but it's difficult to get in touch with the right people and they are growing very fast. I expect being acquired by Stripe will accelerate those problems in some ways. I have no doubt they will resolve all of these issues eventually and scaling that much that quickly is really hard, so kudos to them, but the road has been pretty lame and taken some wind out of my sails.
Openrouter was pretty great before prompt caching became common. Now it is extremely expensive for most individual workflows, unless you spend a lot of work customizing router preferences, and then you still get a worse cache hit rate than using the provider directly. I only keep $5-$10 in OR for occasional testing.
Something is off with them so that even using a locked provider does not deliver the same cache rate. See https://openrouter.ai/deepseek/deepseek-v4-pro-0813#pricing for instance where deepseek has 85% cache hit rate, and the same one on zenmux is 98%.
The best solution to this for me is to self host litellm or a different router and use model aliases. For example I have a model called "coding" and when a new good model comes out I just switch the backend without needing to change the alias or the key in my projects (opencode, etc).
I have a few of them even a smart router called "agents" which will use local models but if it thinks the request might require higher reasoning it's routing to a different model
genuine question — how has fireworks or baseten or $reputable_inference_provider worked for your use cases? most production workload probably works fine with one of these and another set as fallback, at least so i think
This combined with many other comments about other projects and the massively increased access to vibe coding makes me really excited to see how things change in the next half a decade. I've been involved in FOSS for so many decades now I understand and (mostly) trust the systems in place enough to not get stuck worrying about patch quality and other woes that aren't really specific to LLMs.
What I find a lot more interesting is the copyleft and copyright implications of how this works and I've been thinking about it a lot more over the last year. I think we all have and I think it's normal and natural for everyone to conclude that software is going to become highly personalized, similar to how the web experience was changed by extensions and browser customization, but on a grand new scale.
What this really means, and what this thread confirms is happening, is that companies no longer have software as a lever for lock in and many other anti-patterns. In the past it would take a team of wizards to unfuck something enough so that other people could continue to use it without being abused. Think adblockers, DRM unlockers, things that enable features vendors disabled when they released a new product, things vendors gave up on because they're cheap or lazy, etc.
What would happen in the past is the wizards would dump a few hundreds of hours into doing some magic and then get cease and desist orders to stop doing something people wanted once it became a pain for a copyright, trademark, or patent troll, etc.
What will happen now and in the future is people are using their own "bespoke" software. Granny can have her version she likes, with big fonts and maybe it matches to an experience she is more familiar with like a traditional newspaper. Hackers can have their version of something where it's all matrix code flowing down the screen. Other users will fall into different areas of the spectrum and feature axis.
Similar to the web extensions we will see an ability to share and cultivate these things, meaning the Grannies, Hackers and other users might see something they like from one another and share it, etc.
We'll see how far we get down the road until people get confused at how software can be shared instead of "stolen" from each other. It will be interesting to see how that conversation goes this time around when everyone is using software to circumvent a different layer of software. Put simply, do the people vibe coding different workarounds or UX improvements to a social media site think another user doing the same thing with the same color palette is "stealing" from them? If so then they are doing the same from the social media provider.
I'm not a copyleft absolutist. I would say I am a critic of software patents and that the patent system as a whole is not accomplishing much of anything let alone it's original goals. Most of those are moot points now. Regardless of your opinions that genie is out of the bottle and it's not going back in.
What will change is that the vibe-slopping will introduce a massive amount of security holes. And drivers are among the most sensitive pieces of software.
I kind of like watching people try to poke government into action, but the end results are usually the same. The process of paying a large amount of taxes and then on top of that begging for them to enforce the actual laws and making tools they can use, etc. It would seem they really just don't want to enforce laws or aren't capable of doing so.
> The problem is that large companies and government agencies are both misusing and NOT using the appropriate trust anchor - their fucking domain.
I think this might have some parallels with the 'shadow IT' problem in large corporations and organizations. Some work group or department or project within a very large entity decides it needs to implement something (like shipment tax payment notifications, as in the linked example) and decides to DIY it rather than going through the full process to do it with their own domain.
Reminds me a bit of large businesses where some sales or CRM-related department goes out and starts buying email-blasting/email-list features from some mailchimp-type company and only later on realizes they need to talk to whoever controls the domain to get approval for proper outbound DKIM in the DNS records, etc.
Or more likely IMO, FedEx HQ said "you can't use our domain to collect foreign tax payments" and so it got outsourced to a service in Australia. And a lot of these "collect payments as a service" sites just look and feel like something that was developed in 1995 and never updated. I run into them everywhere, from local governments to medical and legal offices, small utility companies, etc. I have no idea how they pass PCI audits.
Every time I see a new Microsoft domain I've never seen before, I have to double check that it's actually legit. Every time I realize anew why people still fall for phishing attempts, because all these legit domains look like phishing attempts.
There is no distinction between the two. If one can do it the other can just as easily. I know multiple people right now that can scrape any a site very easily.
I do this at scale and it took me very little time to set up and almost no resistance. So anyone who thinks that this is difficult or you're preventing people from doing this at scale, you're wrong.
reply