Thanks for the kind words, and sorry BYOK wasn't the fit you hoped for. Quick thing though, you don't actually need BYOK to use DM Champ. You can just pay with credits, and we recently introduced our own in-house AI model that's roughly 4x cheaper than Claude and on par performance wise, so you're not forced into a Claude key at all.

On why BYOK is Claude only: it's on the roadmap and we started building it, but through testing we found that in most cases other models work fine and then suddenly don't. We actually used to run on OpenAI ourselves and it just didn't hold up. We tested GPT, Gemini, Llama, Mistral, the whole list across thousands of real customer conversations. For a small project it's fine if the AI works 9 out of 10 times, but once you're running in production a 10% failure rate is way too much, and that's the failure rate we see even with top-tier frontier models from other providers, not even the cheap ones. That's dozens of broken chats a day, hallucinated replies, broken tool calls.

You can't just plug and play and expect it to work, they all have their own edge cases that need handling. We noticed this back when we were still using Cursor for coding, it supported all these models but just switching between them caused weird issues, like a model claiming it added a file when it hadn't. Claude was the best there. We rechecked all the others recently and those issues still aren't fixed. Supporting 100s of models and having them all actually work reliably is a bit of a myth.

"Cheaper" doesn't really pan out either. The other models don't handle our chat volume and context well, they choke on the amount of history, FAQs, follow-up logic, tool calls and reasoning that goes into a single response, so you end up paying for more retries, more tokens and worse output. Claude with our prompt caching gives the lowest real cost per chat, not the lowest sticker price per token. And our own in-house model now comes in around 4x cheaper than that, which we'll keep optimizing to make even better and more affordable.

So instead of chasing 100s of models, we're putting our focus into our own model, fine-tuned on our own data. Long term that's faster and cheaper for you and means you don't have to mess about picking a model at all, you just go close sales.

Wishing you the best either way!