AI performance comparison and tool monetization

[FrankSzendzielarz]

Hi and first, @kooshi , thanks so much for building this! I was looking for something like this quite some time ago and surprised it wasnt already shipping from MS. In fact as far as I can tell VS 2026 works with some remote LSP and doesn't fully have this kind of semantic aid yet (though I am not sure on that), and is quite new. I had been planning to build this myself, as I did a lot of work with Roslyn a couple years ago and didn't mind diving in. Only last week did I finally get the time to look into it and found your tools here instead.

I use Claude Code a lot, most recently over Avalonia and .NET Core server projects, and the two pain points for me are 1) the file I/O method of working with code is token expensive, as things often get restructured, renamed, reworked and Claude does this without IDE integration 2) Context window limits often result in Claude forgetting architecturl patterns, even when prompted, because it forgets that certain areas of the code already handle something using pattern it does not recall.

So one thing I wonder about, is what benchmarks exist to compare the performance of these coding models given the various levels of IDE or AST integration. So far, as far as I can tell, similar attempts from JetBrains and Cursor work differently, with different limitations. Also, it seems that benchmarks are limited for C# and .net projects, and the popular benchmarks mostly file io driven python codebases. There is a recent C# benchmark I found online, which I would need to look at. (Quote from Claude : Microsoft Research published the first C# benchmark just a few months ago (November 2025): C# (#5 in TIOBE) had no coverage in any existing benchmark. Microsoft They created SWE-Sharp-Bench with 150 tasks from 17 repositories.)

It would be, I think, therefore worthwhile collaborating on using this tool to try and get the benchmark comparisons for with and without semantic model integration. The results could bring significant interest.

So that brings me to monetization. I am just in the process of releasing and trying to garner ML model developers for an early startup, ApiCharge.com, with the first release focused on ML model developers wanting to publish their models as monetizable services, taking all the hosting, pricing and financials pain out of going from training to productionisation. The intro video ApiCharge.com says a bit, but the first technical video on the left is probably more informative at this page here https://mainnet.stellar.apicharge.com/Welcome

In a nutshell, the core of the product is a stablecoin monetizing network proxy that comes baked in with all kinds of features for pricing control, payment processing and so on. Add it into any network architecture, including serving streams, and it monetizes. This I am using to build a library of Docker images of popular servers, initially focused on ML model host servers, with ApiCharge baked in, to democratize API monetization of all kinds. The tooling and client apps are integrated with low cost global GPU marketplaces (5x cheaper than RunPod for example), so users can in a few clicks deploy a host, deploy a server, deploy their model and configure pricing. One feature of ApiCharge is that you can configure per-hour , or per day , or per anything pricing, so pay per use models can be delivered easily.

What I propose then is that if the benchmarks work, to build an ApiCharge Docker image with the tooling built in and a Qwen coder model, on the assumption that the semantic model augmented AI model does much better, delivering Claude Opus like performance at lower cost. Not just in terms of direct code editing, but baking in features like a supervisor agent that is responsible for handling architectural conformity (automatic PR or git commit hook agent) based on models and patterns easily compared with the semantic model. A kind of simple toolchain that makes it easy to deploy either for personal use or for communities commercially, or publicly commercially, with a simple UI to get started with dev quickly. It's the kind of thing that could, if it worked, be like a Cursor for .NET.

Revenue would derive from either the txn fee I currently collect, or directly from the servers that people host and monetise themselves (the Docker images are public). Further, the toolchain produced could be customised with pre made prompts for say, UI dev in XAML etc, and people could publish/deploy their own flavours. Lastly, ApiCharge implements a marketplace where services can be published and discovered, then consumed using the management apps.