This feels like a particularly good moment to start building with open models.

@boltdotnew just put GLM, DeepSeek, and Kimi into Forge with up to 50x more usage.

The model picker is what makes this more useful than simply adding another provider.
You can keep the project and execution environment roughly fixed, change the model, and watch where behavior diverges.

With a much larger token budget and no daily meter, I can ask for another implementation, or compare 2 approaches or throw away the bad ones.