@Zai_org Half my tokens were gone, and not a single one of 10 prompts finished. Every task crashed halfway through with random errors. All I've been doing is filing support tickets. This product is still way too raw.
@LuminaXspace What's the point of speed if the model is dumb as a rock? Especially when there are other models that are only slightly more expensive—or sometimes even cheaper.
@jietang Work for a day, wait a week. Work another day, wait another week. Work half a day, wait a month. That's the whole subscription. What features? Your model burns through tokens so fast it feels like it's charging me even when I'm not using it.
@ornith_ Code generation is the least interesting part. The real test is how well a model debugs existing code.
Most of the job is fixing bugs, refactoring, and navigating large codebases—not writing another Minecraft clone.
@nutlope GLM 5.2 is only cheaper than Opus 4.8 on paper. In reality, it burns through so many tokens that it ends up costing about the same as Opus — maybe even more.
@jun_song GLM 5.2 burns through tokens like crazy. I used 64,490,004 tokens in just 2 days, and now I have to wait 5 days for my quota to reset. In practice, it's not any cheaper than other models because it consumes tokens at an insane rate.
@testerlabor The key isn't the amount of training data, it's the quality. If you're training on garbage code, then even trillions of parameters are basically useless.