Yes, I mainly work with OpenCode, with GLM as an orchestrator agent and DeepSeek Flash as a worker agent. I’m assuming it’s modern? I don’t have the money to use either Fable or Astra, but my assumption is that the few more points they get on the benchmarks still don’t make up for the wild confabulated things they do in my projects.

I can't speak for the other providers but Gemini carbon feels like the difference between gpt3 and gpt5 when compared to Chinese open models.

Both are fine, though not frontier of course. They mostly take unwarranted shortcuts, trying to find the easiest way to reach the literal goal set by you, failing to see the broader picture. But they definitely do not confabulate, as in they don't just make up nonexistent things, they would rather walk extra mile to verify when they're not certain, which is the reason the software they make is trash but it works in the narrow sense. They can be rigid in their writing style do to mode collapse, but I haven't seen this leading to them inventing nonexistent functions or anything like that in a long while. RL makes them do all this, modern LLMs are very far from being next token predictors trained only on text.

> But they definitely do not confabulate, as in they don't just make up nonexistent things

I've definitely seen both Labs' "best" stuff confabulate to this day.

Agreed.

The harnesses / RLHF do a good job of prioritizing mechanisms to check results, which works out pretty well pragmatically for programming.

Nonetheless, I see Opus 5.5 spew out false information daily (never mind its code's unreadability).

I can't speak to Fable or Astra, because my employers, while threatening to fire us if we don't use the bullshit bots, have also refused to light the barrels of cash on fire that are required for those models.

As far as I can see, their fundamental nature has not changed, and truth is not something they have access to - just the called tool's exit code.