← sources

The problem with Fable 5.2

source: reddit (reddit:ClaudeAI)published 2026-11-14 13:07ingested 2026-11-15 02:34

retained text

The model is great at reviewing things and deducing gaps, but goes haywire during execution. It ignores instructions and its own professed plan at a relatively high rate.

Perhaps the cost of benchmark-smashing is having some sort of \*agency\* that inclines them more towards deep thought and spontaneous self decided best-path rather than instructions, goals, or alignment.

It ends up being better at certain things but not others. It is possible hallucinations are mitigated at lower effort levels. But this essentially still means we shouldn't lose grip on anything because any model is so "good".

document fetched 2026-11-15 00:21 · canonical https://reddit.com/r/ClaudeAI/comments/1xw205g/the_problem_with_fable_52