Singularity Gate: Fable 5.2 tops benchmark at 25.8%, but all models stall near 8% on the hardest tasks

queenofartists · reddit · 2026-11-11

Singularity Gate benchmarks whether models can predict paradigm-breaking scientific discoveries made after their training cutoff. New results:

Claude Fable 5.2 leads at 25.8%, burning 28.8% of a week's limit on the $200 plan; API price cuts make it 20% cheaper than Fable 5.1 (fewer output tokens but more context read)
On the 12 Hard tasks — the most paradigm-breaking discoveries — five models cluster between 7.6% and 8.0%, with Opus 5.5 level with Fable 5.2, and Sol level with Astra
Fable 5.2's gain over 5.1 on Hard tasks is within standard error; the author speculates the two were distilled from different Anthropic internal base models (Model 2 / Model 3)
Core concern: as long as the base model behind public models stays the same, bigger models don't improve on truly out-of-distribution science. If the next new-base model also shows no gain, that's a worrying sign for AI in science and the path to the singularity
Per-field: Fable 5.1 answered only 3 of 14 Life Sciences tasks

Full results at singularitygate.org, with public repo and paper.

More from Models