Single Forward Pass Evals on Fable, Opus 5, and GPT-5.6-Sol

Researchers tested newer frontier models—Fable Five, Opus Five, and GPT-5.6-Sol—on reasoning tasks where they had to solve math and logic puzzles in a single neural pass, without revealing their reasoning. The results, per LessWrong, showed substantial jumps over older baselines. Fable Five reached eighty-seven point six percent accuracy on arithmetic problems, well above the roughly sixty percent previous record. GPT-5.6-Sol climbed from fifty-eight point six percent to eighty-three point four percent when given extra tokens. Surprisingly, even meaningless filler tokens—just counting sequences—boosted performance, suggesting these models can perform hidden reasoning without exposing it in their visible output.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton