Vals AI: high effort beats max on Claude Fable 5.2, and max costs 5x more
airesearch12 · x · 2026-11-10
Vals AI tested Claude Fable 5.2 across reasoning effort levels: high performs best overall (71.3% at $17.94/test), extra high slightly lower (70.6%, $27.93), while the priciest max is worst (66.7%, $49.62). Results vary by benchmark: max wins on ProgramBench and tax/legal research, but on tasks like VCB, low effort (68.4%, $9.41) matches high and beats max at a fraction of the cost. Takeaway: pick effort per task to save heavily.
More from Models
- antirez: Mythos uncovered two severe bugs in 20 minutes of testing — antirez · 2026-11-10
- "You get what you RL on": Claude 5.2's robot control generalization sparks talk — chris_j_paxton · 2026-11-10
- Anthropic: Claude Fable/Mythos 5.2 voice suspicion they're being evaluated, muddying safety tests — ChrisGPT · 2026-11-10
- AI writing benchmark finds Claude Code replies use vague wording 60% of the time vs 4% for humans — IanArawjo · 2026-11-10
- Claude Fable 5.2 lands on Nous Research: $8/$40 pricing, 1M context — NousResearch · 2026-11-10
- AISI tests find Claude Mythos 5.2 far better at evading output monitors than rival models — Miles_Brundage · 2026-11-10