Miles Brundage: Fable 5.1 is far more token-hungry in real workflows than in evals
Miles_Brundage · x · 2026-09-03
Miles Brundage notes that while Fable 5.1 is token-efficient on well-defined evals, its default behavior in actual workflows seems to produce many millions more tokens than version 5 did — a notable gap between benchmark performance and real-world usage cost.
More from Models
- Pangram claims third-party validation as only reliable AI text detector, disputed — jackinwarsaw · 2026-09-03
- The AI release window is no longer 24 hours — it's a full week and accelerating — eyishazyer · 2026-09-03
- Meta's chief AI officer calls Muse Spark 1.3 competitive with Fable 5.1, beats GPT-5.6 Sol — eyishazyer · 2026-09-03
- NVIDIA's 75B hybrid MoE Nemotron-3-Puzzle is now runnable locally in llama.cpp — jacek2023 · 2026-09-03
- Dev: GPT models lead at computer use, Anthropic far behind despite Fable's coding lead — CtrlAltDwayne · 2026-09-03
- GPT-6 spotted in branch name, rumored to focus on computer use capabilities — scaling01 · 2026-09-03