Subscription Claude models deliver far fewer thinking tokens, measured five ways
_AustinCalvert_ · x · 2026-10-09
User Lon reports that after Anthropic made Fable 5 permanently available in subscription plans, the model felt noticeably dumber — and measured five different ways, the August version delivered dramatically fewer thinking tokens than July. The retweener argues logging actual thinking-token counts would shock users, that delivered tokens fall far short of published benchmark numbers, and calls silent token reduction the top method vendors use to nerf models and protect compute margins.
More from Models
- Dev complaint: models are still horrible at code reuse — snowclipsed · 2026-10-09
- Perplexity Decider tops DecisionBench as the lowest-cost decision model — victormustar · 2026-10-09
- LightOnOCR-3 released: 0.8B/1B/4B open models for OCR, layout, charts under Apache 2.0 — antoine_chaffin · 2026-10-09
- Perplexity Decider tops DecisionBench with 93.9% accuracy, 534ms latency, and lowest cost — AravSrinivas · 2026-10-09
- Google has been leaking the Argon codename in plain sight since late 2025 — Arteux · 2026-10-09
- User's math prompt to Claude gets blocked by safety system: "Am I about to get swatted?" — Super_Range45 · 2026-10-09