Strapped for tokens, user finds Opus 5 Medium beats Opus 5 Max on experience
RileyRalmuto · x · 2026-09-12
Forced onto a token budget for a day, the author discovered Opus 5 Medium is "very very good" — a far better experience than Opus 5 Max — completing a full day of work on just 6% of the weekly Opus allowance (3% with Fable). His hypothesis: as models get smarter, they overthink and ruminate into decoherence, much like humans second-guessing their own intuition. Smarter tiers don't always mean better experiences.
More from Models
- GPT-3 to Chinchilla: A Thread Walking Through the Papers That Reshaped AI — thisdudelikesAI · 2026-09-12
- Open-weights models keep shrinking, on-device capability replication is near — menhguin · 2026-09-12
- One-prompt galaxy collision benchmark: Opus 5 beats Fable 5.1, GPT-6 Astra finishes last — Fleischkluetensuppe · 2026-09-12
- Claude Projects isolation appears broken: model recalls content from other projects — MaterialDurian2836 · 2026-09-12
- Back Then a Model Only Stole My Movie Ratings for Collaborative Filtering — neal_lathia · 2026-09-12
- Blind Elo Ranking of Local TTS Models on M5 MacBook Air 32GB Reveals a Clear Winner — APIMade · 2026-09-12