Claude Opus 5 flips between 1/3 and 2/5 before finally settling on an answer
Sauers_ · x · 2026-07-25
- The screenshot shows Claude Opus 5 working through a probability puzzle and repeatedly changing its mind before settling.
- It cycles through multiple candidate answers, including 1/3 and 2/5, and keeps reconsidering from scratch.
- The point of the clip is not a final benchmark result but the model’s visibly unstable internal decision process.
More from coding & agent
- A GitHub repo gathers 100+ free open-source AI agents and RAG apps, with 127K stars — Saboo_Shubham_ · 2026-07-25
- Teams strip 80% of Claude Code’s system prompt in a new context-engineering guide — EricBuess · 2026-07-25
- A blunt rebuttal says autonomous agents are being oversold for long-horizon hacking tasks — ctjlewis · 2026-07-25
- Opus 5 model card shows 5-agent coding teams reach 0.6 score 2.2× faster — OfirPress · 2026-07-25
- Devin adds Claude Opus 5 as FrontierCode 1.1 shows near-Fable performance at half cost — _sholtodouglas · 2026-07-25
- A week using Claude Code, Codex, and Gemini CLI showed the same repo-breaking patterns — AIcademy-academy · 2026-07-25