DeepSeek V4 Pro hits OpenRouter with notable output variance across reasoning levels
Simon Willison · rss · 2026-08-13
The latest DeepSeek Pro model is now available via API on OpenRouter. The author's tests reveal a fascinating behavior: the model produces significantly different outputs (like a 'pelican riding a bicycle') across its low, medium, and high reasoning levels—a variance rarely seen in other models.
Interestingly, the model's benchmark scores had a chaotic dissemination path: initially shared in an official WeChat group, copied to Reddit where it was deleted by mods for being 'low-effort', and finally preserved as an ASCII-art table on Hacker News.
More from Fun
- AI circles share a collection of cursed 3D face renders and game glitches — repligate · 2026-10-16
- Sabine Hossenfelder fires back: academia's impact metric turned physics into a paper mill — skdh · 2026-10-16
- Meme: Clicking 'Yes to all' on Claude Code without reading anything — jnack · 2026-10-16
- Minecraft AI mod turns region selection + prompts into autonomous building agent — Acclynn · 2026-10-16
- Throwaway prop code in movies is getting eerily accurate, and AI is why — mike64_t · 2026-10-16
- Distillation copies logits, not hidden CoT — open-sourcing 1.3 leaks IP, not reasoning data — teortaxesTex · 2026-10-16