Testing LongCat-2.5 on Real Long-Horizon Work via Claude Code
FellMentKE · x · 2026-11-04
Blogger FellMentKE argues the better way to evaluate models is to give them a small piece of real work and judge the outcome, not just the code. Using LongCat-2.5 with Claude Code, he tested whether one prompt could produce a page he could open and interact with in a browser. He notes the model's training includes sequences up to 1M tokens, targeting sustained work on complex coding tasks.
More from coding & agent
- Native image reading beats OCR for deep research agents, cutting tool calls — RulinShao · 2026-11-04
- Open Terminal Agent TMax v1.1 Ships With Up to 6-Point Gain on TB2, Fully Open-Sourced — RulinShao · 2026-11-04
- Prized adds GPU rentals to dev boxes; agent rents a GPU overnight — thisguyknowsai · 2026-11-04
- Tuned Harness Beats Newest Model: 9-Month-Old Model Wins by ~10 Points — VictorSanh · 2026-11-04
- Podcast: Building Enterprise AI Platforms — Memory, Guardrails, Observability — Al_Grigor · 2026-11-04
- Manychat engineer to demo a verifiable AI analytics agent, Nov 9 — Al_Grigor · 2026-11-04