Claude Code caps Kimi K3.1 at 200K: 148 compactions across 95 trials vs Opus's 797K
zainhas · x · 2026-11-22
Developer zainhas analyzed peak context usage across 66 tbench4.1 trials and found Kimi K3.1 Pro keeps auto-compacting at 200K inside Claude Code.
- Kimi natively supports 1M tokens, but Claude Code treats it as a 200K model: 148 compactions across 95 trials, with a median trigger around 168K.
- By contrast, Opus 5.5 had zero compactions, with a median peak of 290K and usage up to 797K.
- The author asks whether disabling auto-compaction would improve the model's performance.
More from coding & agent
- Gemini 4 Neon builds an interactive 3D neural character demo, no NVIDIA GPU needed — Eric520CC · 2026-11-22
- RLVR trains models to maximize task reward — long-term code maintainability is on you — bendee983 · 2026-11-22
- Free online Lean 4 book packs 23 chapters from BM25 ranking to local LLMs with Ollama — danbri · 2026-11-22
- Engineer says he independently arrived at the same 'knowledge factory' design across domains — hwchase17 · 2026-11-22
- Running Inkling 2 Mini locally on Mac: 20GB model, ~16 tok/s on M4 Pro — jacob_posel · 2026-11-22
- Stanford CS329Z Guest Lecture: Alex Shaw & Ryan Marten on Evaluating AI Agents with Terminal-Bench — stanfordnlp · 2026-11-22