Claude Opus 4.5 Hits 70.2% on SWE-bench While Wiping Context After Every Command
cephaloform · x · 2026-11-08
An experiment shows Claude Opus 4.5 can wipe its conversation after every single shell command on SWE-bench Verified and still score 70.2%, matching its 70.0% full-history baseline while cutting peak context from 130K to under 9K tokens.
The attached NVIDIA paper "You don't need long context" finds generating on long contexts costs 3.6–7x more per token (GLM 5.2: 3.6x, Kimi K2.6: 5.6x, Qwen 3 Coder Next: 7x), argues frontier models barely lose accuracy on long-horizon tasks with one action per context, and proposes semantic state space models (SSSMs).
More from Models
- JEPA seen as 'the big thing' in vision, but insiders say nobody actually cares — gabriberton · 2026-11-08
- Grok 4.7 scores 35, near-bottom on Artificial Analysis as critic slams Grok Bot pricing — Hesamation · 2026-11-08
- Grok on X web starts showing weekly usage limits as free traffic gets reined in — huangyun_122 · 2026-11-08
- NVIDIA researcher confirms work on post-training chat models without breaking strengths — ziqiao_ma · 2026-11-08
- Viral "finest LLMs" tier list ranks fictional future models from Gemini 4 Carbon to GPT-6 — iamfakhrealam · 2026-11-08
- Alibaba's Qoder Launches Stealth Model Qoduck, Free for Individual Users — Aiden_Tech_Ai · 2026-11-08