Kimi K3.1 Pro leaks: 65.7 on Terminal-Bench 4.0 near Opus 5.5 at a fraction of the cost
realmrfakename · x · 2026-10-24
Benchmark charts for Moonshot's Kimi K3.1 Pro are circulating, drawing reactions from @teortaxesTex and others. K3.1 Pro tops SWE Marathon (51.7), Agents' Last Exam (43.2), OfficeQA Pro v2 (64.4), Kimi Code Bench 3.0 (61.2) and Humanity's Last Exam w/ tools (79.0), and scores 65.7 on Terminal-Bench 4.0, just behind Opus 5.5's 66.4.
The cost curve is the real story: at max settings K3.1 Pro hits 65% TB4 for roughly $4-5 per task, far below Claude Opus 5.5 ($13) and Claude Sonnet 5.5 ($22), while DeepSeek-V4.1-Flash-1017 and GLM-5.3 trail badly. Commenters call it somewhat benchmaxxed and 'soft' but note that if it's token-efficient, Moonshot has secured a spot among the live frontier players. A second chart details K3.1 Pro building a human walking-dynamics simulator via large-scale RL, finding and fixing 36 shoulder moment-arm errors itself, and training a simulated body to walk 11.2 seconds continuously.
Related event: Moonshot Releases Kimi K3.1 Pro, Open-Weight Model Rivaling GPT-6(33 posts)→
More from coding & agent
- Dev experiments with autoencoders over functions; low-dim search may be key to local FunSearch — cephaloform · 2026-10-24
- Agent with SSH access successfully upgrades an ancient PHP 5.6 server end-to-end — DanielLockyer · 2026-10-24
- DeepMind researcher demos a playable 3D racing game built right on her desk — thursdai_pod · 2026-10-24
- Cut LLM costs: route requests, disable thinking, cache answers, escalate only on low confidence — colinmcnamara · 2026-10-24
- NVIDIA's OpenShell: Apache-2.0 runtime sandbox for autonomous agent fleets, 18.6k stars — petrusenko_max · 2026-10-24
- Claude Code Routines runs overnight, screens 380 clinics into 30 outreach drafts — Arindam_1729 · 2026-10-24