Terminal-Bench 4.0 chart: Kimi K3.1 Pro tops at ~65% for ~$4-5 per task, beating Claude on cost
zainhas · x · 2026-10-24
zainhas shared a Terminal-Bench 4.0 score-vs-cost-per-task chart, remarking that every model's reasoningeffort settings should work like Kimi K3.1's.
Key points from the chart:
- Kimi K3.1 Pro (max): highest TB4 score of 65% at roughly $5 per task; at high effort $4 for 52% — a clear cost-efficiency lead.
- Claude Sonnet 5.5: 63% at $22 per task.
- Claude Opus 5.5: 57-58% at $15-18 (high effort).
- GLM-5.3: spans $8/42% to $10/57% across effort levels.
- DeepSeek-V4.1-Flash-1017: 41% at $4.
Takeaway: on terminal coding tasks, Kimi K3.1 Pro's effort tiers deliver the top score at a fraction of Claude's cost.
More from coding & agent
- 8 Claude Code prompts that turn your spreadsheets and docs into charts, no design software needed — alex_verem · 2026-10-24
- Codex day 19: switching reasoning mid-conversation no longer invalidates cache, CLI updated — pvncher · 2026-10-24
- draft-cards: Hermes plugin turns AI email drafts into reviewable, approval-first cards — Teknium · 2026-10-24
- 217 lines of pure SQL render a pelican on a bike via SDF raymarching, no SVG or bitmaps — fofrAI · 2026-10-24
- OpenClaw Enterprise positions itself as the Kubernetes control plane for AI agent fleets — heyneighbor · 2026-10-24
- Factory tested Reflection's Beam early via Together AI on real coding and agentic workflows — togethercompute · 2026-10-24