DeepSeek harness slashes tokens, time and cost vs Codex, but tanks on FrontierCode
teortaxesTex · x · 2026-10-23
TeortaxesTex reviews the DeepSeek Harness (DSH) on Artificial Analysis' Coding Agent Index:
- Strong overall: vs Codex, DSH slashes token usage, turn count, execution time and cost
- Terminal-Bench 4.0: raises scores while cutting tokens (DSH + DeepSeek V4.1 Flash scores 35 at just $76 total cost, 21.4 min/task — a standout budget combo)
- Weak spot: heavy losses on FrontierCode v1.1 (25 vs 56 for the leader); author @tianyi suggested optimizing this
- Leaderboard top: Claude Code + Opus 5.5 (62) at a steep $3.9K total cost; best value clusters around Codex+GPT-6 Astra, GLM-5.3, and Kimi K3 combos
Verdict: already a good harness, with the cost-efficiency route worth watching.
More from coding & agent
- OpenDS-World launches live-web deep search benchmark, frontier models trail humans by large margins — Xiaomin Yu · 2026-10-23
- Naive RAG beats complex agentic memory on 6 of 7 benchmarks, XJTU study finds — XianJiaotongUniversity · 2026-10-23
- AI Assistant Offers to Draft Your Replies and Track Friends Daily — "Sounds Like Hell" — cnakazawa · 2026-10-23
- Dev uses Grok Build to model 140 types of network effects into a Moog-style synthesizer gallery — seanmcdonaldxyz · 2026-10-23
- Why Zig's explicit allocators make it a great fit for building databases — arpit_bhayani · 2026-10-23
- How to build a Company OS with Claude Code and Slack, after 11 days of building — aakashgupta · 2026-10-23