Claude Code ranks last on Terminal-Bench 2.0 as simpler harnesses beat it
OwariDa · x · 2026-11-16
OwariDa rebuts claims about Claude Code's leaked source with benchmark data. The leaked 3,167-line runHeadlessStreaming function in print.ts is just ops plumbing for headless/SDK mode (JSON pipes, rate limits, MCP lifecycle) — not a symbolic reasoning kernel. More notably, Claude Code is a weak harness: on Terminal-Bench 2.0 with the same Opus 4.6 model it scores 58.0%, last among 10 entries, while Terminus 2 — which only sends keystrokes to a tmux session — scores 62.9%. Simpler harnesses clearly outperform it.
More from coding & agent
- Full YouTube walkthrough of solo operator's Business Brain AI workspace — aryanXmahajan · 2026-11-16
- A 32KB, no-engine, no-polygon 3D demo with falling objects goes viral — IgorCarron · 2026-11-16
- Dad Builds 3D Puzzle Game for His 5-Year-Old With Fable 5.2: 30 Levels in One HTML File — JohnnyFFM · 2026-11-16
- User cancels $147 of subscriptions in a week with Grokbot, which pays and enters credentials — DominiqueCAPaul · 2026-11-16
- Frustrated researcher codes his own procurement, accounting and HR suite for free — arjunrajlab · 2026-11-16
- Google Research: Repeating the Prompt Twice Boosts LLM Accuracy from 21% to 97% — udmrzn · 2026-11-16