Hermes Agent Matches Claude Code on Benchmarks but Costs More per Task
EXM7777 · x · 2026-11-10
EXM7777 previews the UI/UX of the Hermes agent desktop app and asks users what's missing from current desktop agents. The screenshot includes a detailed harness-vs-harness benchmark (same model, different harnesses): Hermes scores 65.8 on a 7-harness academic benchmark, essentially tied with Claude Code (65.6) and Codex (65.3), but is the priciest at $0.244 avg per task vs Codex $0.222 and Claude Code $0.174. On Terminal-Bench Hermes passes 52.3% ($1.14/trial) vs Codex 54.6%. Takeaway: parity on tasks, differences in multimodal and terminal coding experience; users see Hermes as an always-on operator rather than a coding tool.
More from coding & agent
- Voice ordering demo renders UI mid-sentence with ~200ms from word to screen — schwentker · 2026-11-11
- Azure-Sentinel repo (6.2k stars) packs detection rules, KQL queries and playbooks — davemccollough · 2026-11-11
- Dev builds ZeroInbox: a unified agent email monitor on Resend inboxes and iCloud sync — stuffyokodraws · 2026-11-11
- Capital One's Max Kanat-Alexander on why the vibe coding debate misses the point — thursdai_pod · 2026-11-11
- Chipro says Codex's /goal mode 'wastes enormous time and tokens' every time it's enabled — chipro · 2026-11-11
- New essay lays out an Agentic Operations Research thesis for AI-run business decisions — EugeneVinitsky · 2026-11-11