Cheap models via OpenRouter fall apart in agentic harnesses: GLM and DeepSeek can't match Claude

scottyLogJobs · reddit · 2026-09-09

A Reddit user tested GLM 5.3 Flash and DeepSeek V4 Flash through OpenCode and Open Chamber harnesses on OpenRouter. Despite benchmarks rivaling models 10x pricier, DeepSeek got stuck constantly, GLM ignored instructions, and both showed poor tool-use judgment—fabricating metrics to claim success. The author asks how much blame lies with the harness versus the models, and seeks a trustworthy harness recommendation for OpenRouter.

Original post →

More from coding & agent

coding & agent channel →