Descriptions:
GLM 5.2, the open-source model from THUDM, has generated significant attention for bringing near-frontier performance to local and cloud deployments. In this episode of the Greg Isenberg Show, host Greg Isenberg sits down with practitioner Amir to break down exactly how to integrate GLM 5.2 into popular coding environments including Cursor, Codex CLI, and Claude Code — all routed through OpenRouter — and why the model represents a meaningful step forward for open-source AI.
Amir walks through GLM 5.2’s key specifications: a 1-million-token context window, an 81-point score on TerminalBench 2.1 (just four points behind Claude Opus 4.8), and a 62.1% score on long-horizon task evaluations versus Opus’s 69.2%. A standout cost comparison: generating comparable output to Opus 4.8 runs roughly $0.44 via GLM 5.2 on OpenRouter versus $2.38 — nearly a 5x difference that compounds when running continuous automation tasks. The model can also run entirely locally if hardware supports it, though it is resource-intensive.
The episode also explores model chaining — what OpenRouter calls “fusion models” — pairing a heavier reasoning model for planning with GLM 5.2 for execution. Amir is candid about current limitations around image understanding and tool-use modalities, and offers practical workarounds. For developers building startups or looking to reduce AI infrastructure costs without sacrificing capability, this is a useful ground-level primer on where open-source models stand today relative to closed alternatives like Claude Opus or GPT.
📺 Source: Greg Isenberg · Published June 23, 2026
🏷️ Format: Podcast
