GLM-5.2 vs Opus 4.8 in Claude Code: Near-Parity Output at a Fraction of the Cost

Key Takeaways

• Across 40+ naively-prompted creative scenes, GLM-5.2’s output often looked cleaner than Opus 4.8 — Saraev argues benchmarks are saturated, so you have to judge by "taste."
• GLM struggled most on interactive mini-games (physics and difficulty tuning); Opus kept an edge there.
• Run GLM-5.2 in the Claude Code harness by pointing the ANTHROPIC_BASE_URL variable at an OpenRouter (or Z.AI) endpoint — same harness, different model under the hood.
• GLM has no built-in web search; wire in Exa AI ("set up exa.ai for web search") to give the agent browsing.
• Four cost routes: Z.AI’s coding plan (light/pro/$80-a-month max), OpenRouter pay-per-token (his pick), dedicated hosts (Fireworks / DeepInfra / GMI), or self-hosting a heavily-quantized 2-bit build (~82% accuracy) on a 256GB Mac.

Commands & Code Mentioned

export ANTHROPIC_BASE_URL=<your OpenRouter endpoint>   # point Claude Code at GLM-5.2
brew install opencode      # run GLM-5.2 in the Open Code harness
brew install crush         # run GLM-5.2 in the Crush harness
glm                         # launch the Claude Code instance wired to GLM-5.2
# prompt: "set up exa.ai for web search within this GLM directory"