Developer Tools

Databricks benchmark reveals GLM 5.2 rivals Opus 4.8 at lower cost

Price-per-token is a poor guide; task-level cost varies wildly.

Deep Dive

Key Points
  • GLM 5.2 achieved top-tier quality (statistically tied with Opus 4.8) at $1.28/task vs $1.94 for Opus 4.8.
  • Sonnet 5 cost $2.09/task despite being cheaper per token, because it used 1.9x more tokens than Opus 4.8.
  • Simple harnesses (like Pi) often outperformed complex ones; models cluster into three capability tiers for routing tasks.

Why It Matters

Teams can cut coding agent costs by matching task complexity to model tier, choosing open models like GLM 5.2.

📬 Get the top 10 AI stories daily