Updated
Updated · Databricks · Jul 8
Databricks Benchmarks AI Coding Agents on 10-Plus-Language Codebase, Finds $1.28 GLM 5.2 Near Top Tier
Updated
Updated · Databricks · Jul 8

Databricks Benchmarks AI Coding Agents on 10-Plus-Language Codebase, Finds $1.28 GLM 5.2 Near Top Tier

3 articles · Updated · Databricks · Jul 8

Summary

  • Databricks said its internal benchmark of AI coding agents on real engineering tasks found three capability tiers and showed no single vendor dominates the cost-performance frontier.
  • GLM 5.2 emerged as a standout open model, statistically tied with Opus 4.8 on quality while costing $1.28 per task versus $1.94, prompting Databricks to consider it a daily-driver option.
  • Task-level costs often diverged from token pricing: Sonnet 5 was about 1.7x cheaper per token than Opus 4.8 but cost $2.09 per task versus $1.94 and scored 81% against 87%.
  • Harness design also materially changed efficiency, with Pi cutting cost by more than 2x in some runs by sending about 3x less context per turn while preserving quality.
  • The benchmark was built from recent human-written pull requests across a multi-million-line, 10-plus-language codebase, with sealed git history and test-based grading to avoid leakage and overfitting.

Insights

How are open-source AI tools now outperforming tech giants on cost and quality?
Why do cheaper AI models often end up costing more for real-world coding tasks?
Is the framework controlling an AI more crucial than the model itself for efficiency?