A comparative analysis of GLM-5.3 and GPT-5.6 Sol on the DeepSWE benchmark reveals that a cascading routing strategy outperforms either model used in isolation. By running GLM-5.3 first and escalating to GPT-5.6 Sol only when tests fail, the system solves 85.9% of tasks at $6.61 per task.

  • GLM-5.3 costs $3.99 per rollout compared to Sol's $8.37, making it 2.1x cheaper and yielding 17 solved tasks per $100 versus Sol's 9.
  • GPT-5.6 Sol leads in single-shot accuracy (72.7% pass@1) and reliability (61 tasks solved four-for-four), while GLM-5.3 ties at pass@2 and leads at pass@4 (87.6%).
  • The models diverge significantly with a 0.43 per-task correlation, covering 106 of 113 tasks between them, which enables the effective cascade.
  • GLM-5.3 exhibits cleaner failure modes, breaking existing tests in only 11% of failures compared to Sol's 20%, reducing the need for heavy regression gating.

This approach leverages GLM-5.3 as a lower-cost front end and Sol as a verifier-gated escalation, achieving higher coverage than Sol alone while remaining cheaper than a single Sol rollout.