A comparison of DeepSeek-V4 Flash 0731 and GPT-5.6 Luna on the DeepSWE benchmark reveals that while GPT-5.6 Luna is the stronger engineer, a cascade strategy using both models achieves higher accuracy at lower cost.

  • GPT-5.6 Luna leads DeepSWE pass@1 decisively at 67.2% versus DeepSeek's 53.3%, maintaining its lead at every equal attempt count.
  • DeepSeek-V4 Flash is the cheapest model on the board at $0.10 per rollout, delivering 532 solves per $100 compared to Luna's 110.
  • Running DeepSeek first and escalating to Luna only on failure solves 78.9% of tasks at $0.385 each, beating Luna alone in both accuracy and cost.
  • DeepSeek fails more cleanly, breaking the repo's existing test suite in 9% of failures versus Luna's 15%.

The cascade approach leverages DeepSeek's low price to clear roughly half the tasks, allowing the flagship model to focus on harder problems for a combined result that outperforms either model used alone.