DeepSeek V4 Pro 0813 vs GPT-5.6 Sol on DeepSWE: Cost, Coding, and Routing

In the DeepSWE benchmark a two‑stage approach of running DeepSeek V4 Pro 0813 first and falling back to GPT‑5.6 Sol when the tests fail achieves 83 % success at an average cost of $3.35 per task, compared with 62.8 % at $0.24 for Pro alone or 72.7 % at $8.37 for Sol alone. Although GPT‑5.6 Sol has higher single‑shot accuracy (pass@1 ≈ 72.7 %) and faster rollouts, it is roughly 35× more expensive; DeepSeek V4 Pro offers stronger value when combined in a cascade. The cascade’s cost‑effectiveness comes from Pro’s low unit price while still achieving 60 % lower spend per solved task than Sol.

Source: Together AI Blog | 2026-08-18