Together AI compared DeepSeek V4 Pro 0813 with GPT-5.6 Sol on 113 software-engineering tasks from DeepSWE. Each model received four runs per task, producing 904 rollouts in the August 18, 2026 study.
GPT-5.6 Sol's single-attempt pass rate was 72.7%, compared with 62.8% for DeepSeek V4 Pro. Sol also finished a typical run faster in Together's measurements, with a median of 17 minutes against 35 minutes for Pro.
The price gap went the other way. Together reports average costs of $8.37 per Sol rollout and $0.24 for Pro in the tested configurations. With up to four tries per task, Pro reached 88.5% pass@4, ahead of Sol's 85.8%. Pass@4 counts a task as solved when at least one of four attempts passes, so it does not describe the reliability of a single attempt.
Together then evaluated a simple routing policy: send a task to the cheaper Pro model first, run the tests, and escalate to Sol only after a failure. This test-gated cascade solved 83.0% of tasks at an average $3.35 per task, according to the study.
The results apply to these model variants, settings and DeepSWE tasks. The routing policy depends on tests that can reliably decide whether an attempt succeeded.