DeepSeek V4 Pro 0813 vs Claude Fable 5 on DeepSWE: Cost, Coding, and Routing

Together AI published results from 904 DeepSWE rollouts comparing DeepSeek V4 Pro 0813 (max) and Claude Fable 5 (max) across all 113 DeepSWE tasks, four trials each, using published per-trial records. DeepSWE is described as a benchmark testing software engineering ability across task types and programming languages.

On first-attempt accuracy, Claude Fable 5 leads at 69.7% pass@1 versus DeepSeek V4 Pro 0813's 62.8%, a 7-point gap. The lead narrows under retries: the two are level at pass@2 (78.5% versus 77.1%), and Pro leads at pass@4 (88.5% versus 84.1%).

Cost differs sharply. Together AI reports $0.24 per rollout for Pro against $21.63 for Fable, a 90x gap, or 260 solves per $100 for Pro versus 3 for Fable. Median rollout time is roughly even: 31 minutes for Fable versus 35 for Pro. Fable produced 115k output tokens over 79 steps; Pro used 146 steps.

Failure profiles also differ. Each model regressed the existing test suite in 11% of failures. Fable had the larger big-miss share (18% versus Pro's 10%); Pro's failures were near misses more often (66% versus 57%).

By domain, Fable won 6 of 8 task types, including data modeling and serialization at 88%, and language internals at 78%. Pro won stateful reactivity (66 versus 64) and concurrency and durability (58 versus 45). By language, Fable led Rust (85% versus 65%), Python, Go and JavaScript; Pro led TypeScript (61% versus 57%).

Per-task correlation was 0.39. Both solved 88 tasks; Pro alone solved 12, Fable alone solved 7, and 6 defeated both, covering 107 of 113 tasks. Together AI reports a Pro-first cascade that escalates to Fable on test-suite rejection solves 82.7% of tasks at $8.28 each; the reverse order costs $21.71 for the same accuracy.