Together AI's benchmark on the DeepSWE coding suite shows DeepSeek-V4 Flash 0731 (released July 31) at $0.10 per task vs. GPT-5.6 Luna at $0.61—a roughly 6x cost difference. While GPT-5.6 Luna leads on single-shot accuracy (67.2% vs. 53.3% pass@1), the economics rewrite at scale: DeepSeek's pass@2 (70.1%) already matches Luna's first-try quality for about one-third the cost. Per-token pricing favors DeepSeek sharply on output ($0.18 vs. $1.20 per million tokens), though Luna's newly discounted input ($0.20 vs. $0.09) has narrowed the input gap.
DeepSeek fails more gracefully: when it breaks, it corrupts existing test suites in only 9% of failures vs. Luna's 15%—the regression signature of OpenAI's lineage models. By domain, Luna dominates reasoning-heavy tasks (program analysis, concurrency, language internals, ~30-point gaps), but DeepSeek holds its own on structured work—SQL builders, config parsing, schema-shaped problems—where it even leads. DeepSeek's JavaScript performance lags sharply (35% vs. 60%), a risk flag for JS-heavy codebases.
A cascade strategy—running DeepSeek first and escalating only on failure—solves 78.9% of DeepSWE tasks at $0.385 each: more accurate than Luna alone, 37% cheaper. For architects scaling coding agents, this benchmark underscores DeepSeek's cost-per-solve advantage and the value of verification-gated multi-model routing. Luna's remaining edge is multimodal input and OpenAI's ecosystem depth; DeepSeek's is open weights (MIT license) and agentic optimization.