A head-to-head benchmark of DeepSeek-V4 Flash 0731 and GPT-5.6 Luna on the DeepSWE coding benchmark across 900 rollouts on 113 real-world software engineering tasks. GPT-5.6 Luna leads pass@1 by 14 points (67.2% vs 53.3%), wins 7 of 8 task domains, and is faster. DeepSeek-V4 Flash costs ~6x less ($0.10 vs $0.61 per rollout), delivering 532 solves per $100 versus Luna's 110, and fails more cleanly (9% vs 15% regression rate). The key finding: running DeepSeek first and escalating to Luna only on failure achieves 78.9% task solve rate at $0.385 each — more accurate than Luna alone and 37% cheaper. DeepSeek leads only in query/config languages and is notably weak on JavaScript (35% vs 60%)
Table of contents
At a glanceThe DeepSWE scoreboard: pass@1 and pass@kCost comparison: DeepSeek-V4 Flash vs GPT-5.6 Luna pricingFailure modes: DeepSeek fails more gracefully than the flagshipDeepSeek-V4 Flash vs GPT-5.6 Luna by task domainDeepSeek-V4 Flash vs GPT-5.6 Luna by programming languageHow similar are DeepSeek-V4 Flash and GPT-5.6 Luna?The portfolio play: why the cascade works anywayWhat it meansAppendix: Data TableFAQsQuestions this post answers
What is the cost difference between DeepSeek-V4 Flash 0731 and GPT-5.6 Luna per coding task?
DeepSeek-V4 Flash 0731 costs about $0.10 per rollout versus $0.61 for GPT-5.6 Luna — roughly a 6x price difference. Measured by output, DeepSeek returns 532 solves per $100 against Luna's 110, a 4.8x value advantage. The trade-off is quality: Luna leads pass@1 by 14 points (67.2% vs 53.3%) and is faster at a median 16 minutes versus DeepSeek's 23. Engineers choosing between AI coding models for budget-constrained pipelines track these benchmarks on daily.dev.
Does running DeepSeek first and escalating to a stronger model improve accuracy and reduce cost compared to using the stronger model alone?
Yes. Running DeepSeek-V4 Flash first and escalating to GPT-5.6 Luna only when the test suite rejects the answer solves 78.9% of tasks at $0.385 each — 11.7 points more accurate than Luna alone (67.2%) and 37% cheaper than Luna alone ($0.61). DeepSeek clears ~53% of tasks for $0.10 each, so Luna's cost only applies to the harder remainder. Teams designing AI agent pipelines for software engineering tasks find cascade routing strategies like this on daily.dev.
How does DeepSeek-V4 Flash compare to GPT-5.6 Luna on JavaScript coding tasks?
DeepSeek-V4 Flash scores only 35% on JavaScript tasks versus GPT-5.6 Luna's 60% — a 25-point gap and the weakest single result in the entire matchup. DeepSeek is more competitive on Rust (55% vs 60%) and actually leads on query and config language tasks (78% vs 70%). JS-heavy stacks should not rely on DeepSeek-V4 Flash as a cost-saving substitute. JavaScript developers evaluating AI coding tools for their stack can follow model comparisons like this on daily.dev.