DeepSeek-V4 Flash 0731 vs GPT-5.6 Luna on DeepSWE: Cost and Coding

DeepSeek-V4 Flash 0731 vs GPT-5.6 Luna on DeepSWE: Cost and Coding Webflow Analyze/Optimize tracking bridge -->

💰 Announcing our Series C. Intelligence should be abundant, not expensive →

🤝 Together AI & Y Combinator announce partnership to deliver the first dedicated YC GPU cluster →

⚡ On-demand B200s now available on Together GPU Clusters →

🚀 Now serving MiniMax-M3 for efficient inference →

All blog posts

Model Library

Published 8/6/2026

DeepSeek-V4 Flash 0731 vs GPT-5.6 Luna on DeepSWE: Cost and Coding

The open-weight model setting the new price-intelligence bar

Authors

Zain Hasan, Shobhit Dixit

Table of contents

40+ Models Chosen for Production...40+ Models Chosen for Production...40+ Models Chosen for Production...

Key Takeaways

While GPT-5.6 Luna is the stronger engineer on every quality measure, DeepSeek-V4 Flash 0731 is cheap enough that a DeepSeek-first cascade beats Luna alone on both accuracy and cost. GPT-5.6 Luna leads DeepSWE pass@1 decisively at 67.2% vs 53.3%, a 14 point gap, and holds the lead at every equal attempt count. DeepSeek-V4 Flash is the cheapest model on the DeepSWE board: \$0.10 per rollout vs \$0.61, delivering 532 solves per \$100 against Luna's 110. DeepSeek-V4 Flash fails more cleanly, breaking the repo's existing test suite in 9% of failures vs Luna's 15%. Running DeepSeek-V4 Flash first and escalating to Luna only on failure solves 78.9% of tasks at \$0.385 each: more accurate than Luna alone and 37% cheaper.

Available now · Open weights

Run DeepSeek-V4 Flash on Together AI

The cheapest model on the DeepSWE board, about ten cents a task. Serve the cheap first stage at production scale without frontier token prices.

Open the playground

‍ DeepSeek-V4 Flash 0731 is the cheapest model on the entire DeepSWE board: about ten cents a task! GPT-5.6 Luna, a solid upper-tier flagship, runs \$0.61 a task, roughly six times DeepSeek’s price. So the question is not which one wins the leaderboard (Luna, comfortably) but what six-times-cheaper buys you, what it costs you, and whether the two together beat either one alone. We ran DeepSeek-V4 Flash 0731 (max) against GPT-5.6 Luna (max) on all 113 DeepSWE tasks: real, long-horizon feature requests from live open-source repos, four trials each, graded pass/fail by a hidden test suite. That is 900 rollouts in total from the published per-trial records (452 on DeepSeek's side, 448 on Luna's). Every figure below comes from this run, so it can differ from other public scorecards. At a glance

DeepSWE · Head to Head

DeepSeek-V4 Flash vs GPT-5.6 Luna at a glance

Model Pass@1 Avg cost Out tok Steps

gpt-5.6-luna [max] 67% ± 4% $0.61 73k 102

deepseek-v4-flash [max] 53% ± 4% $0.10 108k 153

Luna is ~6x the cost for ~14 points more accuracy; DeepSeek uses more output tokens and steps to get less far, but at \$0.10 it is the cheapest run in the set by a wide margin.

The DeepSWE scoreboard: pass@1 and pass@k Luna wins single shot clearly: 67.2% pass@1 to DeepSeek's 53.3% under DeepSWE's official scoring, a 14 point lead, or about 26% more accurate in relative terms. At equal attempt counts Luna stays ahead at every k (81.6 vs 70.1 at two attempts, 90.3 vs 80.5 at four). On raw solving ability, this is not a close fight. But DeepSWE is exactly the kind of workload where you can fan out several attempts in parallel, and there the economics rewrite the picture. DeepSeek's pass@2 (70.1%) already edges Luna's single shot (67.2%), and two DeepSeek attempts cost about \$0.20 to Luna's \$0.61. If a verifier can pick the winning run, the cheap model matches the flagship's first-try quality for a third of the price, before any of the routing tricks below even come into play.

Cost comparison: DeepSeek-V4 Flash vs GPT-5.6 Luna pricing At \$0.61 a task, Luna returns 110 solves per \$100 to DeepSeek's 532: a 4.8x value edge for the cheap model. The cost per rollout gap is about 6x. The one thing the cheap model does not buy you is speed, and this is the twist versus the usual "flash" model story: DeepSeek is the slower of the two here, taking a median 23 minutes and 148 steps to Luna's 16 minutes and 92. It grinds (and emits more output, 104k median tokens to Luna's 70k). DeepSeek's advantage is money, not time. That wall-clock gap is partly an artifact of how new the model is; expect it to shrink as inference engines get tuned for it.

Failure modes: DeepSeek fails more gracefully than the flagship The other thing the cheap model quietly wins is discipline. When DeepSeek fails, it breaks the repository's existing test suite in only 9% of failures. Luna does so in 15%: the GPT-family regression signature, the same 15 to 20% we see across Sol and the other OpenAI-lineage models. Both fail mostly by near miss (DeepSeek 69%, Luna 66%), but the more expensive model is the one more likely to disturb code that already worked. If you deploy Luna, gate it behind a full regression run; DeepSeek needs that guardrail less.

DeepSeek-V4 Flash vs GPT-5.6 Luna by task domain Classify the 113 tasks by what the code actually is, and Luna wins 7 of 8 domains. Its biggest edges are exactly the reasoning-heavy work: program analysis (69 vs 33), concurrency and durability (70 vs 38), language and runtime internals (86 vs 59), roughly a 30 point gap in each. This is where model capability actually shows up, and DeepSeek falls off hard. DeepSeek holds exactly one domain, and it is a telling one: query and config languages, 78 vs 70. The SQL builders, window functions, keyset pagination, config parsers. Structured, schema-shaped, convention-following work is where the cheap model is genuinely competitive, even ahead. Everywhere the task demands holding a hard invariant across a whole system, Luna's capability separates.

DeepSeek-V4 Flash vs GPT-5.6 Luna by programming language Luna wins all five, but the margins tell you where to be careful with DeepSeek. It is respectable on Rust (55 vs 60) and Go (62 vs 79), but its JavaScript is a collapse: 35 vs Luna's 60, the weakest single cell in the entire matchup, and a 25 point hole. DeepSeek's Python is also soft (49 vs 65). If your stack is JS-heavy, the cheap model is a false economy; if it is config, query, or Rust, DeepSeek closes most of the gap.

How similar are DeepSeek-V4 Flash and GPT-5.6 Luna? Less than the cheap-tier pairs. Per-task correlation is 0.50, and the diversity is...