---
title: "DeepSeek-V4 Flash 0731 vs GPT-5.6 Luna on DeepSWE: Cost and Coding"
description: "On the DeepSWE coding benchmark, GPT-5.6 Luna achieves 67.2% pass@1 versus DeepSeek-V4 Flash’s 53.3%, a 14-point gap. DeepSeek costs $0.10 per task, roughly six times less than Luna’s $0.61, delivering 532 solves per $100 versus Luna’s 110. A cascade that runs DeepSeek first and escalates only on failure solves 78.9% of tasks at $0.385 each, beating Luna’s solo accuracy and cutting cost by 37%. DeepSeek fails more gracefully (9% broken tests vs 15%) but is slower and dramatically weaker on JavaScript (35% vs 60%). Luna dominates seven of eight task domains, while DeepSeek leads only on query and config languages, illustrating how inexpensive models can combine with flagships for superior price-performance."
---

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

[Together AI](https://yomu.fyi/company/together-ai) · Aug 6, 2026

**Type:** Benchmark

## Summary

On the DeepSWE coding benchmark, GPT-5.6 Luna achieves 67.2% pass@1 versus DeepSeek-V4 Flash’s 53.3%, a 14-point gap. DeepSeek costs $0.10 per task, roughly six times less than Luna’s $0.61, delivering 532 solves per $100 versus Luna’s 110. A cascade that runs DeepSeek first and escalates only on failure solves 78.9% of tasks at $0.385 each, beating Luna’s solo accuracy and cutting cost by 37%. DeepSeek fails more gracefully (9% broken tests vs 15%) but is slower and dramatically weaker on JavaScript (35% vs 60%). Luna dominates seven of eight task domains, while DeepSeek leads only on query and config languages, illustrating how inexpensive models can combine with flagships for superior price-performance.

## Approach / What changed

Cascade strategy: run DeepSeek-V4 Flash first; if the test suite rejects the answer, escalate to GPT-5.6 Luna. This uses the cheap model for easy tasks and the expensive model only for the hard remainder.

## Takeaways

- GPT-5.6 Luna leads DeepSWE pass@1 at 67.2% vs 53.3% for DeepSeek-V4 Flash, a 14-point gap sustained at every attempt count.
- DeepSeek-V4 Flash costs $0.10 per task versus Luna’s $0.61, delivering 532 solves per $100 against Luna’s 110.
- A cascade that runs DeepSeek first and escalates to Luna on failure solves 78.9% of tasks at $0.385 each, beating Luna alone in accuracy and cost.

## FAQ

### How much cheaper is DeepSeek-V4 Flash per task than GPT-5.6 Luna?

DeepSeek-V4 Flash costs about $0.10 per task, while GPT-5.6 Luna costs $0.61, roughly six times more. This gives DeepSeek 4.8x more solves per dollar.

### Which model is faster on wall-clock time?

GPT-5.6 Luna is faster, taking a median 16 minutes and 92 steps, compared to DeepSeek's 23 minutes and 148 steps. The gap may shrink with inference engine tuning.

### What failure mode does DeepSeek have over Luna?

DeepSeek fails more gracefully, breaking the repo's existing test suite in only 9% of failures versus Luna's 15%, making it less disruptive to working code.

**Tags:** [AI](https://yomu.fyi/topic/ai), [Cost Optimization](https://yomu.fyi/topic/cost-optimization), [DeepSWE](https://yomu.fyi/topic/deepswe), [Inference](https://yomu.fyi/topic/inference), [LLMs](https://yomu.fyi/topic/llm)

- Source: [Together AI](https://www.together.ai/blog/deepseek-v4-flash-0731-vs-gpt-5-6-luna-on-deepswe-cost-and-coding)
- Source URL: https://www.together.ai/blog/deepseek-v4-flash-0731-vs-gpt-5-6-luna-on-deepswe-cost-and-coding
- Ingested by Yomu: 2026-09-19T10:13:45.141Z

[Read original post](https://www.together.ai/blog/deepseek-v4-flash-0731-vs-gpt-5-6-luna-on-deepswe-cost-and-coding)