DeepSeek V4.1 Flash vs Opus 5: 77% vs 74% on DeepSWE at ~1/60 the cost per task
The author cites a benchmark chart showing V4.1-Flash-1017 at 77% on DeepSWE v1.1 for about $0.20 per task versus Opus 5 at 74% and about $12, arguing cheap runs make parallel agent experimentation affordable.
60x cheaper than @claudeai's Opus 5, while scoring HIGHER on DeepSWE. this is the kind of progress that can genuinely change how we build with AI. @deepseek_ai's DeepSeek V4.1-Flash-1017 scores 77% on DeepSWE v1.1 at roughly $0.20 per task in the benchmark chart, compared to Opus 5 at around 74% and $12 per task. obviously, one benchmark does not tell you which model is better at every coding task. Opus still leads on several other evaluations in the comparison. but think about what this cost difference means in practice. when each agent run is expensive, you think twice before asking the model to explore multiple approaches, investigate an obscure bug, or attempt a task that might fail. when the cost drops this much, you can afford to experiment more, run more tasks in parallel, and let agents handle smaller jobs that would not have justified the expense before. and that is why i care so much about the cost per successfully completed task, not just benchmark scores or token prices. "Cheap intelligence changes how much work you can afford to delegate." we are getting closer to a point where running an agent on every small engineering problem becomes the default, rather than something you reserve for difficult tasks.
This post quotes another post: x.com/deepseek_ai/status/2111375640717033777