DeepSeek's two-tier V4 lineup sharpens the cost-quality tradeoff in late July 2026. V4-Flash ($0.14 input / $0.28 output per 1M tokens) is positioned as the high-volume workhorse, while V4-Pro ($0.435 input / $0.87 output per 1M tokens) targets hardest reasoning and coding. On output tokens, Flash costs roughly 68% less than Pro, or about 12.4× cheaper per unit. The distinction matters because most developers pay for output, not input, when the context is already loaded.
On standard software-engineering benchmarks, V4-Flash achieves ~79% on SWE-bench Verified (resolving real GitHub issues) versus ~80.6% for Pro. The -0731 post-training upgrade specifically tuned Flash for agentic tasks: Terminal-Bench 2.1 at 82.7, Toolathlon Verified at 70.3, and DeepSWE at 54.4 — competitive with much larger closed-source models. DeepSeek noted the upgrade as a post-training refinement, not a new architecture, and kept the model name (deepseek-v4-flash) stable so existing integrations get the improvement automatically.
Compared to Western frontier models, Flash remains 95% cheaper on input and 98% cheaper on output versus Claude Sonnet 4.6, and 85% cheaper than V4-Pro itself on a per-token basis. Cache-hit pricing drops Flash input to $0.003 per 1M, making repeated-prefix workloads nearly free after the first pass. DeepSeek announced a future price increase (effective date and magnitude TBA), but the current rate card stands.
For engineering teams building high-volume agentic systems, long-document RAG, or coding assistants, V4-Flash's low output cost and sufficient reasoning capability flip the equation away from single-model consolidation toward intelligent routing: V4-Flash by default, Flash@max reasoning mode for most work, and Pro escalation only for the hardest minority. That routing strategy typically cuts bills 60–80% while maintaining quality on 95% of tasks.