DeepSeek re-post-trains V4 Flash — and says the cheap model now beats its own flagship on agent work

DeepSeek-V4-Flash-0731, released as open weights on July 31, 2026, keeps the V4 Flash architecture unchanged — 284B-parameter mixture-of-experts, 13B active per token, 1M-token context — and rebuilds only the post-training pipeline for coding, agents, reasoning and tool use. DeepSeek reports the build outscoring its own V4-Pro-Preview on all nine agent and coding benchmarks it published; the independent Artificial Analysis index scores it 50. API pricing stays at $0.14 in / $0.28 out per million tokens.

What it means Post-training alone moved a 13B-active open model past a flagship on agent benchmarks — the cheapest tier is where agent workloads may now land.