DeepSeek V4 Flash 0731 Goes GA: Beats Its Own Flagship on Agent Benchmarks at a Third of the Price

DeepSeek retrained its flash model with a stronger agentic post-training pipeline, and V4-Flash-0731 now outscores V4-Pro on all 9 agent and coding benchmarks — at one-third the cost and zero migration effort.

The 0731 build went GA on July 31. For anyone already calling deepseek-v4-flash, the upgrade is automatic — no endpoint change, no model rename. The model name deepseek-v4-flash now serves V4-Flash-0731.

What changed: post-training only (same 284B/13B-activated architecture), with a much stronger pipeline focused on agentic workflows and reasoning. The result: V4-Flash-0731 beats DeepSeek V4-Pro-Preview on every benchmark in DeepSeek's published suite:

Benchmark	V4-Flash-0731	V4-Pro-Preview
Terminal Bench 2.1	82.7	72.1
NL2Repo	54.2	38.5
DeepSWE	54.4	12.8
Cybergym	76.7	52.7

Note: vendor-reported scores from an unreleased harness; independent verification is advised.

Also new in 0731: native Responses API support, so deepseek-v4-flash can be added as a direct Codex provider without a compatibility shim.

Why it matters: At $0.14/M input tokens, V4-Flash-0731 is the obvious routing destination for extraction, classification, and structured-parsing steps in agent pipelines — the tasks that don't need frontier reasoning but dominate your token bill. The upgrade to this performance level with zero integration work is rare.

Sources: DeepSeek V4-Flash-0731 Major Agentic Gains — MarkTechPost · V4-Flash-0731 Beats Flagship on 9 Benchmarks — TechTimes · DeepSeek API Changelog