Back to blog

Monday, August 10, 2026

DeepSeek V4-Flash-0731 Is Live — and the Old DeepSeek Models Are Gone

Posted by

TH

The Scout

@promptgenius

cover

DeepSeek shipped an update to V4-Flash-0731 and, as flagged in their release notes, the old deepseek-chat and deepseek-reasoner model names are now fully retired. If you're still calling the legacy names, you're not getting what you think.

What Changed

The update. deepseek-v4-flash now routes to DeepSeek-V4-Flash-0731. The calling method is unchanged — just use deepseek-v4-flash and you get the latest build. Flash's reasoning "closely approaches V4-Pro" while staying the fast, cost-effective option, and it performs on par with Pro on simple agent tasks — which is exactly why it's the recommended subagent model in agent setups.

The retirement. deepseek-chat and deepseek-reasoner were scheduled for full retirement after July 24. That's now complete. Legacy names no longer map to the models they once did — DeepSeek's docs state the old names previously routed to V4-Flash thinking/non-thinking, and that window is closed.

What You Should Use Now

TaskModel
Default / non-thinkingdeepseek-v4-flash
Thinking / reasoningdeepseek-v4-flash with thinking enabled, or deepseek-v4-pro
Max-quality agent workdeepseek-v4-pro
Long context (1M)deepseek-v4-pro[1m]
Cheap subagentsdeepseek-v4-flash

Both V4 models support 1M context and dual thinking/non-thinking modes via the thinking parameter. Agent integrations (Claude Code, OpenCode, Copilot) use the Anthropic-compatible endpoint — see our DeepSeek V4 + Claude Code setup for the exact env vars.

Why It Matters for Agent Users

The retirement is the kind of silent breaking change that bites production setups. If your CI, scripts, or agent configs still reference deepseek-chat or deepseek-reasoner:

  1. Audit your configs — grep for deepseek-chat and deepseek-reasoner in env files, agent configs, and CI.
  2. Replace with the V4 names — deepseek-v4-flash / deepseek-v4-pro.
  3. Verify the thinking mode — if you relied on deepseek-reasoner for reasoning, you now enable thinking explicitly ("thinking": {"type": "enabled"}) rather than via model name.

The good news: the price-to-performance equation keeps getting better. Flash at its new build is the cost play for bulk agent work; Pro is the reasoning tier. For cost-optimization patterns across the lineup, see DeepSeek V4 Flash vs Pro and Cost Optimization Patterns.

Sources: DeepSeek API docs — V4 Preview release notes and model listing. Related: DeepSeek V4 with Claude Code.