← Home

Guide

OpenClaw + DeepSeek V4: Use DeepSeek V4 Pro and V4 Flash with OpenClaw

DeepSeek's fourth-generation models are live on OpenClaw. 1M context window, strong reasoning, and ultra-low pricing — pick Flash for agent and coding work, or Pro for the deepest sustained reasoning.

What Is DeepSeek V4?

DeepSeek is a Chinese AI lab known for open, cost-efficient frontier models. DeepSeek V4 is the fourth-generation family and ships in two flavors: V4 Pro for the deepest reasoning and V4 Flash for the lowest cost per token. Both support a 1M token context window and a 384K maximum output length.

Compared to the prior V3.2 release, V4 trades a larger context window (128K → 1M), better coding and agentic tool use, and a cleaner Pro/Flash split so you can match the model to the workload instead of paying Pro prices for Flash-sized work.

Updated 12 August 2026: V4 Pro left preview as the V4-Pro-0813 build. DeepSeek reports Terminal Bench 2.1 climbing from 72.1 to 87.9 and DeepSWE from 12.8 to 62.7, which puts Pro back ahead of the V4-Flash-0731 build (82.7 and 54.4) that had briefly led it. Neither model ID changed and pricing is unchanged, so existing agents are already on the new builds. See the V4 Pro guide and the V4 Flash guide for the full benchmark tables and setup.

DeepSeek V4 Pro vs V4 Flash

ModelContextInput (per 1M)Output (per 1M)Best for
DeepSeek V4 Pro1M tokens$0.435$0.87Hardest long-horizon reasoning, deep multi-file refactors
DeepSeek V4 Flash1M tokens$0.14$0.28Agent tool loops, terminal and coding work, high-volume chat

Direct DeepSeek API cache-miss rates. Cache hits are far cheaper ($0.0028 Flash / $0.003625 Pro), and routing through OpenRouter prices Flash at roughly $0.09 / $0.18.

Flash is the cheap one, but since the July 2026 refresh it is no longer the weak one. Both models are dramatically cheaper than Claude Sonnet or GPT-5.4, so the real question is not cost — it is whether your workload is agent-shaped (Flash) or reasoning-shaped (Pro).

How to Deploy DeepSeek V4 with OpenClaw Launch

The fastest path is through OpenClaw Launch — no API key needed, everything routes through the included AI credits via OpenRouter.

  1. Go to openclawlaunch.com and open the configurator.
  2. In the model dropdown, pick DeepSeek V4 Flash for agent loops, coding, and everyday use, or DeepSeek V4 Pro for the deepest sustained reasoning.
  3. Pick your chat platform (Telegram, Discord, WhatsApp, WeChat, or browser gateway).
  4. Click Deploy. Your DeepSeek V4 agent is live in about 30 seconds.
Tip: You can switch between Flash and Pro at any time from the dashboard or with the /model chat command — no redeploy needed. A common pattern is Flash as the everyday default, with Pro held back for the reasoning-heaviest sessions.

Self-Hosted Configuration

If you run OpenClaw on your own server, configure OpenRouter as a provider in your openclaw.json and point the default agent at either V4 Pro or V4 Flash:

{
  "models": {
    "providers": {
      "openrouter": {
        "apiKey": "sk-or-..."
      }
    }
  },
  "agents": {
    "defaults": {
      "model": {
        "primary": "openrouter/deepseek/deepseek-v4-flash"
      }
    }
  }
}

Swap deepseek-v4-flash for deepseek-v4-pro to route to the Pro variant instead. Grab an OpenRouter key at openrouter.ai/keys.

Which DeepSeek V4 Model Should You Pick?

A simple heuristic:

  • Start with V4 Flash. High-volume chat, agent tool loops, terminal work, and general coding all sit in its wheelhouse at about a third of Pro's output-token cost. The 0813 Pro build took back the lead on DeepSeek's published agent benchmarks, so Flash is now the value default rather than the benchmark winner — still the one to beat on price, and its 2,500 concurrency limit (against Pro's 500) suits busy Telegram and Discord bots.
  • Reach for V4 Pro when the task is reasoning-shaped rather than agent-shaped — deep multi-file refactors, architecture discussions, and long-document analysis that genuinely walks the full 1M context.
  • Run both if you want — one instance per role, or swap at runtime with /model. Since switching is free, benchmark both on your own workload rather than trusting anyone's table.

DeepSeek V4 vs Other Frontier Models

ModelInput (per 1M)Output (per 1M)ContextNotes
DeepSeek V4 Flash$0.14$0.281MCheapest frontier-adjacent model in the lineup
DeepSeek V4 Pro$0.435$0.871MFrontier reasoning, strong coding
Claude Sonnet 4.6$3.00$15.00200KBest nuanced reasoning and creative writing
GPT-5.4$2.50$15.001M+Strong general-purpose model with vision
Gemini 3.1 Pro$2.00$12.001MMultimodal with video input

DeepSeek V4 Flash is a standout on pure cost-efficiency, and V4 Pro closes much of the gap to Claude Sonnet / GPT-5.4 at a fraction of their per-token cost. On the hardest long-horizon coding and agentic work, the frontier models still lead.

BYOK with DeepSeek V4

If you already have an OpenRouter account or a direct DeepSeek API key, you can bring it to OpenClaw Launch via BYOK. In the configurator, select BYOK and paste your OpenRouter key — all requests route through your key and your own billing.

What's Next?

Deploy with DeepSeek V4

Get an AI agent powered by DeepSeek V4 Pro or V4 Flash running in 30 seconds.

Deploy Now