Does OpenClaw support DeepSeek V4 Pro?

Yes. V4 Pro is in the OpenClaw model picker on managed hosting, and you can BYOK with a DeepSeek key or route via OpenRouter.

When should I pick V4 Pro over V4 Flash?

Pick V4 Pro for sustained multi-step reasoning, agent and terminal work, and multi-file refactors that walk the full 1M context. Since the V4-Pro-0813 GA build shipped on 12 August 2026 it also leads Flash on DeepSeek's published agent benchmarks — 87.9 versus 82.7 on Terminal Bench 2.1 and 62.7 versus 54.4 on DeepSWE. Pick V4 Flash when cost per turn dominates: it is roughly a third of Pro's output price and carries five times the concurrency.

What changed in DeepSeek V4 Pro 0813?

DeepSeek V4 Pro left preview on 12 August 2026 as DeepSeek-V4-Pro-0813. DeepSeek reports Terminal Bench 2.1 rising from 72.1 to 87.9, DeepSWE from 12.8 to 62.7, CyberGym from 52.7 to 83.3 and DSBench-Hard from 31.1 to 67.2. The model ID, the 1M context window and the $0.435 input / $0.87 output per million pricing are unchanged, so existing agents moved to the GA build with no migration.

← Home

Guide

OpenClaw + DeepSeek V4 Pro: Deploy with the Flagship

DeepSeek V4 Pro is DeepSeek's flagship reasoning model — longer chain-of-thought, sustained multi-step analysis, and frontier-class quality at DeepSeek prices. This guide shows three ways to wire it into OpenClaw.

What Is DeepSeek V4 Pro?

DeepSeek released the V4 model family in 2026 with two variants: V4 Flash (low-cost, agent-tuned) and V4 Pro (flagship reasoning). V4 Pro is the strongest reasoning option in the line — longer thinking budget, deep multi-step tool use, and competitive with Claude Sonnet on agent tasks at a fraction of the price.

V4 Pro Left Preview on 12 August 2026

DeepSeek shipped the GA build, DeepSeek-V4-Pro-0813, on 12 August 2026. The model ID is unchanged — you still request deepseek-v4-pro and are served the new build — but the agent scores moved a long way from the preview. These are DeepSeek's own published numbers, not independent evaluations:

Benchmark	V4-Pro-Preview	V4-Pro-0813
Terminal Bench 2.1	72.1	87.9
DeepSWE	12.8	62.7
CyberGym	52.7	83.3
DSBench-Hard	31.1	67.2
Automation Bench	12.8	31.8

That flips the guidance we gave between 31 July and 12 August. The 0731 V4 Flash build briefly led the V4 Pro preview on agent benchmarks (Terminal Bench 82.7, DeepSWE 54.4); the GA Pro build is now ahead of it again on both. Flash is still the value pick for high-volume chat and cheap tool loops — it is about a third of Pro's output cost and carries five times the concurrency — but tool-heavy work is no longer a reason to avoid Pro. V4 Pro remains text-only: send images or video to a vision model instead.

DeepSeek V4 Pro Pricing

Model	Best For	Notes
`deepseek/deepseek-v4-flash`	High-volume chat, cheap tool loops	$0.14 in / $0.28 out per M (0731 build)
`deepseek/deepseek-v4-pro`	Deepest reasoning, agent and terminal work, multi-file refactors	$0.435 in / $0.87 out per M, $0.003625 cache hit (0813 build)

Rates as published on 12 August 2026. DeepSeek has said an overall API price rise is coming, so check DeepSeek pricing before you budget.

For context, Claude Sonnet 4.6 sits around the upper end of frontier-model pricing. V4 Pro delivers comparable reasoning depth at a fraction of that cost — check DeepSeek's pricing page for the exact current ratio.

Option 1: OpenClaw Launch Model Picker (Easiest)

DeepSeek V4 Pro is available out of the box on OpenClaw Launch — no DeepSeek account or API key required.

Go to openclawlaunch.com and click Deploy Now.
Select DeepSeek V4 Pro in the model picker.
Connect Telegram, Discord, WhatsApp, WeChat, or web chat.
Hit Deploy. Your container is live with V4 Pro in ~30 seconds.

Option 2: BYOK with Your DeepSeek API Key

If you have a DeepSeek account, route through your own key:

Get a key from platform.deepseek.com.
Go to dashboard → API Keys on OpenClaw Launch.
Add the DeepSeek key. OpenClaw routes V4 Pro traffic directly to DeepSeek with your key.

Option 3: Self-Host with DeepSeek

# In openclaw.json
{
  "models": {
    "providers": {
      "deepseek": {
        "apiKey": "sk-...",
        "baseUrl": "https://api.deepseek.com"
      }
    }
  },
  "agents": {
    "defaults": {
      "model": {
        "primary": "deepseek/deepseek-v4-pro"
      }
    }
  }
}

Option 4: V4 Pro via OpenRouter

If you want one key across all providers, OpenRouter exposes V4 Pro as deepseek/deepseek-v4-pro:

export OPENROUTER_API_KEY=sk-or-...
# Then in OpenClaw config:
"primary": "openrouter/deepseek/deepseek-v4-pro"

V4 Pro vs V4 vs V4 Flash for Agent Workloads

Workload	Recommended	Why
Customer support chat, FAQ bots	V4 Flash	Fast and cheap; doesn't need deep reasoning
Long-horizon research / multi-file refactors	V4 Pro	Longer chain-of-thought; sustained reasoning depth
Always-on Telegram bot for personal tasks	V4 Flash	Lowest cost for high-volume chat
Agent tool loops, terminal and coding tasks	V4 Pro	0813 build scores 87.9 on Terminal Bench 2.1 against Flash's 82.7
Complex analysis and planning	V4 Pro	Reasoning depth competes with Claude Sonnet

Switching at Runtime

/model deepseek/deepseek-v4-pro      # flagship reasoning
/model deepseek/deepseek-v4-flash    # agent loops, coding, high-volume chat

When V4 Pro Is the Right Pick

Research, planning, or coding tasks where sustained reasoning depth matters
Long-document analysis that walks the full 1M context
You want Claude Sonnet-level quality at DeepSeek prices
Quality per turn matters more than cost per turn — if you are running an always-on bot where token cost compounds, V4 Flash is the cheaper fit.

What's Next?

OpenClaw + DeepSeek V4 — DeepSeek V4 overview
OpenClaw + DeepSeek — provider overview
DeepSeek V4 on Ollama — self-host on local hardware
OpenClaw + OpenRouter — single key for all providers
All supported models