Use DeepSeek with Claude Code

Claude Code doesn’t require a Claude model: it works with any backend that speaks the Anthropic Messages API. CheapestInference exposes DeepSeek V4 Flash through an Anthropic-compatible endpoint, so you can run Claude Code on DeepSeek with unlimited usage during your reserved time blocks — a flat monthly fee instead of per-token billing.

Setup (2 minutes)

1. Subscribe to the Core Pool and copy your subscriber key (sk-...) from the dashboard.
2. Export these variables in your shell (or put them in your shell profile):

export ANTHROPIC_BASE_URL="https://api.cheapestinference.com/anthropic"
export ANTHROPIC_AUTH_TOKEN="sk-..."             # your subscriber key
export ANTHROPIC_MODEL="deepseek-v4-flash"
export ANTHROPIC_SMALL_FAST_MODEL="deepseek-v4-flash"
export CLAUDE_CODE_AUTO_COMPACT_WINDOW=1000000   # DeepSeek V4 Flash's 1M window

3. Start claude as usual. Every request now runs on DeepSeek.

Why DeepSeek V4 Flash for Claude Code

• 1M-token context: the whole repo fits. Setting CLAUDE_CODE_AUTO_COMPACT_WINDOW to the full window means long sessions compact far less often than on a 200K model.
• Fast: V4 Flash is DeepSeek’s speed-optimized line — agentic loops with many tool calls stay snappy.
• Flat price: coding agents are token-hungry; a long Claude Code session can burn millions of tokens. During your time blocks there is no per-token meter — see Unlimited Subscriptions.
• Thinking blocks: DeepSeek’s reasoning surfaces through standard Anthropic thinking content blocks, so Claude Code renders it natively. See Reasoning for toggling it.

Per-project configuration

Pin one project to DeepSeek without touching your global shell profile with .claude/settings.json in the project:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.cheapestinference.com/anthropic",
    "ANTHROPIC_AUTH_TOKEN": "sk-...",
    "ANTHROPIC_MODEL": "deepseek-v4-flash",
    "ANTHROPIC_SMALL_FAST_MODEL": "deepseek-v4-flash",
    "CLAUDE_CODE_AUTO_COMPACT_WINDOW": "1000000"
  }
}

FAQ

Can Claude Code use DeepSeek?

Yes. Claude Code talks to any endpoint that implements the Anthropic Messages API. Point ANTHROPIC_BASE_URL at https://api.cheapestinference.com/anthropic, set ANTHROPIC_MODEL to deepseek-v4-flash, and Claude Code runs on DeepSeek — tool calls, streaming and thinking blocks included. No plugin or proxy on your machine.

Is DeepSeek good enough for Claude Code?

For most day-to-day coding, yes: V4 Flash handles multi-file edits and tool loops well, and its 1M window holds far more of the repo than a 200K model. For the hardest refactors, flagship-class models like Kimi K3 leave a bigger quality margin — same setup, different model id.

What does it cost to run Claude Code on DeepSeek?

A flat monthly subscription to the Core Pool — see current prices. Usage during your reserved time blocks is unlimited, so a heavy Claude Code week costs the same as a light one. On per-token APIs the same agentic session bills every one of its millions of tokens.

Does /compact and auto-compaction work?

Yes. Set CLAUDE_CODE_AUTO_COMPACT_WINDOW=1000000 so Claude Code compacts against DeepSeek’s real 1M window instead of assuming a Claude-sized one.

Notes

• Your Core Pool subscription also covers MiMo v2.5 — switch models by changing ANTHROPIC_MODEL, no new key needed.
• Want a stronger model in the same setup? The same two variables work with GLM 5.2 (Frontier Pool) or Kimi K3 (Flagship Pool).
• General setup, model table and troubleshooting: the main Claude Code guide.