DeepClaude — Run Claude Code with DeepSeek V4 Pro for 17x Less

Claude Code is powerful but expensive — Anthropic’s Max 20x plan costs $200/month, and heavy sessions can burn through tokens fast. DeepClaude is an open-source script that redirects Claude Code’s API calls to DeepSeek V4 Pro (or any Anthropic-compatible backend), preserving the full agent experience at 17x lower cost. The tradeoff: no vision, no parallel tool use, and benchmark performance within 0.2 points of Opus rather than matching it.

How It Works

DeepClaude is a bash/PowerShell script — no fork, no rewrite. It temporarily overrides two environment variables:

ANTHROPIC_BASE_URL=<deepseek-or-openrouter-endpoint>
ANTHROPIC_AUTH_TOKEN=<your-api-key>

Claude Code’s entire agent loop, file I/O, subagent orchestration, and git integration continue unchanged — only the model behind the API changes.

┌─────────────────────┐
│   Claude Code CLI    │
│  (unchanged UX)      │
└──────────┬──────────┘
           │ API calls
           ▼
┌─────────────────────┐
│    DeepClaude        │
│  (env var override)  │
└──────────┬──────────┘
           │ redirected to
           ▼
┌─────────────────────────────────┐
│  DeepSeek V4 Pro / OpenRouter   │
│  / Fireworks AI / any backend   │
└─────────────────────────────────┘

Cost Comparison

DeepSeek’s automatic context caching is the key: after the first request, the system prompt and file context are cached at fractions of a cent per million tokens — making agent loops (where context is mostly stable) extremely cheap.

Note: DeepSeek V4 Pro’s 75% promotional discount expires May 31, 2026, after which prices quadruple to $1.74/$3.48 per M tokens. Even post-promotion, it remains ~7x cheaper than Anthropic on input and ~7x cheaper on output.

Performance: How Close Is DeepSeek V4 Pro?

Within measurement noise on SWE-bench — but SWE-bench is one benchmark. Real-world differences may be larger on tasks requiring deep reasoning, multi-step planning, or edge cases. Note that Opus 4.7 is now the current Anthropic flagship, so the gap against the latest model may be wider.

What You Lose

Setup (5 Minutes)

# Clone
git clone https://github.com/aattaran/deepclaude.git
cd deepclaude

# Make executable
chmod +x deepclaude.sh

# Set your API key
export DEEPSEEK_API_KEY=your-key-here
# Or: export OPENROUTER_API_KEY=your-key
# Or: export FIREWORKS_API_KEY=your-key

# Run Claude Code through DeepClaude
./deepclaude.sh

The tool supports switching backends during a session via custom Claude Code commands or the --switch flag.

Caveat: DeepClaude is a young third-party wrapper that depends on Claude Code’s internal API format. It may break if Anthropic changes request formats, headers, or tool schemas. It is not an officially supported integration.

Alternative Backends

DeepClaude supports any Anthropic-compatible API:

When to Use DeepClaude vs Anthropic

Real-World Use Cases

• Budget-conscious developers — Run Claude Code’s full agent loop for ~$3/session instead of $20+.
• Open-source contributors — Use Claude Code on personal projects without a $200/month subscription.
• Team exploration — Let team members try Claude Code workflows before committing to Anthropic pricing.
• Batch processing — Run long automated sessions (linting, refactoring, docs) where per-token cost dominates.

How LearnAI Team Could Use This

• Cost modeling exercise — Students calculate real costs of AI-assisted development across different backends, learning that model choice is an engineering decision, not just a quality decision.
• API compatibility lab — Teach how API-compatible backends enable model swapping — a pattern used throughout production AI systems.
• Benchmark vs. real-world discussion — Use the 0.2-point SWE-bench gap as a starting point: when does a small benchmark difference matter in practice?

Links

• GitHub: aattaran/deepclaude
• Setup Guide: MindStudio
• Course: FindSkill.ai