DeepClaude: Claude Code Agent Loop Powered by DeepSeek V4 Pro
Summary
- • Open-source tool swaps Claude Code's model backend for DeepSeek V4 Pro
- • Cost drops from $200/month to $20-80/month depending on usage intensity
- • DeepSeek V4 Pro scores 96.4% on LiveCodeBench at $0.87/M output tokens
- • All Claude Code capabilities preserved except image input and parallel tool use
Details
DeepClaude routes Claude Code's agent loop through DeepSeek V4 Pro via environment variable injection
The tool sets ANTHROPIC_BASE_URL and related env vars per session, transparently redirecting Claude Code's API calls without modifying the CLI itself. On exit, original settings are restored. This approach requires no fork or patching of Claude Code.
Output token cost drops 17x: from $15/M (Anthropic) to $0.87/M (DeepSeek V4 Pro)
Monthly spend for light users falls from $200 (capped subscription) to roughly $20, and heavy users pay around $50 versus the same $200 cap. DeepSeek's context caching drops cached-turn costs further to $0.004/M — making repetitive agent loops dramatically cheaper than even the base rate implies.
DeepSeek V4 Pro scores 96.4% on LiveCodeBench, justifying the quality-for-cost substitution
The LiveCodeBench score is cited to justify that coding task quality is comparable to Claude Opus despite the significant price difference. DeepSeek is the largest open-weight model available at time of release.
Four backends supported: DeepSeek (China), OpenRouter (US), Fireworks AI (US), and native Anthropic
DeepSeek's own servers offer the lowest price and automatic context caching but are hosted in China. OpenRouter provides US/EU-hosted routing at the same price with lower latency. Fireworks AI is the fastest but costs roughly 4x DeepSeek. Anthropic remains available as a fallback flag.
Context caching on DeepSeek reduces repeat-turn agent costs by up to 120x per source
After the first request in a session, the system prompt and file context are cached at $0.004/M tokens versus $0.44/M uncached. For multi-step autonomous coding loops where the same large context is re-sent repeatedly, this makes DeepClaude substantially cheaper than the headline per-token rate.
Limitations: no image/vision input; parallel tool use unsupported in practice
DeepSeek's Anthropic-compatible endpoint does not support vision input, ruling out use cases requiring screenshot or image analysis. Parallel tool use is technically supported by DeepSeek but Claude Code does not invoke it.
Product Launch = tool release, Financials = cost/pricing data, Stat = benchmark or quantitative claim, Tech Info = backend/configuration details, New Tech = novel capability, Insight = tradeoff or limitation analysis
What This Means
For AI practitioners who rely on Claude Code for autonomous coding tasks, DeepClaude offers a practical way to bypass both the $200/month subscription cost and its usage caps by substituting a competitive open-weight model at a fraction of the price. The approach is architecturally clean — no forking Claude Code, just environment variable redirection — which means it stays compatible with future Claude Code updates. The main tradeoff is model hosting location (DeepSeek's cheapest tier runs on Chinese servers) and the loss of vision input, which matters for teams working with UI screenshots or diagrams. This release reflects a broader pattern of the open-source community treating frontier proprietary tool UX as separable from the underlying model, accelerating cost arbitrage across the AI tooling stack.