- DeepClaude is an open-source tool surfaced May 4, 2026 that lets Claude Code run on DeepSeek V4 Pro, OpenRouter, or Fireworks AI instead of Anthropic‘s API.
- DeepSeek V4 Pro is reported at $0.87/M output tokens versus Anthropic Opus at $15/M — roughly 17x cheaper at headline rates.
- The tool runs a localhost:3200 proxy that intercepts API calls and routes to the chosen backend; Claude Code’s tool loop, file editing, bash execution, subagent spawning, and git operations all still work.
- Limitations: image input, parallel tool use, MCP server tools, and Anthropic’s prompt caching are unsupported through the compatibility layer.
What Happened
An open-source tool called DeepClaude surfaced on Hacker News on May 4, 2026. The tool routes Claude Code’s API calls to DeepSeek V4 Pro by default, with optional backends for OpenRouter and Fireworks AI, while preserving Claude Code’s full agent loop. The author cites DeepSeek V4 Pro’s reported 96.4% on LiveCodeBench and pricing at $0.87/M output tokens versus Anthropic’s $15/M for Opus, framing the project as “swap the brain while keeping the body.”
Why It Matters
Claude Code costs $200/month with usage caps in Anthropic’s plans, and Uber‘s recently disclosed AI budget burn (the company exhausted its 2026 AI allocation in four months on Claude Code and Cursor, per CTO comments covered earlier this week) shows how quickly per-engineer costs scale at large organizations. DeepClaude tests a specific commercial proposition: that 80% of routine Claude Code workflows can run on a Chinese open-weight model 17x cheaper, with an option to switch back to Anthropic for the 20% of complex reasoning where the gap matters. If the trade-off works, the per-seat economics of agentic coding change materially.
Technical Details
DeepClaude works by setting environment variables Claude Code already reads — ANTHROPIC_BASE_URL, ANTHROPIC_AUTH_TOKEN, ANTHROPIC_DEFAULT_OPUS_MODEL, ANTHROPIC_DEFAULT_SONNET_MODEL, ANTHROPIC_DEFAULT_HAIKU_MODEL, and CLAUDE_CODE_SUBAGENT_MODEL — to point at the chosen backend. A proxy runs on localhost:3200 and intercepts API calls; a control endpoint at /_proxy/mode lets users switch the active backend mid-session without restarting.
Backend pricing as published by the project: DeepSeek (default) at $0.44/M input and $0.87/M output, served from China with automatic context caching that drops repeat-turn costs to $0.004/M; OpenRouter at the same DeepSeek rate but US servers; Fireworks AI at $1.74/M input and $3.48/M output (US, fastest inference); Anthropic at $3.00/M input and $15.00/M output (the original Claude Opus path).
What works: file reading/writing/editing via Claude Code’s Read/Write/Edit tools, bash and PowerShell execution, glob and grep search, multi-step autonomous tool loops, subagent spawning, git operations, project initialization (/init), and thinking mode. What does not work through the compatibility layer: image and vision input (DeepSeek’s Anthropic-compatible endpoint does not support images), parallel tool use (forced to sequential), MCP server tools, and Anthropic’s cache_control prompt caching (DeepSeek has its own automatic caching instead).
Cost projections published by the author: light usage (10 days/month) at roughly $20/month vs Anthropic’s $200 cap (90% savings); heavy usage at roughly $50/month (75% savings); auto-loop heavy usage at roughly $80/month (60% savings).
Who’s Affected
Individual developers running Claude Code at home or on side projects gain a 60-90% cost-cut option for routine work. Engineering organizations that have hit Anthropic’s usage caps gain a fallback path. Anthropic faces direct competitive pressure on its Claude Code monetization at the exact moment Uber’s disclosed budget overrun has put pricing in the spotlight; the company’s response will be a useful signal on whether Anthropic considers the routing tool a meaningful threat or a niche workaround. DeepSeek and OpenRouter gain inbound demand from Claude Code users they would not otherwise reach. Cursor, Continue, Aider, and Codeium — all of which compete with Claude Code — have a sharper price benchmark to compare against.
What’s Next
Independent measurement of how DeepSeek V4 Pro performs on real Claude Code workflows compared to Anthropic Opus will determine whether the 80/20 framing holds up in practice. Watch for whether Anthropic responds with pricing changes, terms-of-service updates against backend swaps, or technical changes to the Claude Code client that prevent third-party routing. If the project survives those reactions, expect similar tools for other Anthropic-compatible workflows — and direct competitive responses from xAI’s Grok Code, Cursor’s planned in-house models, and the open-source coding-agent stack.
