TL;DR — Claude Code always calls Anthropic's API with whatever model you've configured. Configure it with
model: autopointed at CodeRouter, and a phase-aware router decides per-call which model runs — keeping Opus for genuine planning / hard debug phases, moving the 70% of bread-and-butter calls to Sonnet 4.6 or DeepSeek V3. Same Claude Code UX, 60–80% lower bill.
Why Claude Code is expensive
Claude Code is purpose-built for long agentic coding sessions — Plan mode, long tool chains, file-read/write, Bash execution. It's probably the most efficient coding agent at output quality per engineered result.
The problem is cost per token:
- Claude Code defaults to Opus 4.7 — $15 input / $75 output per 1M tokens.
- Agentic tool use means every step reships the growing context. A 50-step session with 80K accumulated context = ~$4M of context tokens re-processed at Opus rates.
- Plan mode (super useful) runs on Opus too. Fine for the planning call. Wasteful for the subsequent implementation calls that follow the plan.
A typical full Claude Code day with 3–5 planning sessions and 100+ implementation steps easily hits $30–80/day direct to Anthropic. That's $700–2400/month for one heavy user.
How phase-aware routing reduces this
Claude Code's workflow actually maps beautifully onto phase detection:
| Claude Code mode | Phase detected | Router's pick |
|---|---|---|
| Plan mode (<plan_mode> tag) | plan (confidence 0.95) | Claude Opus 4.7 ← stays |
| Tool call: Read followed by user asking a question | plan (multiple file reads = review) | GPT-5.2 / Sonnet 4.6 |
| Tool call: Bash → output contains Error: / traceback | debug | Sonnet 4.6 / DeepSeek R1 |
| Tool call: Edit / Write with implementation intent | implement | Sonnet 4.6 / DeepSeek V3 |
| Tool call: Write to a *_test.py or test_*.ts file | test | DeepSeek V3 / Kimi K2.5 |
| Tool call: Write to a .md file | document | Haiku 4.5 / Gemini Flash |
The router uses Claude Code's own system prompt to fingerprint it with 95% confidence, then uses Plan mode tag + tool-result inspection to pick the right phase. No config needed beyond model: auto.
Setup — Claude Code proxy configuration
Claude Code doesn't currently have a native custom-endpoint field, but it respects standard Anthropic environment variables. Point them at CodeRouter's Anthropic-compatible endpoint:
# Install Claude Code if you haven't already
$ npm install -g @anthropic-ai/claude-code
# Point it at CodeRouter
$ export ANTHROPIC_BASE_URL=https://coderouter.io/api/v1/anthropic
$ export ANTHROPIC_API_KEY=cr_your_coderouter_key
# Start Claude Code
$ claude
Or persist via your shell profile:
# ~/.zshrc or ~/.bashrc
export ANTHROPIC_BASE_URL="https://coderouter.io/api/v1/anthropic"
export ANTHROPIC_API_KEY="cr_your_coderouter_key"
When you next run Claude Code, it will connect to CodeRouter. The router accepts Anthropic's native message format and internally maps to whichever provider the phase detection chose — Anthropic for Opus / Sonnet / Haiku calls, OpenAI / DeepSeek / Google for non-Anthropic picks.
Verifying routing is active
After your first Claude Code session:
- Open the CodeRouter dashboard → Analytics.
- Look at the Phase Distribution chart. You should see a mix — if 100% of requests show as one phase (e.g. all "implement"), detection isn't picking up your agent's patterns and you should check your Claude Code version.
- Check the Models Used chart. A healthy distribution: ~10% Opus 4.7, ~40% Sonnet 4.6, ~25% DeepSeek V3, ~15% Haiku 4.5, ~10% others. All-Opus means the phase detector is treating everything as
plan— bug, report to us. - Weekly Savings card shows what Opus-for-everything would have cost.
Real-world cost example
Claude Code daily session logs from an internal test (~200 tool calls, 4 hours of use):
| Setup | Cost for that day | |---|---| | Claude Code → Anthropic direct (Opus 4.7 default) | $42.18 | | Claude Code → Anthropic direct (Sonnet 4.6 model flag) | $8.94 | | Claude Code → CodeRouter | $3.81 (blended across phases) |
That's 91% vs. Opus-direct, or 57% vs. the already-optimized Sonnet-direct config. Plan mode calls still hit Opus (as they should), but the 160+ implementation/test/docstring/refactor calls that followed got distributed across cheaper models.
FAQ
Will Plan mode still work?
Yes, and that's the whole point — CodeRouter recognizes Plan mode via the <plan_mode> system prompt marker and keeps routing it to Opus. What changes is the post-plan implementation calls that Claude Code makes by the dozen.
What about Claude Code's native MCP tool ecosystem? Unaffected. MCP servers are local; only the LLM calls route through CodeRouter.
Does routing to DeepSeek/Gemini break Claude Code's tool-call format?
No — CodeRouter translates Anthropic's tool_use blocks into OpenAI's tool_calls format for non-Anthropic providers and back. Claude Code sees a consistent Anthropic-shaped response regardless of which upstream served it.
What about Claude Code on Bedrock / Vertex? If you're on Bedrock, keep your existing setup — no CodeRouter needed since Bedrock pricing differs. We integrate with native Anthropic API, not with cloud-provider resellers.