How to use Claude Code with DeepSeek (v4-pro) via powapi

Why use DeepSeek under Claude Code?

Anthropic ships a fantastic CLI, but two things bite real users:

• Cost. Sonnet 4.6 is $3 / 1M input and $15 / 1M output. A busy day on Claude Code can easily burn through $5–$15 in tokens. The Pro plan's included quota disappears faster than most people expect.
• The weekly cap. Anthropic introduced weekly limits on top of the 5-hour rolling window. Hit them mid-week and you're locked out until the next reset.

DeepSeek V4-pro hits ~80% of Sonnet's SWE-bench score at roughly 1/6th the price. The catch: Claude Code talks Anthropic wire format, and DeepSeek (until recently) didn't. powapi bridges the gap — Claude Code keeps thinking it's hitting Anthropic, we route each turn to whichever model fits best.

Setup — 30 seconds, 2 environment variables

Sign up at powapi.io and mint a key from the Keys page. Then point Claude Code at our endpoint by setting two environment variables — either inline or in your shell rc file:

export ANTHROPIC_BASE_URL=https://api.powapi.io/claudecode
export ANTHROPIC_API_KEY=pk_live_yourkeyhere

# launch Claude Code as usual
claude

That's it. Claude Code now routes every request through powapi. The dedicated /claudecode endpoint applies routing rules tuned for the Claude Code traffic profile (long system prompts, parallel sub-agents, tool calls, file edits).

Prefer Cursor, Cline, Roo, Aider, OpenCode or Continue? Each has its own endpoint — see the tool endpoints reference in the docs.

How the routing actually picks a model

powapi runs a small classifier on every incoming turn that scores the request along four axes:

• Difficulty — short edit vs multi-step reasoning vs architectural question.
• Context length — how much of the conversation actually matters for this turn.
• Modality — text only, or does it include images, PDFs, or large diff blocks?
• Tool intensity — single function call vs the model orchestrating 6 tools in sequence.

The classifier itself runs on a fast, cheap model (Llama-3.3 served by Cerebras) and takes <100ms. Based on its score, we pick from:

• DeepSeek V4-flash — trivial turns (renames, single-file edits, "what does this function do" lookups).
• DeepSeek V4-pro — the workhorse. Handles the vast majority of Claude Code traffic.
• Anthropic Sonnet 4.6 — escalation for long-context, vision, or turns where the verifier flagged a low-quality first attempt.

Every routing decision is recorded in your dashboard at /usage, so you can see exactly which models actually answered each request and what they cost.

Cost & latency numbers from production

Real numbers from powapi's production database after ~1500 Claude Code requests in May 2026 (one heavy daily user):

The Sonnet escalations are 2% of requests but 64% of total cost — that's the whole point of cascading. You pay Sonnet prices for the turns that need it, DeepSeek prices for everything else.

When does Sonnet take over?

Three triggers escalate a request to Sonnet, in priority order:

1. Modality. Vision input or PDF parsing? Direct route to Sonnet — DeepSeek V4 is text-only.
2. Context length. Over ~120K tokens in flight? Sonnet's 200K window handles it cleanly; DeepSeek >128K runs into degraded recall.
3. Quality verifier. A small judge scores DeepSeek's response. If it's vague, contradicts the prompt, or hallucinates a function signature, we retry once with a critique prompt on the same adapter (cheap) and then escalate to Sonnet if still low.

The cascade is transparent — Claude Code sees a single successful response and never knows it came from a different provider on the inside.

Sub-agents, tool calls, and vision

Claude Code lean heavily on three Anthropic features. Here's how each one survives the DeepSeek bridge:

• Sub-agents. Each agent launched by Claude Code is a separate API request to powapi, so each one gets its own routing decision. A diff-summarizer agent might land on flash, a refactor agent on V4-pro, a "review this 50-file PR" agent on Sonnet — all in the same parent session.
• Tool calls. DeepSeek V4-pro implements the full Anthropic tool-use schema including parallel_tool_use. Round-trips work the same as native Anthropic.
• Vision & PDFs. Routed straight to Sonnet. Costs about 6× more than text turns but Claude Code uses vision sparingly (mostly for screenshots when debugging UIs), so it barely moves the monthly bill.

FAQ