DeepSeek V4 Pro with Claude Code: No-Code Setup

DeepSeek exposes an Anthropic-format endpoint at https://api.deepseek.com/anthropic that speaks the Messages API^[1]. Because Claude Code talks to that protocol, it can use DeepSeek as a drop-in backend — no code changes, no forks^[2].

The August 13 GA release notes confirm Claude Code and GitHub Copilot can use DeepSeek directly as a backend^[2]. V4 Pro 0813 is the flagship option here: 96.40% SWE-bench Verified at $0.022 per test^[5].

Set two environment variables and Claude Code routes every request to DeepSeek. The same variables work for Claude Code CLI, GitHub Copilot, and any Anthropic-SDK client^[1][2].

• ANTHROPIC_BASE_URL must include /anthropic — the root https://api.deepseek.com is the OpenAI-format endpoint^[1].
• ANTHROPIC_API_KEY is your DeepSeek key from platform.deepseek.com.
• For GitHub Copilot: the same two variables apply via its Anthropic-compatible provider slot^[2].

# terminal / shell profile
export ANTHROPIC_BASE_URL="https://api.deepseek.com/anthropic"
export ANTHROPIC_API_KEY="<DeepSeek API Key>"

# start Claude Code — no further config needed
claude

DeepSeek maps Claude model names automatically. claude-opus* (and other high-end names) resolve to deepseek-v4-pro; unsupported model names fall back to deepseek-v4-flash^[1]. You can usually leave Claude Code's model setting untouched.

If you want to force a specific model, pass it directly in the Messages API model field — deepseek-v4-pro and deepseek-v4-flash both work on the Anthropic endpoint^[1].

Tool calling, multi-turn conversations, JSON output, and thinking-mode reasoning all work through the Anthropic endpoint^[1][3]. Two caveats: Anthropic-format requests use a reasoning block with effort levels (none/low/high/max) that mirror DeepSeek's reasoning_effort^[3], and legacy Anthropic-only features like MCP-brokered auth are not applicable.

In practice, V4 Pro through Claude Code behaves like a frontier coding agent with DeepSeek pricing: tool-call-heavy sessions bill at $0.66/$1.98 per 1M off-peak after the 8/16 pricing change^[4].

The mental model for the endpoint: DeepSeek speaks Anthropic's Messages wire format, so anything that expects an Anthropic client — Claude Code CLI, Copilot's Anthropic slot, custom Anthropic-SDK apps — treats it as a drop-in. Your DeepSeek key works as the bearer token, and DeepSeek's own docs confirm Claude Code and GitHub Copilot as supported backends^[1][2].

Cost is where this setup surprises people: the same Claude Code session that would burn $1.29 per task on Claude Opus 5 runs for about $0.022 per task on V4 Pro^[5]. If your team already standardized on Claude Code's interface, switching the backend is two environment variables — the UI, keybindings, slash commands, and agent loop stay exactly the same.

• Works: agent tool loops, file editing, terminal commands inside Claude Code.
• Works: reasoning blocks with effort none/low/high/max^[3].
• Works: streaming responses in Anthropic SSE format^[1].
• Not applicable: Anthropic-specific account features, quota bundles, or Opus-exclusive tools.

Against real Claude models, V4 Pro 0813 sits just below the top: 96.40% SWE-bench Verified versus Claude Opus 5's 97.00% — at 1/59th of the per-test cost ($0.022 vs $1.29)^[5]. On BenchLM's composite, Claude-class models lead by a wider margin on long-horizon agentic work^[6].

The pragmatic setup for many teams: Claude Code UI + DeepSeek backend for volume work, and switch ANTHROPIC_BASE_URL back to Anthropic only when you need maximum autonomous agent behavior. Both routes share the same Claude Code interface, so the switch is two environment variables^[1][2].

Before committing to the switch, run a 50-task A/B on your own repo: same prompts against both backends, and compare pass rate, tokens per task, and wall-clock time. The SWE-bench delta (96.4% vs 97.0%) is inside the noise band for most real workloads, but long-horizon autonomy still favors Anthropic on independent agentic evals^[5][6].