DeepSeek Provider for Claude Code

DeepSeek-V4-Pro is a coding model with a 1M token context window. DeepSeek-V4-Flash is the cheaper, faster tier. Both sit behind a native Anthropic-compatible endpoint, so Claude Code connects without an adapter or a proxy.

Model specs

The template maps V4-Pro to the Opus and Sonnet roles, and V4-Flash to the Haiku and small/fast roles. Subagent work also runs on V4-Flash, so background calls bill at the cheaper rate.

Thinking mode is enabled with REASONING_EFFORT: high and 32,000 thinking tokens. Effort level is set to max.

Setup

1. Create an account at deepseek.com and generate an API key
2. Run the setup command:

claude-multi add deepseek

3. Paste your API key when prompted

The template sets the base URL, model mappings, thinking parameters, and output limits.

When to pick DeepSeek

Pick DeepSeek if you want per-token billing rather than a monthly plan. You pay for what you use and nothing when you’re idle.

V4-Pro takes the complex coding work. V4-Flash is quick enough for autocomplete, shell commands, and subagent calls, where latency matters more than depth.

If you would rather have a fixed monthly cost, the GLM Coding Plan is the subscription alternative.

Pricing details

DeepSeek is pay-per-token only, with no subscription plan. V4-Flash (the haiku-tier model) costs less than V4-Pro, so background work and subagent calls stay cheap while heavy reasoning goes to the larger model. deepseek.com has the current per-token rates.

Related providers

• GLM - a subscription plan instead of per-token billing
• MiniMax - 1M context, 512K output
• Qwen - Alibaba’s three-tier coder models, pay-per-token