DeepSeek V4 supports Anthropic-compatible routing into Claude Code and Cowork for ~90% lower cost

• DeepSeek's own coding agents guide documents Anthropic-compatible environment-variable overrides, and aibuilderclub_’s setup thread showed that path working inside Claude Cowork with DeepSeek V4.
• In aibuilderclub_’s verification screenshot, Cowork exposed four DeepSeek V4 variants in the model picker, including Flash, Flash 1M, Pro, and Pro 1M, after a local restart.
• Cost is the whole pitch here: jasonzhou1993 framed the swap as roughly 90 percent cheaper, while nummanali described it as about 10x below Claude API pricing.
• ollama turned the same compatibility story into one-line commands, adding ollama launch claude --model deepseek-v4-pro:cloud plus wrappers for Hermes Agent, Codex, OpenClaw, and OpenCode.
• The caveat is that compatibility is still harness-dependent: in the LocalLLaMA thread, one commenter pointed to DeepSeek's docs and local overrides, while another claimed the Claude Code binary itself remains locked down.

You can read DeepSeek's coding agents guide, open the Ollama model page, and compare that against the LocalLLaMA thread where users were already arguing over how far Anthropic-compatible routing actually goes. The odd part is how quickly this split into two tracks: GUI guides for Claude Cowork, and launcher commands that make Claude-style harnesses look like just another frontend over a different model backend.

The core mechanic is simple. DeepSeek documents Claude Code support through Anthropic-style environment variables in its coding agents guide, and aibuilderclub_ used that same pattern to route Claude Cowork to DeepSeek V4.

That matters less as a DeepSeek launch detail than as a harness detail. The interface stays Claude-shaped, but the model call can move underneath it when the endpoint speaks Anthropic's protocol.

After the local apply step, aibuilderclub_’s screenshot showed four DeepSeek variants inside Cowork's picker:

1. DeepSeek V4 Flash
2. DeepSeek V4 Flash 1M
3. DeepSeek V4 Pro
4. DeepSeek V4 Pro 1M

The same post said DeepSeek V4 Pro answered correctly once selected. Around that proof point, jasonzhou1993 and nummanali both pushed the price angle, describing the setup as a way to keep the Claude-style desktop UX while cutting model spend by about an order of magnitude.

Ollama packaged the compatibility story as launcher commands rather than a manual config guide. In ollama's thread, the new cloud model could be invoked four ways:

• ollama launch claude --model deepseek-v4-pro:cloud
• ollama launch hermes --model deepseek-v4-pro:cloud
• ollama launch codex --model deepseek-v4-pro:cloud
• ollama launch openclaw --model deepseek-v4-pro:cloud
• ollama launch opencode --model deepseek-v4-pro:cloud
• ollama run deepseek-v4-pro:cloud

The attached benchmark image also positioned DeepSeek V4 Pro as competitive on coding and agentic evals, while the Ollama model page made the packaging concrete enough to try immediately.

The community discussion was already less tidy than the setup threads. In the LocalLLaMA post, one commenter pasted ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN overrides for local backends, another linked the same DeepSeek coding agents guide, and another argued that Claude Code itself is still locked to Anthropic.

That split is useful evidence on its own. "Claude-compatible" currently means different things depending on the wrapper: Cowork guides showed a working desktop reroute, Ollama exposed a Claude launcher over DeepSeek cloud, and Reddit users still disagreed on whether stock Claude Code behaves like a generic Anthropic-client frontend without extra tooling.