Can You Run Claude Code on DeepSeek's Flash Model?

Hey hey hey, welcome to Agentic Pod! I'm Liam.

And I'm Stella! Liam, DeepSeek's docs now list Claude Code as a supported agent. Like, Claude Code, running on a DeepSeek model? That's real?

It's real, and it takes about three settings. DeepSeek's docs call the model deepseek-flash, and they publish the exact setup. Today: what changed, the steps, and the gotchas nobody mentions.

Gotchas. Okay. I love a gotcha.

Uh, you will. Let's go.

So what's new? I thought DeepSeek already had a Flash model.

It did. The new part is the name. The docs now say deepseek-flash is served by DeepSeek-V4.1-Flash-1017. The old names, deepseek-v4-flash and the vision one, still work, but they route to the new model.

And they bill at Flash price. Right?

Right. You also get deepseek-v4-pro, the bigger one. And DeepSeek says its Anthropic-style endpoint works with any tool that speaks that format.

Wait, any tool? Not just Claude Code?

The docs list Claude Code, Codex, OpenCode, OpenClaw, Hermes, Reasonix, CodeBuddy and Qoder. And there's something called DeepSeek Harness, in developer preview.

Okay, so Claude Code is just one door. Show me how to walk through it.

Step one, you need Node eighteen or newer. Then install Claude Code with npm, and check the install by asking it for its version.

Step two is the API key?

Yep. Make a key on the DeepSeek platform. Then you set a few environment variables. The first is the base URL. It points Claude Code at DeepSeek's Anthropic-style endpoint, instead of the normal Anthropic address.

And my key goes where?

In the auth token setting. Then the model setting is deepseek-flash.

That's it?

The docs also set the Opus, Sonnet and Haiku slots to deepseek-flash, the subagent model, and a bigger auto compact window. I'd copy that block exactly, uh, every last character.

You say that like it's a cake recipe.

It kind of is. Skip one ingredient and the cake changes.

Step three?

cd into your project, and type claude. That's the whole thing.

Give me a first prompt I can try.

Try this. "Read this repo and list the three riskiest functions, with one test idea each. Don't change any files."

Read-only, so my repo's safe. Smart.

Yeah. And once that works, you hit the part that surprised me.

Which is?

The name mapping. If a tool asks for a model starting with claude-opus, DeepSeek quietly maps it to deepseek-v4-pro, and bills you at the Pro price.

So an Opus slot can cost more than I expect.

Yep. Haiku and Sonnet names map to Flash. And any model name that doesn't start with claude gets a four hundred error on that endpoint.

Okay, what else is lurking?

Some fields are ignored. Prompt caching hints, MCP servers passed in the request, and a few content types like documents are not supported.

Hold on. MCP servers are ignored?

In the API request, yes. Check how your own MCP setup behaves before you trust it with real work.

Anything on web search?

If the model decides to search, Claude Code calls it through DeepSeek's API. That adds token cost for summaries, plus the new Search API billing.

So, watch the meter. And prices?

The pricing page is your source, not me. Numbers have been moving, so check it the day you start.

Okay. But is it actually good?

I can't promise it matches the model you use today. DeepSeek's own claim is that V4.1 Flash beats its old Pro model, and that's their testing, not mine.

So run your own test.

Exactly. Take one real task, like a bug you fixed last week, and give it to both. Compare the diff and the bill.

Who should bother?

Anyone burning a lot of tokens on routine work. Test generation, refactors, repo questions. And people who want a second option when one provider has a bad day.

And who shouldn't?

If your project depends on MCP servers or document inputs, test those first.

Okay, honest verdict. And no hedging this time.

Worth an afternoon. Three settings, easy to undo, and you learn what your own work really needs. The setup block and links are in the description. What task would you test first? Tell us in the comments. See you next time!