Pointing Claude Code at a non-Anthropic model: what you lose

The short answer

Claude Code's features live in three layers, and only the middle one is at risk:

• Client layer (subagents, MCP, file editing, slash commands): runs on your machine. Unaffected by the model.
• Protocol layer (Messages format, streaming events, cache_control, thinking blocks): whatever the gateway implements.
• Model layer (reasoning quality, vision input, cache hits): belongs to the model. Change the model, change the capabilities.

The real trade-off happens at the model layer, not the protocol layer. That sets the right expectation: you are replacing the model, not Claude Code.

What breaks, what does not

Item by item

Image input stops working

When you paste an image, Claude Code encodes it as an image block and puts it in the message. A text-only model either rejects the parameter or silently drops it — and the silent case is worse, because the model starts guessing at what the picture showed.

Workaround: route vision-dependent tasks to a model that accepts images; pure text refactors do not need it.

Extended thinking never appears

Extended thinking is a model-side reasoning block from Anthropic, surfaced in the Messages protocol as a thinking content block. A gateway can pass such blocks through, but the model has to produce them — with DeepSeek that path is empty.

The alternative is to use a reasoning model directly. deepseek-v4-pro-0813 is a reasoning model; its thinking happens inside the model rather than being exposed as blocks. You get "thinks better", not "you can watch it think".

Prompt caching is partially available

This is the easiest item to misjudge, because there are two layers:

• Billing: the catalog has a cached-input price. DeepSeek V4 Pro is ¥4.5 input versus ¥0.45 cached at off-peak — a tenfold gap, which means cache hits are genuinely metered and billed.
• Protocol: Anthropic marks cache breakpoints with cache_control, and that does not map exactly onto upstream cache implementations. The hit rate is decided upstream; the client cannot force it.

So caching does work and does save money, but do not budget using Anthropic's hit rates.

WebSearch is simply unavailable

Claude Code's WebSearch is not a local tool. It is a server-side tool provided by Anthropic. Once the request goes through a third-party gateway, the upstream does not recognise the tool definition and the call fails or returns unimplemented.

Workaround: use a local MCP fetching tool instead — that is client-layer, so it survives the model swap.

What is unaffected

Worth stating plainly, because it decides whether the swap is worth it:

• Editing ability: reading and writing files, multi-file diffs, running tests — all client-side.
• Subagents and parallel work: scheduled by Claude Code itself, not by the model vendor.
• The MCP ecosystem: databases, browsers, internal systems keep working.
• Permissions and hooks: local configuration, model-independent.

In other words, Claude Code's value as an agent harness is fully preserved. What you lose is some properties of Claude the model.

Minimising the loss

1. Match the model to the task: long-file understanding and refactors on deepseek-v4-pro-0813 (1M context, reasoning); typo fixes and commit messages on deepseek-v4.1-flash, roughly ten times cheaper.
2. Handle vision separately: do not hand screenshots to a text model. Describe them, or switch to a vision model for that task.
3. Do not depend on thinking blocks: when you want to see the reasoning, ask for a plan first and then execution.
4. Discount your cache expectations: estimates based on Anthropic hit rates come in low.

Verification checklist

Three checks that the connection is really live:

Set your own KimiSeek API key in the same terminal that will run curl, replacing the placeholder. A credential configured in your client does not automatically populate this shell variable:

# Replace the placeholder with your own key from the dashboard.
export KIMISEEK_API_KEY="ks-your-key"

curl https://api.kimiseek.app/v1/messages \
  -H "x-api-key: ${KIMISEEK_API_KEY:?Set KIMISEEK_API_KEY before running this command}" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4-pro-0813",
    "max_tokens": 512,
    "messages": [{"role": "user", "content": "Hello"}]
  }'

• 200 with a content array → the protocol layer works.
• Claude Code answers "hello" → the client config is right.
• Ask it to read and edit a file → the tool-call chain works.

If any step fails, match the error text in the error reference.

When to go back to the official service

Saving money is not always the right call. Three cases argue for staying official:

• Heavy vision use (mockup to code, UI review)
• WebSearch-driven live research
• Extreme sensitivity to cache hit rates (very long system prompts, high call volume)

For everything else the savings are real. Concrete numbers are in the no-credit-card route.

Versions and verification

• Claude Code: 2.x
• Verified: 2026-10-26
• Sources: Anthropic documentation (Messages protocol, extended thinking, cache_control, server-side tools); KimiSeek catalog capability and pricing fields