Claude Code on a third-party model: API Error 400 'type' must be in ["enabled", "disabled", "auto"]

What the error looks like

⏺ API Error: 400 'type' must be in ["enabled", "disabled", "auto"]

400 ValidationException: "thinking.type.enabled" is not supported for this model. Use "thinking.adaptive" and "output_config.effort" to control thinking behavior.

The two messages are the same disagreement seen from opposite sides. The first comes from a host that knows only the older thinking forms; the second from a gateway that has moved to the newer one and now refuses the older form for current Claude models. A switcher in the middle (cc-switch, a local proxy) does not change what Claude Code sends, so it cannot fix this on its own unless it rewrites the field.

What Claude Code actually sends

{
  "model": "deepseek-ai/DeepSeek-V4-Flash",
  "max_tokens": 32000,
  "stream": true,
  "thinking": { "type": "adaptive", "display": "omitted" },
  "output_config": { "effort": "..." },
  "context_management": { "..." : "..." },
  "system": [ { "type": "text", "text": "...", "cache_control": { "type": "ephemeral" } } ],
  "tools": [ "Agent", "Bash", "Edit", "Read", "Write", "... 21 tools" ],
  "messages": [ "..." ]
}

We captured this with a logging proxy between Claude Code 2.1.275 and our endpoint. The thinking field is adaptive whatever the model name, output_config.effort comes with it, and the request is streamed. A user in the cc-switch thread measured the same from a SenseNova endpoint: all 92 failing requests in their proxy log were streaming ones, and direct calls to the host with type: "adaptive" reproduced the rejection.

Fix 1: send the older form and keep thinking

{
  "env": {
    "CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING": "1"
  }
}

With CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1 Claude Code goes back to {"type": "enabled", "budget_tokens": N}. Two users in the thread confirmed this stops the error; we confirmed the variable exists in Claude Code 2.1.275. Put it in ~/.claude/settings.json under env, or export it in the shell. If you use cc-switch, note the report that its proxy mode rewrites settings.json on restart and the block has to be added again.

Fix 2: switch thinking off

{
  "env": {
    "CLAUDE_CODE_DISABLE_THINKING": "1",
    "MAX_THINKING_TOKENS": "0"
  }
}

No thinking field is sent, so nothing can be rejected. The model then answers without its reasoning phase, which is faster and cheaper for routine edits and worse for hard problems. One person in the thread objected to the trade, which is why Fix 1 is listed first.

Fix 3: a host that accepts both forms

export ANTHROPIC_BASE_URL="https://api.inferenceapis.com"
export ANTHROPIC_AUTH_TOKEN="your-key"
export ANTHROPIC_MODEL="deepseek-ai/DeepSeek-V4-Flash[1m]"
claude

Our Messages route treats enabled, adaptive and auto the same way: the model's reasoning stays on and streams back as thinking blocks. No field, or disabled, switches it off. Nothing in Claude Code needs changing. The models are DeepSeek V4 Flash, V4.1 Flash and V4 Pro, GLM 5.3 and 5.3 Flash, Kimi K2.7 Code, MiniMax M3 and GPT-OSS 120B, on GPUs in the United States, billed per token from a prepaid balance. The Claude Code guide has the full setup and what each model did in our test.

Two related symptoms on third-party hosts

• A tool-use turn ends with a thinking block and no text (cc-switch #3654). Claude Code returns to the prompt with "Cooked for 36s" and no summary, and the session log shows a final assistant message with stop_reason: end_turn and only thinking content. That is the host closing the message after the reasoning without a text block, often after a streaming finish_reason it did not understand. Asking the model to continue makes it say the work is done. In our eight test runs every final message had a text block; if you see this with us, send us the session id.
• "deepseek-v4-flash is not a model this version of Claude Code recognizes, so auto-compact will keep this session within 200k tokens" (cc-switch #6221). This is Claude Code being careful, not an error, but compaction at 200k also discards the cached prompt prefix and makes the next turns cost full price. For a 1M-context model, append [1m] to the model name (ANTHROPIC_MODEL="deepseek-ai/DeepSeek-V4-Flash[1m]"; Claude Code strips it before sending) or set CLAUDE_CODE_MAX_CONTEXT_TOKENS to the real window.

Frequently asked questions

Why did this start suddenly?

Claude Code changed what it sends, not the hosts. The thread's first report is July 30, 2026, and one user who downgraded to 2.1.153 still saw it, so the change predates that build. Anthropic's current models take the adaptive form and the client now sends it for every model name.

Does ANTHROPIC_MODEL=deepseek-v4-flash make Claude Code send the older form?

No. The captured request above was for a DeepSeek model name and still carried type "adaptive". Only the environment variables, or a proxy that rewrites the field, change it.

Is "auto" a real value?

Not in Anthropic's documentation, which lists enabled, disabled and adaptive. Some DeepSeek-format Messages endpoints accept auto, which is why it appears in their error message. We accept it and treat it as enabled.

I use cc-switch. Can it fix this for me?

In its proxy mode it could rewrite thinking.type before forwarding; as of the thread it does not, and its restart overwrote the settings.json workaround for one user. Check the project's newer releases, or set the variable in your shell where cc-switch cannot touch it.

Where Inference APIs fits

Our /v1/messages route accepts enabled, adaptive and auto and leaves the model's reasoning on for all three, so Claude Code works with its defaults. Claude Code 2.1.275 ran a full fix-the-tests task through it on eight open models on 2026-09-18; the guide has the setup and the per-model timings.

Related entries

• OpenRouter 402: "Insufficient credits" and "Free BYOK request cap exceeded for this month"
• "thinking is enabled but reasoning_content is missing in assistant tool call message": Kimi K2 tool loops fail on the second turn
• OpenCode Go 403 RegionError: "only available hosted in China and requires explicit opt in"