Run Claude Code with a different model provider: the variables per vendor, what CC Switch and magpie write, and what breaks

curl -sS -w '\n%{http_code}\n' -X POST "$ANTHROPIC_BASE_URL/v1/messages" \
  -H "Authorization: Bearer $ANTHROPIC_AUTH_TOKEN" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{"model": "claude-sonnet-4-6", "max_tokens": 1, "messages": [{"role": "user", "content": "."}]}'

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.deepseek.com/anthropic",
    "ANTHROPIC_AUTH_TOKEN": "<your DeepSeek API key>",
    "ANTHROPIC_MODEL": "deepseek-flash[1m]",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "deepseek-flash[1m]",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "deepseek-flash[1m]",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "deepseek-flash",
    "CLAUDE_CODE_SUBAGENT_MODEL": "deepseek-flash",
    "CLAUDE_CODE_EFFORT_LEVEL": "max",
    "CLAUDE_CODE_AUTO_COMPACT_WINDOW": "786432"
  }
}

What CC Switch and magpie write to disk

A switcher does steps 2 and 3 for you, and it rewrites your config each time you switch. CC Switch v4.0.11, released 13 October with 144,049 GitHub stars, and magpie v0.1.1233, released 14 October with 12,231 stars, are both MIT-licensed and touch the same file.

Sources: CC Switch config file docs, magpie reference, magpie README.

Three side effects to know. CC Switch does not save the model you picked with /model back to the previous provider. magpie's docs say a key in ANTHROPIC_AUTH_TOKEN signs Claude Code out of claude.ai while it runs through magpie, which turns off claude.ai's plan limits in /usage, its connectors, voice and /teleport, and that Remote Control stays off because Claude Code offers it only on Anthropic's own address. magpie also reads your existing Claude Code and Codex sign-ins to offer them as providers (magpie accounts discovery off stops it) and sends one usage ping a day unless DO_NOT_TRACK=1 is set.

What goes wrong, and the fix

The cache misses and the bill jumps

Prompt caching is where a provider swap costs or saves the most, because a long session resends the whole context every turn. Anthropic's caching docs warn that a gateway which drops the cache markers "bills as uncached input on every turn". DeepSeek ignores the markers and caches on its own, best-effort, so hits are not guaranteed. OpenRouter keeps a session on one provider for its cache until 10 minutes of inactivity. Ollama does not support the markers, which costs nothing on a local model.

Input cost of one turn that resends a 100,000-token context, at each vendor's list price on 14 October 2026 (output and cache writes left out):

Prices: Anthropic, DeepSeek, Kimi, Z.ai, MiniMax. DeepSeek halves both prices off-peak. A missed cache on DeepSeek V4 Pro ($0.132) or even V4.1 Flash ($0.03) costs more than a hit on Sonnet 5.5 ($0.01), so a cheaper model with a broken cache can cost more than the Claude model it replaced. Kimi K3 lists input above Sonnet 5.5, at $3 against $2 per million tokens.

Fix: watch the hit ratio in /usage for the first sessions, and do not switch provider in the middle of a session: CC Switch's docs note that the new model has to rebuild its prompt cache.

Background calls fail with "invalid thinking"

Some models only run with thinking on. Kimi's docs show kimi-k2.7-code, which their own setup puts on the Haiku tier, rejecting requests sent with thinking off with 400 invalid thinking: only type=enabled is allowed for this model, and tell you to keep thinking on in Claude Code. MiniMax's M3.1 Flash Preview also cannot turn thinking off. CC Switch hit this when its Routing mode wrote the Haiku tier as claude-haiku-4-5, a name Claude Code treats as able to run without thinking, and fixed it on 12 October.

Fix: keep thinking on, or put a model that accepts thinking off on ANTHROPIC_DEFAULT_HAIKU_MODEL, and update the switcher.

/context shows a 200k window on a 1M model

Claude Code assumes 200,000 tokens for a model name it does not know, so it compacts early or reports a nearly full context before the first message (issue #94487). Without a matching auto-compact window, it can also run into the vendor's real limit.

Fix: use the [1m] suffix the vendor's docs give, or set CLAUDE_CODE_MAX_CONTEXT_TOKENS and CLAUDE_CODE_AUTO_COMPACT_WINDOW to the model's real window (env vars). An ID with [1m] ignores CLAUDE_CODE_MAX_CONTEXT_TOKENS unless CLAUDE_CODE_DISABLE_1M_CONTEXT=1 is also set.

Requests fail after a Claude Code update

New Claude Code releases send new tool schemas and beta fields that other vendors' endpoints may reject. Anthropic's troubleshooting table gives CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1 for an Extra inputs are not permitted error. For a schema error that starts with one release, check the vendor's and the switcher's issue trackers before downgrading.

Images and web search stop working

DeepSeek V4 Pro does not take images; only Flash does. Claude Code's WebSearch tool runs on Anthropic's search backend, which DeepSeek's endpoint serves itself at extra token cost; Z.ai and MiniMax offer search through MCP servers instead.

When the vendor's own harness is the better choice

If you will run one vendor's model full time, its own harness avoids most of these failures: ZCode for GLM, DeepSeek Harness (a preview), Kimi Code for Kimi, or a model-agnostic one such as Pi, all among the AI coding tools developers list. The swap makes sense when you want Claude Code's hooks, skills and plugins and a cheaper model for part of the work, which is the trade the model routing post prices. On Stackness, as of 14 October 2026, 8 real profiles list Claude Code, 1 lists Kimi K3 and none list CC Switch or magpie, after removing bot accounts (data sources), so these are early setups.

Checklist

[ ] Claude Code 2.1.268 or later
[ ] curl test returns 200 or an unknown-model error, not 401
[ ] ANTHROPIC_BASE_URL without /v1, key in ANTHROPIC_AUTH_TOKEN (Kimi Code: ANTHROPIC_API_KEY)
[ ] Opus, Sonnet, Fable, Haiku and subagent models all set, in ~/.claude/settings.json or settings.local.json
[ ] Haiku-tier model works with the thinking setting you run
[ ] [1m] suffix or CLAUDE_CODE_MAX_CONTEXT_TOKENS matches the real window
[ ] /status shows the base URL and the auth token line
[ ] cache hit ratio checked in /usage; bill read on the vendor's dashboard
[ ] no provider switch mid-session