MagicTools
Back to Articles
Claude GuidesAugust 6, 20264 views3 min read
claude-codeclaude-code-labmodel-configurationsettingsjsonanthropicmodelcli

Which Model Is Claude Code Actually Using? settings.json vs ANTHROPIC_MODEL vs --model, Tested

Why this deserves scrutiny

A wrong model configuration does not error — the session runs fine, just on a different model. In GitHub issue #82466, a user configured Fable in ~/.claude/settings.json, ran multi-agent workloads for a full day silently on Sonnet, and had to discard everything.

The official docs describe the model field in settings.json, the ANTHROPIC_MODEL environment variable, the --model flag, and the /model command separately — but never one table saying who wins when they coexist. Here are the measured results.

Method: hard evidence from modelUsage

In headless mode with --output-format json, the modelUsage key in the result is the model ID you were actually billed for — more reliable than any UI display or the model's self-description (what a model says about itself can be shaped by system prompts and does not count as evidence):

claude -p "reply ok" --output-format json | python3 -c \
  "import json,sys; print(list(json.load(sys.stdin)['modelUsage'].keys()))"

The tested priority matrix

Environment: macOS + Claude Code v2.1.220. Each row is an independent experiment in a clean directory, evidence taken from modelUsage:

Configuration Actually used Conclusion
Project .claude/settings.json set to haiku, nothing else haiku ✅ project-level settings honored
settings set to claude-fable-5[1m] (1M-context suffix) claude-fable-5[1m] ✅ the [1m] suffix form works too
settings haiku + env ANTHROPIC_MODEL=sonnet sonnet env var overrides settings
settings haiku + flag --model sonnet sonnet flag overrides settings
env haiku + flag --model sonnet sonnet flag overrides env var

The resulting priority chain (high → low):

--model flag  >  ANTHROPIC_MODEL env  >  project settings.json  >  built-in default

It matches the Unix intuition that configuration closer to the invocation wins — but until now that was intuition. Now it is evidence.

About that Windows issue

Issue #82466 reports a different layer: the user-level ~/.claude/settings.json (note: not project-level) being ignored on Windows 11 + PowerShell, and /model <exact-id> in interactive sessions replying "Kept model as X" and refusing to switch. The reporter ruled out env vars, project overrides, and every known on-disk location, and suspects some client-side UI state that never hits disk.

We could not reproduce it on macOS (every layer worked, as the table shows), and the issue has no official response at the time of writing. If your model looks wrong on Windows: do not assume you misconfigured it — collect evidence with the modelUsage method above, then add your environment details to that issue. Every cross-environment report moves the fix closer.

Practical recommendations

  • CI / scripts: pass --model explicitly — highest priority, smallest scope, immune to any persisted state
  • Project-wide default: use the project's .claude/settings.json (checked into the repo, shared with the team) — verified reliable
  • Verification: whenever a model "feels wrong", run one --output-format json call and read modelUsage; in interactive sessions, run /model with no argument and trust the highlighted item in the picker — the issue reporter found the "Kept model as X" text misleading, while the picker highlight reflects the real current state
  • Cost-sensitive setups: modelUsage also carries token counts and cost, handy for auditing bills

Scope and boundaries

  • Test environment is in the scope card above; the matrix covers headless (-p) with project-level settings / env / flag. The user-level ~/.claude/settings.json and interactive /model switching are outside this article's tested scope (the former is unsafe to mutate on a production machine, the latter is discussed in the issue)
  • Bedrock / Vertex channels have their own model ID schemes and env vars; conclusions here apply to direct Anthropic API/subscription only
  • Priority ordering is stable design semantics and should hold across versions; we will update the status here once the Windows issue is fixed
📚 Browse all Claude tutorials →

Related Articles

Claude Code Says No conversation found to continue — Your -p Sessions Are Being Filtered Out

You run a headless claude -p command, then type claude --continue to pick it up interactively — and get No conversation found to continue, even though the session file sits right there on disk. This is a regression introduced in v2.1.90: the --resume picker's 'hide -p/SDK sessions' filter was mistakenly applied to --continue, which should continue the latest session unconditionally. We reproduced it end to end on macOS + v2.1.220 (the GitHub issue reports Fedora, so both platforms confirm), and verified two workarounds: headless -p --continue has always worked, and prefixing CLAUDE_CODE_ENTRYPOINT=sdk-cli lets interactive --continue bypass the filter with full history loaded.

claude-codeheadless+4
claudeAug 6, 20263 min
3

The Claude Code Hooks Stdin Trap: Python Heredocs Eat Your Hook JSON

The docs say hooks receive JSON via stdin — true. But parse it with a python heredoc (python3 - <<'EOF') and you hit a silent failure: the heredoc redirects python's stdin to the script itself, so json.load(sys.stdin) reads nothing. A best-practice hook then swallows the exception and exits 0 — no error anywhere, just a hook that 'somehow does not work'. The real debugging session, the two-line fix, and a lesson: silently fault-tolerant code is your enemy at debugging time.

claude-codehooks+5
claudeAug 6, 20263 min
20

I Put My Opus 5 Config on the Recommended Diet, Then Ran 18 A/B Trials: 21% Cheaper, and I Can't Prove It's Better

"Opus 5 verifies itself, so delete your fallback prompts" is advice you hear everywhere. I took it, then isolated both configs with CLAUDE_CONFIG_DIR and ran 18 headless trials on identical tasks. Output tokens dropped 21% and wall clock 20-29%, with zero counterexamples. But two of the rewritten rules never fired at all, and one result points the other way: the old config's most thorough run covered a strict superset of what the new one found. Why cheaper and better are separate questions.

prompt-engineeringclaude-code+4
claudeAug 5, 20266 min
36

Claude Code's Auto Mode Judges a Model With a Model — When the Model Is Down, You Can't Even Run cat

Auto permission mode calls your session model to judge whether each Bash command is safe — the judge and the worker are the same model. When it goes unavailable you land in a counterintuitive half-paralysis: reading files works, but you can't run a single cat. This is a log of a real debugging session in which I proposed three entirely reasonable hypotheses and knocked all three down with controlled experiments, leaving exactly one dependable way out.

claude-codepermissions+5
claudeAug 3, 20268 min
82

Try these free online tools

JSON FormatterFormat, minify, and validate JSON. Supports drag-and-drop upload, live preview, and error hints.JS / CSS FormatterFormat and minify JavaScript, CSS, and HTML code online. Beautify or compress for production. 100% local.Diff CheckerCompare two texts line by line. Highlights additions and deletions. Unified and split view.Schema Markup GeneratorGenerate JSON-LD structured data for Google rich results. Supports Article, FAQ, Product, LocalBusiness, and more.Regex TesterTest regular expressions in real time with match highlighting, group details, and flag controls.Word CounterCount words, characters, sentences, paragraphs and estimate reading time. Real-time, works offline.

Published by MagicTools