`claude setup-token` tokens rejected with "API Error: 400 status code (no body)" on first request — Max 20x account, fresh mints, reproduces locally and in CI

Summary

Every long-lived OAuth token minted by claude setup-token (correct sk-ant-oat01- prefix) is rejected by the API with API Error: 400 status code (no body) on the very first request — while interactive Claude Code on the same machine, same account, works perfectly. Reproduces identically in GitHub Actions (claude-code-action@v1, claude_code_oauth_token input) and locally (CLAUDE_CODE_OAUTH_TOKEN=<token> claude -p "..."). Multiple fresh mints over several hours on 2026-08-15 (~18:00–20:00 UTC) all fail the same way.

Environment

• Account: Claude Max 20x (organizationType: claude_max, billingType: stripe_subscription, subscription active since 2026-02)
• Local: Claude Code 2.1.232 on Windows 11 (Git Bash) — interactive use works
• CI: ubuntu-latest, anthropics/claude-code-action@v1 (also reproduced with the CLI pinned to 2.1.232 via path_to_claude_code_executable)
• Timing note: mints occurred hours after the Aug 14 20:00 – Aug 15 00:11 UTC OAuth incident on status.claude.com was marked resolved

Failure signature

{"type":"system","subtype":"init","message":"Claude Code initialized","model":"claude-sonnet-5"}
{"type":"result","subtype":"success","is_error":true,"duration_api_ms":0,"num_turns":1,
 "total_cost_usd":0,"input_tokens":0,"api_error_status":400,
 "result":"API Error: 400 status code (no body)"}

Assistant message model is <synthetic>, is_api_error_message: true, error: "unknown". Fails in ~250 ms at zero cost — the first request never gets a body back.

Ruled out (each tested)

1. Token value corruption — token piped via stdin with CR/whitespace stripping; multiple careful re-pastes; same result.
2. Wrong token type — confirmed sk-ant-oat01- prefix (not sk-ant-api03/sk-ant-sid01).
3. CI environment — reproduces locally with only CLAUDE_CODE_OAUTH_TOKEN set; action debug logs show the token delivered to the execute phase, no Bedrock/Vertex/base-URL overrides, anthropic_api_key empty.
4. Model — identical 400 with default claude-sonnet-5 and explicit --model claude-haiku-4-5-20251001.
5. CLI version — identical 400 on 2.1.233 (CI default) and 2.1.232 (pinned; the exact version that works interactively on the same machine).
6. Account mismatch — setup-token browser approval performed in the same claude.ai account the CLI is logged into (verified via oauthAccount metadata).
7. Stale mint / revocation by newer mints — every attempt used the newest mint, pasted immediately.
8. Workspace API-key policy (#54502) — not applicable; consumer claude_max org, the workspace setting does not exist.

Expected

A freshly minted setup-token token authenticates for user:inference on a Max account, as documented for CI use.

Possibly related

• anthropics/claude-code-action#1316 (oat01 token 400s in action, works locally — our case fails in BOTH, so possibly a different or wider issue)
• anthropics/claude-code-action#1281 (Max-plan OAuth token failures)
• The Aug 14–15 OAuth incident — is token issuance fully recovered?

Ask

Is long-lived token issuance currently degraded for Max accounts, and is there server-side logging for the 400-no-body path that can identify what the request is missing?