Help
FAQ
Find the cause by symptom. Each error message is explained in Errors. If the answer is not here, check Service status first, then follow Feedback and contact to send us the time of the failed request from your usage logs.
Connecting and setup
First launch says Unable to connect to Anthropic services
The first-run onboarding contacts Anthropic's services, which are often unreachable from mainland China. Open ~/.claude.json (on Windows, %USERPROFILE%\.claude.json) in an editor, add "hasCompletedOnboarding": true at the top level, save, and run claude again.
Still using the official account
Run /status and check the Base URL. If it is wrong, check the settings file path, open a new terminal, or run /logout to leave the claude.ai login.
The VS Code extension does not use CoreCode
Make sure env is in ~/.claude/settings.json, then quit VS Code completely and reopen it. If you signed in to the extension with a claude.ai account, sign out first.
Model does not exist
Use a model name listed on the Pricing page, without a vendor prefix.
No output for a long time
Extended thinking can go minutes without text; that is normal. The gateway's read timeout is 10 minutes.
Errors
401 invalid key
The key was copied incompletely or has been deleted. Copy it again from the API Keys page, and make sure ANTHROPIC_API_KEY is not also set.
403 insufficient quota
Your balance is used up. Top up on the Wallet page.
403 only requests from Claude Code are accepted
Keys in the Claude 满血（仅限CC）group accept requests from the Claude Code CLI only; calls straight from the Anthropic SDK, curl or other clients are refused. To call from those, create another key in the Claude 满血（可外接）group.
429
The key's request rate or concurrency limit was exceeded. Wait a moment and retry; contact us if you need more concurrency.
An error, and not sure whether CoreCode is at fault
Check Service status first. If it looks normal, follow Feedback and contact to send us the time of the failed request from your usage logs.
Billing and caching
Low cache hit rate, higher spend than expected
The most common cause is auto mode, followed by switching models mid-session or pausing for more than 5 minutes. See Caching and billing for how to investigate.
Many Non-stream requests in the usage logs
Claude Code's main conversation always streams, so non-stream requests are usually auto mode's checks before tool calls. See The cost of auto mode.
I only use Opus, but the bill shows Haiku or Sonnet
Background summaries, subagents and auto mode's classifier use other models. See Other models on the bill.