Error messages

API Error: Claude can't help with this. Start a new session to continue.

API Error: Claude can't help with this. Start a new session to continue.

Claude Code prints this when Anthropic's API refuses to answer. A safety classifier, a second model that screens requests against the Usage Policy, looked at the conversation and said no, and Claude Code turned that into one line:

API Error: Claude can't help with this. Start a new session to continue.

Learn more: https://www.anthropic.com/legal/aup

You probably searched for the older wording. Until v2.1.219 the same refusal read "Claude Code is unable to respond to this request, which appears to violate our Usage Policy", and 489 issues in the anthropics/claude-code repo quote that phrase as of September 25, 2026. We went through the 2.1.282 binary to see what it says now, and the words "Usage Policy" don't appear in it at all.

The two wordings in version 2.1.282

Which one you see depends on whether Claude Code has another model it could hand you to. When there isn't one, you get the short line above, with the model's name in place of "Claude" whenever one is recorded. That's why reports quote "Sonnet 5 can't help with this". In an interactive terminal the last line reads "Send feedback with /feedback or learn more:" instead, pointing at the same page.

When a fallback model does exist, the refusal is longer and names the model that flagged you:

API Error: <model>'s safeguards flagged this message (https://www.anthropic.com/legal/aup). This sometimes happens with safe, normal conversations. Claude Code can't respond to this message with <model>.

Double press esc to edit your last message, or try a different model with /model.

A biology flag swaps the middle sentence for one saying the safeguards "can sometimes flag legitimate biology tasks". Either version can end with a Details: line holding the category the API reported, such as [bio], and a Request ID. Keep that ID, since it's how anyone at Anthropic finds your request.

What sets it off

Anthropic's API documentation says newer models run classifiers that can decline a request, and the reply then carries a stop reason of "refusal" plus a category: cyber, bio, frontier_llm (work that could help build a competing AI model), reasoning_extraction or general_harms, which covers the rest of the Usage Policy. The category can also be empty. For almost every category the docs add that benign work can trigger it too, which is a fairly candid thing to put in a reference table.

Claude Code's own error reference makes the point that matters most here. The check reads the whole conversation, not only your last prompt. What surprised me was how literally that plays out. In a July 2026 report on version 2.1.197, the refused message was "ja starten", German for "yes, go ahead", sent deep into a long session about a dental milling machine's touch screen. The category was bio, and the next three requests were refused as well. A September report on Sonnet 5 describes the same thing: every turn after the first block failed, a plain "are you there?" included.

So sending another message rarely helps, and neither does quitting and coming back with --continue or --resume, because the transcript on disk still holds whatever tripped it. It can even fire on your very first message. That request carries your CLAUDE.md and git status, and a repository full of security or biology material can set it off before you've asked for anything.

What gets you working again

Press Esc twice or run /rewind to go back to a checkpoint before the turn that caused it, then take the work in a different direction. If you can't tell which turn it was, /clear starts a fresh conversation in the same project, and the old one stays available in /resume. In non-interactive mode (-p), where there's no rewind, start a new session without --continue. The docs also say checks vary by model, so --model sometimes gets a different answer.

Interactive sessions often never show you this line at all. With the Switch models when a message is flagged setting on (it's on by default), Claude Code reruns a flagged request on a fallback model and leaves a notice in the transcript instead. To see whether your own setup is the trigger, launch with claude --safe-mode, which leaves out CLAUDE.md, skills, MCP servers and hooks.

Rephrasing, in the docs' sense, means saying more plainly what you're actually building. If the work is legitimate and keeps getting refused, run /feedback and include the Request ID, since the docs name that ID as the thing to quote when you think a refusal is wrong. On the API, a refusal that comes before any output is billed only in the bio, frontier_llm and reasoning_extraction categories (per the API docs in September 2026).

Cybersecurity flags have their own message on Anthropic's API, with a verification program attached, and that's on the cyber safeguards page. On Amazon Bedrock, Google Cloud's Agent Platform and Microsoft Foundry, a cyber flag shows up as this Usage Policy line instead.