Claude Code users report keyword-trigger billing after Opus 4.7 rollout

• A GitHub bug report says the exact case-sensitive string HERMES.md in recent git commit history could push Claude Code requests onto extra-usage billing instead of a Max plan's included quota, burning $200.98 while 86 percent of plan capacity remained, according to homebrewer's HN summary and GitHub issue #53262.
• theo's original repro showed a second trigger, openclaw.inbound_meta.v1, where an otherwise empty repo returned an "out of extra usage" error after a commit message mentioned OpenClaw.
• Anthropic later said affected users would get full refunds plus extra credits, as the HN discussion summary and giffmana's refund screenshot both note.
• The most plausible explanation in community discussion is a brittle anti-abuse heuristic meant to distinguish Claude Code from third-party harnesses like OpenClaw or Hermes, with the fresh HN discussion and homebrewer's core summary both framing the bug that way.
• Separate reports around the same week said Opus 4.7 sessions could burn quota unusually fast or hit limits that did not match visible local usage, according to Issue #54770, Issue #54750, and theo's rebilling complaint.

You can read the full GitHub billing bug, the newer rapid quota consumption report, and Anthropic's own Pro and Max plan help page. The weird detail is that one trigger was the literal filename HERMES.md, while another was JSON mentioning openclaw.inbound_meta.v1 in theo's repro. Anthropic refunded the first bug, but the surrounding threads kept circling back to a bigger reveal: repo history and harness fingerprints appear to be entangled with billing selection.

The cleanest primary record is GitHub issue #53262. The reporter said a recent commit message containing the exact string HERMES.md made Claude Code route requests to extra usage billing, even though their Max 20x plan still showed 86 percent remaining.

The reproduction in the issue is absurdly narrow. add HERMES.md failed, while nearby variants worked:

• add HERMES.md triggered "You're out of extra usage"
• add hermes.md did not
• add HERMES did not
• add HERMES.txt did not

That specificity is why the story spread. The bug report also said the string was picked up from recent git history, not from a live file tree, which matches the issue's own note that Claude Code includes commit content in its prompt context.

A day later, theo's follow-up video post showed a second repro path. In an empty repo, a commit message containing {"schema": "openclaw.inbound_meta.v1"} was enough for claude -p "hi" to return an extra-usage error.

That widened the story from one bizarre filename to a keyword family. the HN page summary and elmean's HN core summary both describe reports that mentioning OpenClaw could either refuse a request or shift it into a more expensive billing path.

The community instantly started poking at the edges. the discussion summary pulled out three failure modes people wanted tested: accidental false positives from ordinary identifiers like OpenClaw(), cheap keyword floods across large repos, and the possibility that a webpage or docs blob could inject the same trigger string into context.

Anthropic's public response, at least on the HERMES bug, was eventually concrete. A Claude Code team member said everyone affected would get a full refund and an extra grant of usage credits equal to their monthly subscription, according to the HN discussion summary.

The original issue says support first rejected compensation for "incorrect billing routing," then engineering stepped in, fixed the overactive anti-abuse system, and closed the issue on April 25. That sequence is recorded in GitHub issue #53262.

The refund story helped on the narrow incident. It did not settle the broader question of whether Anthropic is intentionally classifying some harness-shaped traffic differently from normal Claude Code use.

The strongest community theory is that Anthropic has some server-side heuristic for separating sanctioned Claude Code or desktop traffic from third-party agent harnesses, and that the heuristic leaked into billing. homebrewer's core summary states that theory directly, while the HN discussion summary quotes commenters who linked Hermes and OpenClaw to attempts to stop monthly-plan access from being used as a cheap backend for other tools.

That theory remains community analysis, not an Anthropic policy document. But the pattern in the evidence is hard to miss:

• HERMES.md in commit history changed billing behavior, per GitHub issue #53262
• openclaw.inbound_meta.v1 in a commit message triggered an extra-usage error, per theo's original repro
• commenters in the fresh HN discussion argued the safer architecture would be a hard signature or auth boundary, not silent billing rerouting

The community read that as a product boundary hidden inside prompt context and repo history. Christmas came early for coding agent nerds, just not in the fun way.

The billing-trigger story landed in a week when Claude Code users were already arguing about limits. One open bug, Issue #54770, says Opus 4.7 1M sessions could consume 90 percent or more of an allotment within minutes, apparently because full attached context might be billed every turn. Another, Issue #54750, says Claude Code could show "current session: 100% used" even when the visible local session summary was tiny.

At the same time, bridgemindai's screenshot claimed rate limits had improved sharply, with only 27 percent of the current session and 23 percent of the weekly limit used after a morning of work. theo's rebilling complaint pointed the other direction, saying a rebilling API failure stopped credits from re-upping and locked out downstream users.

Put together, the week produced a messy picture: some users saw better ceilings, others saw quota burn they could not reconcile, and the OpenClaw or HERMES triggers made every billing oddity look more suspicious.

Anthropic's own docs already describe one explicit way Claude Code can jump from subscription usage to API billing. The help page for using Claude Code with Pro or Max says that if ANTHROPIC_API_KEY is set in the environment, Claude Code uses that API key instead of the subscription plan and charges API usage.

The docs also draw a second line that sounds simple on paper and fuzzier in practice this week. Anthropic's costs documentation says Claude Code charges by API token consumption, while also noting that Max and Pro subscribers have usage included in the subscription and that the /usage session cost block is intended for API users.

Those official docs describe a clear, user-controlled switch: subscription auth versus API auth. The HERMES and OpenClaw reports describe something else, where repo context appeared to flip requests into extra billing without any visible auth change.