Anthropic will confirm and fix a server-side Claude context-injection bug that reports millions of phantom tokens and may overcharge usage-billed accounts.
state: resolvedheat: lowuncertainty: highconvergesscott: lowclaude billing-bug context-window server-side-incidentAnthropic
What is this?
Claude users reported requests failing with context-limit errors claiming roughly 5.5 million excess tokens, even for one-word prompts; a separate Claude Code issue reports runaway input-token consumption that depleted API credits. The supplied search summary says Anthropic confirmed and is fixing a server-side bug, but the underlying snippets do not directly establish that confirmation, explain the alleged token injection, or prove that context-error victims were billed for phantom tokens. Other results describe broader alleged Claude billing discrepancies, though their relationship to this specific incident is unclear.
Why it matters to Scott
If confirmed, the incident strengthens Scott’s position that provider token counts and costs require independent receipts, observability, and enforceable spend controls—especially because his projects route model traffic through LiteLLM and retain Opus rescue paths. The operational implication is meaningful, but provisional: the supplied evidence does not establish Anthropic’s confirmation, phantom-token billing, or a fix.
ip:concept.agent-receiptsdev:technology.langfusedev:technology.litellmdev:project.dev-wiki
queries asked of Scott's wikis
- LLM usage metering and billing auditability
- context-window accounting in coding-agent harnesses
- runaway tool context and prompt injection safeguards
- provider incident resilience and spend limits
- token observability for agent workflows
- verifiable AI API billing and dispute mechanisms
Measured heat
no measured readings yet — the hourly heat pass fills this in
How the heat travelled
no chain yet — the hourly chain pass fills this in
Evidence (2) — ⭐ canonical anchor
Interpretation history
2026-07-21T10:23:56Z
The same reporter now says the issue was resolved through a memory-backend fix, shifting this from an active incident to a closed, narrowly documented account failure. There is still no independent corroboration, Anthropic confirmation, or evidence that phantom tokens were actually billed, so the broader billing hypothesis remains unproved.
2026-07-21T10:21:06Z
evidence attached: reddit.post.1v2e9kl — A detailed postmortem independently supports the hypothesis of server-side context injection causing millions of phantom tokens.
2026-07-20T04:31:30Z
grounded: converges/medium — If confirmed, the incident strengthens Scott’s position that provider token counts and costs require independent receipts, observability, and enforceable spend
2026-07-19T11:24:22Z
case created — The report alleges reproducible cross-account failures and ongoing monetary harm that warrants rapid confirmation.
Decision trace
- 07-21 20:23resolveThe same reporter now says the issue was resolved through a memory-backend fix, shifting this from an active incident to a closed, narrowly documented account failure. There is still no independent co
- 07-21 20:21attachA detailed postmortem independently supports the hypothesis of server-side context injection causing millions of phantom tokens.
- 07-21 20:20propose_attachA detailed postmortem independently supports the hypothesis of server-side context injection causing millions of phantom tokens.
- 07-20 15:24feedback_downScott vote via UI
- 07-20 14:31groundIf confirmed, the incident strengthens Scott’s position that provider token counts and costs require independent receipts, observability, and enforceable spend controls—especially because his projects
- 07-19 21:24createThe report alleges reproducible cross-account failures and ongoing monetary harm that warrants rapid confirmation.