OpenAI's API spend limits are soft by default: the dashboard shows a monthly cap but warns 'actual costs may exceed this based on usage,' and calls are only actually cut off when a separate 'Enforce Hard Limit' toggle is enabled — a mechanism documented explicitly in a ZDNET walkthrough recommending it be switched on to stop agent overspending. Community threads spanning 2023 to mid-2026 report charges hundreds to over a thousand dollars above set limits, usually after a stolen or leaked key, and one 2026 thread ties a compromised key to 'OpenClaw-related activity' — agent-driven usage draining a hijacked account. Several users further claim even enforced hard limits have failed or that OpenAI quietly replaced a previously functional project-level cap with what one calls 'illusion.' The returned snippets do not surface iambateman's specific $285 report itself, so that incident's details are unverified here — but the product behavior it describes is independently corroborated.
Converges with the autonomy-budget / manners-vs-physics position that a cap only exists when something deterministic actually enforces it — OpenAI shipping enforcement as a default-off toggle, with community reports of charges far past displayed limits and even flipped hard limits failing, is a dated vendor receipt for exactly that claim, and pushes the real control into his own LiteLLM gateway rather than the vendor console. It touches infrastructure Scott actively runs and pays for (his sustained OpenAI API account routed through the gateway), the stolen-key custody and liability axis of the Agent Provenance Stack, and hands the LLM Report / Agent Token Manifesto cost-control territory a concrete, checkable incident.
ip:concept.autonomy-budgetip:concept.manners-vs-physicsdev:concept.deterministic-agent-control-planedev:technology.litellmwork:technology.openai-apiip:framework.agent-provenance-stackdev:project.llmreportradar:concept.inference-economicsradar:concept.credential-securityradar:aws-project-spend-limitsradar:tokenops-whole-run-budgetradar:claude-subscriber-token-theftradar:codex-bedrock-billing-overchargeradar:github-copilot-billing-header-flaw
queries asked of Scott's wikis
- agent run cost caps budget guardrails
- API key custody secrets in shipped agent apps
- inference economics token cost per run
- local inference vs hosted API cost exposure
- runaway agent loop kill switch rate limiting
- liability for stolen API key spend
now 0 pts/hpeak 8 pts/hcomments 0/hpeers p50momentum: steady1 platformsage 405h
points/hour across evidence · reading as of 2026-10-13 05:35:48.225472+11:00 · deterministic, not a model opinion
2026-10-05T10:35:36Z
The Azure Foundry report ($17k billed against $21k displayed credits) is a second, independent, different-vendor instance of console-displayed numbers not binding actual charges, which broadens the case from an OpenAI soft-default quirk to a corroborated cross-provider pattern — thin and unverified on its own, but the first evidence the failure class is not OpenAI-specific. The flagged $78k story's further climb to 82 points is more manipulation-consistent motion on an already-discounted object and prices nothing.
2026-10-05T10:24:08Z
evidence attached: hn.story.49962785 — Second, much larger independent instance of the same failure class — displayed credits/limits not binding actual charges ($17k on Azure Foundry Claude) — directly bears on whether uncapped-charge exposure is a standing cross-provider cost-control trap.
2026-09-27T04:39:15Z
The triggering velocity spike is the already-discounted $78k Codex story climbing 45→64 points in ~10h — growth consistent with the vote manipulation minimaxir flagged, on a single flagged object whose comments are skeptical rather than additive — so it prices attention, not belief, and I'm declining to raise heat against a 95th-percentile numbers line because there is no periphery: no new platforms, outlets, or independently corroborated incidents. Case meaning is unchanged: the soft-default mechanism stands at the documentation level, the $285 first-hand report stands, and OpenAI has not responded.
2026-09-26T22:57:03Z
The dramatic $78k Codex report that was read as materially strengthening the hypothesis is now credibly challenged on its own thread — a vote-manipulation flag from a reputable participant, fresh-account skepticism, and internally implausible token counts — so the case's corroboration reverts to the documented default-off enforcement behavior plus the modest, consistent $285 first-hand report. The trap is real and established, but there is no escalation, no OpenAI response, and the second report likely adds nothing.
2026-09-26T22:24:09Z
evidence attached: hn.story.49861047 — A second and far larger report of unauthorized agent-driven OpenAI charges ($78k, 826 parallel Codex agents, deleted records) materially strengthens the uncapped-agent-cost-trap hypothesis, with the unverified first-party caveat noted.
2026-09-25T21:34:28Z
grounded: converges/high — Converges with the autonomy-budget / manners-vs-physics position that a cap only exists when something deterministic actually enforces it — OpenAI shipping enfo
2026-09-25T21:26:22Z
case created — Concrete first-hand billing incident with a checkable product-behavior claim that directly threatens uncapped charges on agent workloads and is resolvable by OpenAI's own behavior.