AI · OpenAI · May 18, 2026
$1,300,000 — a swarm of 100 coding agents burned 603 billion tokens in a single month
Vendor
OpenAI
Service
OpenAI API (coding agents)
Amount
$1,300,000 (reported)
Timeframe
One month
Cause
Uncapped agent swarm
Outcome
No resolution reported
Who's exposed: Anyone running fleets of autonomous coding agents without a hard token or spend cap. A hundred agents each re-shipping a full standing prompt and uncompacted history on every request can reach seven figures before anyone checks the meter.
Reported by Tom's Hardware and discussed on r/openclaw. The dollar figure was calculated by OpenClaw's own tooling, so treat it as an estimate rather than an audited invoice. Tom's Hardware
What happened
According to Tom's Hardware, the creator of OpenClaw ran up roughly $1.3 million in OpenAI API tokens in a single month. The bill covered about 603 billion tokens across 7.6 million requests, spread over around 100 coding agents running in parallel. That works out to roughly 79,000 tokens per request. On a coding agent, most of that is not the model reasoning, it is input: the same standing prompt and conversation history re-sent on every loop. The agents ran in a fast mode with no token budget cap, so the spend just kept climbing across the month with nothing set to stop it.
Root cause
An agent swarm was left to run with no ceiling. Each request re-ships the full context, tools, system prompt, and history, so 100 agents looping continuously multiply the same bytes into billions of tokens. Without a per-agent or per-key budget cap, and without structuring calls to reuse OpenAI's cheaper cached input, the cost has no natural stopping point. The number grew quietly until the monthly total reached seven figures.
How to avoid it
•
Set a hard token or dollar budget per API key and per agent, and stop the run when it trips. A swarm with no cap can reach seven figures in a month before anyone reads the meter.
•
Structure calls so the standing prompt and tools hit prompt caching. Most of a coding agent's tokens are re-shipped context, and cached input bills at a fraction of the price.
•
Fork a fresh, minimal context per step instead of carrying full history into every request. Re-sending uncompacted history on every loop is what turns 100 agents into a 603 billion token bill.
CostGoat watches your OpenAI bills, so a surprise like this reaches you as an alert, not an invoice.
More Bill Shock stories
February 25, 2026
$82,314.44 — a stolen Gemini key turned a small monthly bill into a bankruptcy threat in 48 hours
June 7, 2024
$96,000/wk — artists fled Meta's AI policies to Cara, and its serverless bill scaled as fast as its users
November 3, 2023
$121,000 — a buggy auto-translate function called a paid API six billion times in two days
August 28, 2022