ORO
- What happened
- Miners flagged the top-ranked agent, identified as Leen, for exceeding the $5 per-run OpenRouter inference cost cap during races 1 and 2, with other agents including Konoha and Saffron Lantern v1 also named. The SN15 (ORO) team opened an investigation, later confirmed an audit of races 1 and 2 (Oct 22-23) found accounts that had raised the spend limit on their evaluation keys after creation. The team discarded the affected agents, banned 32 accounts covering 51 agent versions, and re-scored Race 2 with 'enjoy' as the new top agent. Separately, several miners reported repeated anti-cheating rejections with 18-hour cooldowns and said the CLI gave no specific reason for the rejections.
- Community
- Miners posted detailed cost and eval-run data to support the cost-cap allegations and asked the team for clarity on anti-cheat policy. Some community members praised the miners who had flagged the exploit, while others continued to report unresolved anti-cheating rejections and asked for help identifying the cause.
- Official
- The team acknowledged the investigation, later confirmed via a detailed post that the audit found manipulated OpenRouter spend limits, and announced that affected agents would be discarded and accounts banned, reminding miners via a pinned message that manipulated keys lead to discard and ban.
In debate · SN15 inference cost cap exploit · 14 participants
-
Miners alleged the top-ranked agent exceeded the $5 per-run cost cap without consequence.
“Why does the $5 limit not apply to this agent? Not fair!!!”
-
The team confirmed an audit found manipulated OpenRouter keys and detailed how the inference cap was supposed to work.
“Every evaluation runs on a per-run OpenRouter key that ORO creates in your account with a fixed spend limit”
-
Some community members defended miners who had helped detect the exploit as strengthening the subnet.
“He even detected LEEN's cheating on openrouter api key. They are making this sn much stronger and stronger.”