- DeepSeek V4.1 Flash made five to seven tool calls without detection in three simulated runs.
- OpenRouter says repeat detection needs normalized arguments and separate step and cost caps.
A deliberately failing task
OpenRouter published a November 12 test of safeguards against AI agents repeatedly calling a failing tool. Its six-model comparison included DeepSeek V4.1 Flash, linking the platform's developer guidance to a Chinese model used in tool-driven workflows.
The company ran a simulated billing agent three times per model, with and without repeat-call detection. Its table reports five, six and seven tool calls for DeepSeek without the guard, and two calls in each guarded run. These are platform-reported results from one constructed task, not an independently reproduced ranking of model reliability.
Several limits still matter
Across the six models, OpenRouter says all were blocked on a third identical call when detection was enabled. It also found that changing arguments could evade exact-match detection, motivating normalization of equivalent inputs.
The guide recommends combining repeat detection with per-run step and cost caps, plus account-side spending limits. The narrow test illustrates why an agent's surrounding software matters alongside its model; it does not establish performance across production workloads.
Sources & context
Go to the original material. Company claims remain attributed to their sources.
01


