The $21 lesson: use decision models for email triage, not LLMs.

The pitch. Jev is a decision model sold as a complement to LLMs: input a state and multiple-choice questions, output answers in parallel with confidence scores. The creator ran 38,000 calls through OpenRouter's decisions endpoint in one week, consuming 55.3 million tokens for $21. The core claim is that Jev is 20 to 200 times faster and 40 to 1,000 times cheaper than an LLM for mechanical choices.

The filtering layer. In practice, Jev replaces the first expensive pass in a second-brain automation. Every incoming email gets a prompt-injection check and a worth-replying decision before any LLM drafts a response. A second use case scores news sources and Reddit posts for relevance and flags likely duplicates against a knowledge base. The pattern is the same: high-volume, low-context decisions happen before the language model runs.

The cost numbers. The strongest evidence is the dashboard: 38,000 calls and 55.3 million input tokens cost about $21 in a week, and Jev answers multiple questions per email in under a second. That is the number worth repeating to a team. It also explains why the creator has shifted some LLM work to cheaper open-weight models like GLM and DeepSeek.

The limit. Jev is not a complete replacement for an LLM. For the news filter, the creator found Jev could skip obvious noise but struggled when an article needed a lot of context, such as whether it fits the channel's audience. That final judgment goes to Claude Haiku or GLM. The same caution applies to prompt-injection detection: Jev's usefulness depends on how well the state and questions are written.

The trade-off. Adoption is not free. The creator iterated on prompts and still sets thresholds, like quarantining only emails with confidence above 0.7. The accuracy claims are personal, not benchmarked. Teams adopting Jev-style models need their own evaluation sets for false positives, especially for security checks, because a cheap wrong answer can be more expensive than an expensive right one.

The signal. The durable takeaway is architectural, not specific to Jev. Put mechanical decisions in a cheap, fast layer and reserve LLMs for judgment that needs context. That pattern reduces token spend, cuts latency, and changes how automations are designed. The video is a useful worked example, not a template to copy blindly.