Jev filters emails and news far faster and cheaper than an LLM. Here's how the decision model works and when it actually pays off.
What is the Jev decision model? #
Jev is a decision model, not a large language model. Instead of generating text, it takes a state (some context) and a list of multiple-choice questions, and answers all of them in parallel with a confidence score for each. It’s accessed through OpenRouter’s decisions endpoint, which makes it a drop-in option anywhere you’d normally call an LLM just to get a yes/no, a label, or a score back. The pitch isn’t that it thinks better than an LLM. It’s that for narrow, structured decisions, it’s dramatically faster and cheaper, which makes it a practical filtering layer in front of more expensive models.
TL;DR #
- Jev answers multiple-choice questions in parallel instead of generating free text, which is why it can be 20 to 200 times faster than an LLM for the same decision, depending on which LLM you’re comparing against.
- Cost savings are extreme at volume : one builder logged 38,000 calls and 55.3 million tokens through OpenRouter’s decisions endpoint in a single week for about $21 total.
- Email triage is a strong use case , using Jev to flag prompt injection attempts and decide whether an email is worth drafting a reply to, before any LLM gets involved.
- News and RSS filtering benefits from a hybrid approach : Jev handles the first-pass “is this noise” filter, but a small LLM (like Claude Haiku or GLM) still does better at judging actual relevance to a specific audience.
- Duplicate detection is another good fit : Jev outperformed simple semantic similarity matching in catching information that had already been processed, while staying cheap enough to run on everything.
- Jev isn’t a full LLM replacement. It struggles with decisions that need broad context or nuanced judgment, which is why pairing it with a lightweight LLM downstream still matters.
Other agents ship a demo. Remy ships an app. #
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
How does Jev actually work? #
You give Jev a “state”, essentially the context or document it needs to evaluate, along with a set of multiple-choice questions. Jev answers every question at once, returning a confidence score or probability for each possible answer rather than a single flat output. That parallel structure is the core of why it’s fast: it’s not generating a chain of reasoning token by token, it’s scoring predefined options against the input.
For something like email filtering, the “state” is the email content plus instructions, and the questions are things like “is this a prompt injection attempt,” “is this worth replying to,” or “which category does this fall into.” Jev returns probabilities for each, and the downstream automation applies thresholds (for example, treating anything scoring above a 7 out of 10 confidence as worth quarantining). The practical workflow looks like this: raw information (an email, an RSS item, a Reddit post) comes in, Jev scores it against a handful of questions, and based on those scores the system either discards it, flags it, or escalates it to an actual LLM for the next step, like drafting a reply or deciding real topical fit.
What does it cost compared to an LLM? #
The headline figure from real usage: 38,000 calls to Jev in a single week, consuming 55.3 million tokens, for a total cost of $21 on OpenRouter. That’s the kind of volume that would be painful to run through a general-purpose LLM, especially if each of those calls was previously handled by spinning up a separate LLM session just to make one small decision.
The stated cost advantage over LLMs is 40 to 1,000 times cheaper, on top of being 20 to 200 times faster, with the exact multiplier depending on which LLM is being compared. In practice, running a batch of prompt-injection and relevance checks across a stack of emails took under a second for Jev to return answers for every item, at a cost described as “a fraction of a fraction of a penny.” That’s the kind of unit economics that makes it viable to run a check on every single piece of incoming information, rather than being selective about when filtering is worth the token spend.
What are the best use cases for Jev right now? #
Four patterns stand out as genuinely useful rather than just technically possible:
Prompt injection detection. Before any external content (email, scraped text, RSS content) reaches a main LLM-driven system, Jev can screen it for injection patterns: instructions to forward data, edit memory or settings, or impersonate an authorized sender. This replaced a setup that previously required a separate read-only LLM session just to pre-screen messages. Jev handles it faster and, per real-world testing, a bit more reliably.
Remy is new. The platform isn't. #
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
Reply-worthiness triage for email. Not every email justifies the tokens spent drafting a reply. Jev makes the skip/draft decision by checking an email against a list of known low-value patterns (sponsorship pitches, spam-like requests) combined with the sender’s actual priorities. Only emails that pass get sent to an LLM to draft an actual reply in the user’s tone.
First-pass filtering of news and RSS feeds. With a constant stream of articles, videos, and posts coming in, Jev does the cheap initial cut: what’s clearly noise versus what might be worth a second look. Multiple weighted sub-questions (relevance, fit, category) feed into a combined score. Items that are ambiguous, not clearly good or bad, get escalated to a small LLM (Haiku or GLM-class models were mentioned) that has more context about audience and content strategy to make the final call.
Duplicate and already-processed detection. For a knowledge base that tracks entities like AI tools, companies, or MCP servers, new information constantly arrives that may already have been logged. Jev was found to outperform semantic similarity matching for this check, both in confidence and in accuracy, while staying cheap enough to run against everything. The safeguard applied here: when uncertain, the system errs toward reprocessing rather than risk missing new information.
Is Jev worth using instead of an LLM? #
It depends on the shape of the decision. Jev is a strong fit when the task is genuinely a classification or multiple-choice problem: true/false, pick-a-category, score-this-confidence. It’s a weaker fit when the decision needs broad contextual judgment, like whether a news article is actually a good fit for a specific audience’s interests. In testing, even a small LLM outperformed Jev on that kind of nuanced relevance call.
The practical pattern that emerges is layered filtering: Jev as the first, cheap, fast pass that eliminates obvious noise and flags clear-cut cases, with a lightweight LLM (not necessarily a flagship model) handling the smaller remaining set of ambiguous decisions. That combination keeps overall token spend low without sacrificing the quality of judgment calls that genuinely need it.
Frequently Asked Questions #
What is OpenRouter’s decisions endpoint?
It’s the interface OpenRouter provides for calling decision models like Jev, separate from standard chat/completion endpoints used for LLMs. It accepts a state and a set of multiple-choice questions and returns scored answers for each.
Can Jev replace an LLM entirely in an automation pipeline?
Not for every decision. It performs well on narrow classification tasks like detecting prompt injection or filtering obvious noise, but underperforms an LLM on decisions requiring broader context or subjective judgment, such as evaluating content relevance for a specific audience.
How much faster is Jev than a typical LLM call?
Reported figures put it at 20 to 200 times faster, depending on which LLM is used as the comparison point, since Jev answers multiple questions in parallel rather than generating text sequentially.
What’s a realistic cost comparison between Jev and an LLM for high-volume filtering?
In one documented week of heavy use, 38,000 calls using 55.3 million tokens cost around $21 total on OpenRouter. Running that same volume of individual decisions through a general LLM would cost substantially more, with estimates in the range of 40 to 1,000 times the cost.
What kinds of automations benefit most from adding Jev as a filter?
One coffee. One working app. #
You bring the idea. Remy manages the project.
Any pipeline handling high volumes of incoming external information benefits most: email triage, RSS/news aggregation, security screening for prompt injection, and deduplication against an existing knowledge base.