Jev
A decision-oriented classification engine: given predefined options, it quickly returns executable typed decisions with confidence scores, emphasizing high speed and low cost, already integrated with Vercel AI Gateway, Cloudflare, and OpenRouter.
About Jev
Jev is an inference service for decision-making and classification: you define a set of possible output options up front, it evaluates each option against the input context, and it returns a typed decision with calibrated confidence scores (such as a primary action, confidence, and fallback routing) that downstream code can execute directly, with no need to parse natural language generated by an LLM. It is not a general-purpose chat model — it cannot write code, generate natural language, or reason step by step — instead, it does one thing, decision routing, fast and cheap: according to official announcements it can be 20-200x faster and cut costs by 40-400x (output tokens are free). Developers have already used it for classifying architectural blueprints (a set of 26 drawings in 2.9 seconds, about $0.0052) as well as in security operations for threat hunting, detection engineering, and incident response. Note that independent third-party testing showed an actual speedup of 25x rather than the advertised 200x, with a defect detection rate of 6 out of 7 that did not exceed the comparison model — validate it against your own workload before adopting.
Best for
Best use cases
Routing chat/LLM output into predefined options (such as classification, verdicts, or choosing the next action)
Automatically classifying large volumes of items, e.g. sorting thousands of system execution records into categories such as "normal administrative activity / software execution / suspicious use / unknown"
In threat hunting, ranking broad hunt results by relevance and scoring users/hosts/processes/sessions by suspiciousness
In detection engineering, performing false-positive analysis on historical alerts and adding context-based scoring to deterministic detection rules
In incident response, compressing large-scale timelines down to the events most relevant to an intrusion and continuously scoring hosts/users to support prioritization
Large-scale table-level classification of structured documents such as architectural drawings
Features
Typed decision output
Evaluates each option for a given input and returns a decision with calibrated confidence scores that code can execute directly.
Classification over predefined options
Predefine the possible output options and let Jev determine which category the input belongs to — ideal for large-scale automated classification workloads.
Structured output for security use cases
In security workflows, instead of asking "investigate this," you define the question and possible outputs and get usable structured probabilities and decisions.
Pros
- Returns typed decision results with confidence scores that downstream code can execute directly, with no need to parse natural language generated by a model
- Officially published input price is $0.042 per million tokens, with output tokens free
- Already integrated with popular gateways: Vercel AI Gateway, Cloudflare, and OpenRouter (in beta)
- Public case study: classifying a set of 26 architectural drawings took only 2.9 seconds and cost $0.0052
Cons
- It is not a general-purpose conversational model: it cannot write code, generate natural language, reason step by step, or show its process, and it never produces output outside the predefined options
- You must define the complete set of options before use; poorly designed options will limit the output
- The officially advertised speed and cost advantages do not hold in every test: independent third-party testing showed a 25x speedup rather than 200x, and its defect detection rate did not exceed the comparison model