Jev is HERE. How to use it
Greg Isenberg · 2026-09-18
💡 Quick Take
1. Use Jev for instant email classification, spam scoring, and reply prioritization, but expect it to function as a fast decision-making engine rather than a conversational text model.
2. Use Jev to build instant quote engines, customer support routing, and lead-qualification triage forms that require split-second responses.
3. Don't use Jev for complex reasoning tasks like stock market or Bitcoin trading predictions where traditional LLMs perform better.
4. Use Jev for automated video clip extraction by passing word-level transcripts through custom classifier decisions to spot high-value moments in seconds.
5. Use Jev for browser automation control to drastically accelerate multi-step tasks like finding airline flights compared to traditional AI agents.
6. Use the Vercel AI gateway to get immediate access and start experimenting with Jev if you want to bypass the initial invite-only waitlist.
📊 Detailed Explanation
The video introduces Jev, a fundamentally new type of AI created by Dooo Almeida (known for research that built ChatGPT), which operates as a high-speed, cost-effective classifier rather than a traditional autoregressive LLM. Unlike conversational models that stream tokens and generate unstructured text, Jev takes an input object (such as an email, support ticket, or video transcript), evaluates it against a strictly defined developer output schema, and returns probabilities and classification choices. For instance, in an email triage demonstration processing 1,700 items, Jev used 4.2 million input tokens and 500,000 output tokens to categorize, prioritize, and assign spam and reply scores across the dataset in mere seconds for a total cost of only 18 cents.
Jev's core architectural distinction is that it acts as an AI decision-maker and traffic cop rather than a text generator. Because it does not perform internal text-based chain-of-thought reasoning or conversational generation, it achieves consistent execution speeds of around 200 milliseconds per query regardless of the input-output structure. Developers define an output schema—which enforces type safety so data can be used directly in code without extra processing—and Jev populates those fields with confidence percentages across predefined options (e.g., true/false scales or categorical buckets).
While Jev excels at routing and categorization tasks, the speakers caution against misapplying it to domains requiring deep model intelligence or complex synthesis. In a demonstration attempting to connect Jev to Bitcoin price signals for buy, hold, or sell decisions, the model performed poorly and lagged behind frontier LLMs like OpenAI's GPT-6 Astra that cross-reference external news. Therefore, users should avoid deploying Jev for financial forecasting or autonomous portfolio management while embracing it for structural routing, triage, and data classification workflows.
Practical use cases highlighted in the discussion include powering contact form lead scoring for agencies, instantly matching service aggregation requests with local businesses to provide immediate quotes, routing customer support inquiries to the correct product teams, and extracting highlight clips from long-form video transcripts. Furthermore, combining Jev with browser automation tools enables extreme efficiency gains—such as navigating flight booking sites from Zurich to London in just over seven seconds—significantly outperforming slower traditional AI agents.
For individuals looking to get started, Jev is initially rolling out via an invite-only waitlist, but developers and early adopters can gain immediate access through the Vercel gateway to test packages and build experimental workflows. Because query execution costs are fractions of a cent, users can load nominal credits like five or ten dollars and experiment extensively over multiple days without hitting financial barriers.
🎯 Tech Expert Opinion
Jev represents a refreshing and pragmatic departure from the exhaustive hype cycle surrounding general-purpose, conversational LLMs. By explicitly stripping away token streaming, open-ended generation, and visible reasoning loops, the creators have engineered a specialized classification engine optimized strictly for high-throughput decision-making. For software engineers and startup builders, the introduction of type-safe output schemas eliminates the tedious data-cleaning overhead typically associated with parsing unstructured AI responses, making it a powerful utility for backend data routing.
The speaker's advice is clear and well-calibrated: use Jev aggressively for structured triage, spam filtering, lead scoring, and browser automation tasks where speed and cost efficiency are paramount, but strictly avoid using it for complex analytical prediction or financial trading. This distinction is crucial because treating Jev like a frontier reasoning model will inevitably lead to failure, whereas deploying it as an automated traffic cop for incoming queues unlocks massive operational leverage.
The economic and performance profile of Jev—processing thousands of records for pennies at sub-second latencies—lowers the barrier to entry for building micro-SaaS tools and internal productivity automations. While early adopters must navigate invite restrictions or utilize gateway access points like Vercel, the underlying mechanics demonstrate a viable blueprint for deterministic AI workflows that developers should test immediately.
Kanal: Greg Isenberg