--- title: Jev skips token generation entirely, and the speed... description: TypeSafe AI launched Jev, a model that skips token generation and instead returns typed decisions with probabilities and confidence scores for structured... canonical: https://daily.dev/posts/jev-skips-token-generation-entirely-and-the-speed-numbers-are-hard-to-ignore-nv8tizgl7 copyright: © 2026 Daily Dev Ltd. terms: https://daily.dev/tos/ twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Jev skips token generation entirely, and the speed numbers are hard to ignore | daily.dev og:description: TypeSafe AI launched Jev, a model that skips token generation and instead returns typed decisions with probabilities and confidence scores for structured... og:url: https://daily.dev/posts/jev-skips-token-generation-entirely-and-the-speed-numbers-are-hard-to-ignore-nv8tizgl7 og:image: https://api.daily.dev/og/posts/Nv8TIzgl7.png og:image:alt: Jev skips token generation entirely, and the speed numbers are hard to ignore og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Jev skips token generation entirely, and the speed numbers are hard to ignore **[Trends](https://daily.dev/sources/trends)** · 3 min read · 5 upvotes · 0 comments ## Summary TypeSafe AI launched Jev, a model that skips token generation and instead returns typed decisions with probabilities and confidence scores for structured questions, framed as a "smart if-statement" for workflows that need fuzzy judgment calls rather than prose. Claimed benchmarks: 70-500ms responses, 20-200x faster and 40-400x cheaper than comparable LLMs, trained via a calibration approach called RLCD. Early reactions from Santiago Valdarrama and Omar Syed praised the confidence-score routing and parallel structured-question design as architecturally novel, while cautioning that the speed and quality claims are self-reported and unverified pending independent benchmarks. ## Content TypeSafe AI shipped Jev, a model that doesn't generate text. It takes structured questions and returns typed decisions with calibrated probability scores in 70-500ms. That's the whole thing. And somehow it's the most-discussed AI release in developer circles right now. The numbers are doing the work. Vercel replaced GPT Luna for command-safety checks and got 18x faster results at better accuracy. One developer classified 1,018 AI research papers for $0.08 total. Another ran 100 emails through a fraud detection pipeline in 1.42 seconds for $0.07, routing uncertain cases to Kimi K3 as a fallback and hitting 96% accuracy. MotherDuck shipped a `prompt_jev()` SQL function that matched GPT-5.6-Terra's accuracy in 40 seconds for $0.50 versus 32 minutes and $37.58 for the LLM. The cost is $0.042 per million input tokens with output tokens free, which is the kind of pricing that makes people do math in public. The pattern everyone's converging on: use Jev as a fast, cheap first pass, then route low-confidence cases to a heavier model. It's not replacing LLMs, it's sitting in front of them. The skeptics have real points though. Hamel Husain asked the obvious question: "What's wrong with calling it a classifier? It's an established term that describes exactly what this is." One analysis argued Jev's speed advantage comes mainly from single-token generation against constrained choices, not a fundamentally new architecture, and that similar results are achievable by prefilling regular LLMs. Assaf Elovic called it "a step forward but not a disruptor," noting similar classification models already exist and LLMs are still required in any harness that uses Jev. A benchmark test on 77-class banking intent classification found Jev at 79% accuracy versus OpenAI Terra at 83.9%, though the gap closed to near-parity on a 7-class version of the same task. The calibration story is where it gets interesting. Jev returns confidence scores, and they're reportedly well-calibrated, meaning a 95% confidence call is actually right about 95% of the time. That's what makes the routing pattern work: you can trust the score enough to automate the easy cases and escalate the rest. Harrison Chase from LangChain called it a "Decision Model" and expects other entrants in the category soon. The creator of Jev explicitly warns against using it for reasoning tasks, calling demos that do so "a bad idea." The architecture details are thin: TypeSafe mentions a "parallel sampler" and a training method called RLCD (reinforcement learning for calibrated decisions), but hasn't published much beyond that. Demand was high enough that TypeSafe had to throttle API access at launch. Cloudflare Workers and the Laravel AI SDK already have integrations. The question now is how long before Anthropic, OpenAI, or Google ships something similar and makes the moat question moot. ## Questions this post answers ### What is Jev and how is it different from a typical LLM? Jev, launched by TypeSafe AI, does not generate text token by token. Instead it takes structured questions and returns typed decisions with probabilities and confidence scores, aimed at use cases like classifying support tickets as urgent or checking policy violations. Claimed performance is 70-500ms response times, 20-200x faster and 40-400x cheaper than comparable LLMs, since output tokens are free. _Developers weighing structured-decision models over text-generating LLMs can track how this plays out on daily.dev._ ### What training approach does Jev use to calibrate its confidence scores? Jev uses an approach TypeSafe AI calls RLCD, which calibrates probabilities to reflect how often the model is actually correct. This lets a confidence score double as a routing signal: when the model isn't sure, the output can be sent to human review instead of trusted directly, which Santiago Valdarrama highlighted as the interesting architectural piece. _Teams designing human-in-the-loop review pipelines can follow calibration approaches like this via daily.dev._ ## Community take How the wider developer community reacted, aggregated from 1 discussion and 17 comments across x (as of 2026-09-21). **TL;DR:** People are impressed by the raw cost and speed of using Jev to classify thousands of papers for pennies, but several push back on the methodology used to validate its accuracy against prior classifications. **Sentiment:** 45% positive · 45% mixed · 10% skeptical **The case for** - Extremely low cost and fast turnaround for large-scale classification tasks (e.g. thousands of papers for cents in under two minutes). - Seen as shifting the practical bottleneck from budget/compute to prompt design and validation. **The pushback** - Validation approach is questioned for only sampling cases where two models disagreed, without a blind or pre-registered method. - Inconsistent numbers across topics raised concerns about bugs or batching issues in the classification pipeline. - Unclear whether large discrepancies with prior tags stem from the new tool's strength or the old taxonomy's ambiguity. **By community** - x (mixed): Enthusiasm about cost and speed is tempered by pointed methodological questions about how the results were validated. **Hottest debate:** Whether the reported agreement/disagreement numbers reflect genuine accuracy gains or an artifact of a flawed sampling and validation methodology. **Open questions** - Was the disagreement sample blind to which model produced which label? - What explains the inconsistent numbers across different topics? - Did ambiguous taxonomy or weak prior classification drive most of the changed labels? **Highlights** > @omarsar0 30 disagreements all accepted is a strong result, but it only samples the stratum where the two models differ. Did you also check a sample where they agreed, and were the labels blind to which model produced which tag? > — [MLaurusevicius on x · 1 points, 1 comments](https://x.com/MLaurusevicius/status/2102071550102933755) > @omarsar0 Agreed on error analysis. The artifact that would help most is the disagreement set itself: which cases you overturned and by what rule, fixed before the run. Otherwise the rule gets fitted to the cases after you see them. > — [MLaurusevicius on x · 1 points](https://x.com/MLaurusevicius/status/2102082318408020445) > @omarsar0 Why aren't these numbers the same? They are for some topics, but quite wildly off on others. > — [stolsvik on x · 1 points, 1 comments](https://x.com/stolsvik/status/2102074951175938328) > @omarsar0 Wait, 75% agreement with the old tags. Either DeepSeek V4 Flash was weak at classification or the taxonomy itself was ambiguous. I'd bet on the latter, but curious which one drove the 579 changes. > — [MilesDigitek on x · 1 points](https://x.com/MilesDigitek/status/2102069952047939829) > @omarsar0 $0.14 for 2,300 papers puts the per-paper cost at 0.006 cents. At that price, the bottleneck shifts completely from budget to prompt design and output validation. > — [jatingargiitk on x · 1 points](https://x.com/jatingargiitk/status/2102073465859956854) **Source threads** - [x](https://x.com/omarsar0/status/2102066232383979749) · 0 points · 17 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#ai-inference](https://daily.dev/tags/ai-inference), [#jev](https://daily.dev/tags/jev), [#llm](https://daily.dev/tags/llm) [View this post on daily.dev](https://daily.dev/posts/jev-skips-token-generation-entirely-and-the-speed-numbers-are-hard-to-ignore-nv8tizgl7) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Jev skips token generation entirely, and the speed numbers are hard to ignore","url":"https://daily.dev/posts/jev-skips-token-generation-entirely-and-the-speed-numbers-are-hard-to-ignore-nv8tizgl7","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/jev-skips-token-generation-entirely-and-the-speed-numbers-are-hard-to-ignore-nv8tizgl7"},"datePublished":"2026-09-15T19:40:24.942Z","dateModified":"2026-10-15T02:22:49.088Z","description":"TypeSafe AI launched Jev, a model that skips token generation and instead returns typed decisions with probabilities and confidence scores for structured...","isAccessibleForFree":true,"articleSection":"Trends","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Trends","logo":"https://media.daily.dev/image/upload/s--ZfSp3asX--/f_auto,q_auto/v1780996004/logos/trends?_a=BAMAMiWQ0","url":"https://daily.dev/sources/trends"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/jev-skips-token-generation-entirely-and-the-speed-numbers-are-hard-to-ignore-nv8tizgl7","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":5},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-agents,ai-inference,jev,llm","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Trends","item":"https://daily.dev/sources/trends"},{"@type":"ListItem","position":3,"name":"Jev skips token generation entirely, and the speed numbers are hard to ignore"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/jev-skips-token-generation-entirely-and-the-speed-numbers-are-hard-to-ignore-nv8tizgl7#faq","mainEntity":[{"@type":"Question","name":"What is Jev and how is it different from a typical LLM?","acceptedAnswer":{"@type":"Answer","text":"Jev, launched by TypeSafe AI, does not generate text token by token. Instead it takes structured questions and returns typed decisions with probabilities and confidence scores, aimed at use cases like classifying support tickets as urgent or checking policy violations. Claimed performance is 70-500ms response times, 20-200x faster and 40-400x cheaper than comparable LLMs, since output tokens are free. Developers weighing structured-decision models over text-generating LLMs can track how this plays out on daily.dev."}},{"@type":"Question","name":"What training approach does Jev use to calibrate its confidence scores?","acceptedAnswer":{"@type":"Answer","text":"Jev uses an approach TypeSafe AI calls RLCD, which calibrates probabilities to reflect how often the model is actually correct. This lets a confidence score double as a routing signal: when the model isn't sure, the output can be sent to human review instead of trusted directly, which Santiago Valdarrama highlighted as the interesting architectural piece. Teams designing human-in-the-loop review pipelines can follow calibration approaches like this via daily.dev."}}]} ```