Article
CommunityJev Guides & Articles
VentureBeat: paying LLMs to write text for decisions needing labels
VentureBeat's Sam Witteveen argues a large share of enterprise AI only needs a label, not generated text. Per TypeSafe, Jev returns choices, scores and yes/no answers in 70 to 500 ms at $0.042 per million input tokens, with output tokens free.
- Format
- Article
- Reading time
- 3 min
- Published by
- Community
- Author
- Sam Witteveen
- Last verified
- Oct 2026
Used inJev for Software & DevTools
Highlights
- Witteveen: many enterprise AI stacks ask a language model to generate text, then parse the text back into a label for decisions like ticket queueing.
- Per TypeSafe, Jev takes application state and typed questions and answers in 70 to 500 ms at $0.042 per million input tokens, with output tokens free.
- Within days Jev had integrations from observability platforms to agent frameworks like Pydantic AI and LangChain, and TypeSafe paused its signup queue.
- Open reproductions: SemIf reads probabilities from a frozen Qwen3.5-4B; Laya uses a 421M-parameter ModernBERT backbone that scored near chance until fine-tuned.
- TypeSafe's headline figures, roughly 194x faster and 445x cheaper, come from its own workflow evaluations and likely sit at the high end, the piece notes.
Watch out
Copy reviewed against the published article on 2026-10-17. The speed, price and integration details are TypeSafe's claims or vendor figures; the open-reproduction numbers come from those projects' developers, and the savings example is a worked illustration rather than a measured result.
Questions about this entry
Short answers drawn from the linked project or page.
That much enterprise AI is quietly deciding things, such as which queue a support ticket belongs in, whether a RAG document is relevant, or which model should handle a request, and that stacks pay LLMs to generate text only to parse a label back out.
Per TypeSafe, it takes application state plus typed questions and returns choices, scores and yes/no answers with probabilities attached, typically in 70 to 500 milliseconds, at $0.042 per million input tokens, with output tokens free.
The piece credits stronger pretraining: zero-shot classification was workable by 2019 but brittle outside training domains, while models like ModernBERT, trained on 2 trillion tokens with an 8,192-token context, handle broader request-time questions without task-specific training.
SemIf trains no model and reads option probabilities from a frozen Qwen3.5-4B, taking about a second for 21 questions and agreeing with generated JSON on 18 of 21. Laya uses a 421M-parameter ModernBERT-large backbone answering in roughly 33 to 40 ms per question on an Nvidia T4.
LinkedIn's Erran Berger described distilling a 7-billion-parameter teacher through a 1.7B intermediate down to a 0.6B student serving production traffic. Shopify's Farhan Thawar said its platform distills frontier models into task models that run 2x to 30x cheaper and faster.
Audit LLM traffic into generation versus decisions such as routing, triage, tagging, approval and scoring. A cited worked example: if $6,000 of a $10,000 monthly AI bill is decisions, moving them to a path costing 5% as much could save $5,700 a month.
Jev is in early access, accepts text only, and serves from the West Coast; TypeSafe says it will take time to establish whether its pricing is sustainable; its headline figures are self-tested and likely high-end; and Pydantic warns that input text can steer Jev's answer.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
Hacker News: Simon Willison's overview of Noul, Choice, and Score, and where a decision model fits.
benwerd53 pointsJev introduces a new shape of LLMRead the thread on Hacker NewsHacker News: swyx posts the Latent Space interview with Diogo Almeida.
swyx4 pointsJev: System One Models for Prod, Not God – With Diogo Almeida, CEO, TypeSafe AIRead the thread on Hacker NewsX: Diogo Almeida in five minutes
X: LLMs vs. Jev, clearly explained
More Jev resources like this
An episode of The Merge (by CodeRabbit) recorded at the first Jev hackathon 11 days after launch; TypeSafe DevRel Allie Laabs explains how Jev replaces prompt-parse cycles with typed choice, score and yes/no outputs.
Jev Guides & ArticlesArticle#community#press#podcast
Community
John Lindquist joins Claire Vo to walk through eight Jev use cases, from a real-time voice to-do app to data deduplication with confidence scores, on the episode billed as Jev week at How I AI.
Jev Guides & ArticlesArticle#community#press#podcast
Community
A BIG CHANGE analysis of the Diogo Almeida Latent Space interview: Jev's most consequential proposition is the unit of automation, a bounded judgment inside an existing process, and the launch claim that Jev can't hallucinate needs a narrow reading because schema matching does not establish truth.
Jev Guides & ArticlesArticle#community#press#explainer
Community
On The MAD Podcast, Matt Turck talks with TypeSafe CEO Diogo Almeida about why intelligence per dollar matters more than enterprise adoption, and why RLVR benchmark gains still leave models 95% Einstein, 5% Mr. Bean.
Jev Guides & ArticlesArticle#community#press#podcast
Community
Langfuse's notebook integration traces TypeSafe Jev System One calls through the OpenInference instrumentor, which patches the Python client in place and emits OpenTelemetry spans.
Jev Guides & ArticlesArticle#community#press#explainer
Community
Kenny Rogers explains Jev as a decision model and shows Choice, Score, and Noul responses from calls he placed through OpenRouter on 21 September 2026.
Jev Guides & ArticlesArticle#community#article#explainer
Community
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.