Latest
Last updated . Newest first; every item links to its original source.
Infron adds Jev 1.13 and Jev 1.14 to its decision-model lineup
Infron announced 17 decision models on its model marketplace for typed outputs and probabilities in classification, routing, scoring and policy checks, listing TypeSafe Jev 1.13 and Jev 1.14 among them; its marketplace lists Jev 1.13 at $0.042 per 1M input tokens with output free.
Jev 1.14 is live on Runware
Runware announced Jev 1.14 on its platform, highlighting the experimental image input and the new Box question that returns where an object sits in an image; its model page lists Jev 1.14 alongside Jev 1.13 via its compatible APIs.
Neon AI Gateway adds Jev 1.14 and OpenJev
Neon announced Jev 1.14 and OpenJev among the models recently added to Neon AI Gateway; its docs list Jev 1.14 as typesafeai-jev-1-14-0, with inference free during the private preview.
AI/ML API launches Jev Router for Agentic Tools
AI/ML API announced Jev Router for Agentic Tools: one MCP endpoint that connects an agent to its catalog of tools and models, with Jev picking the right one for each task. TypeSafe's official account shared the launch.
Source: AI/ML API on X ↗ TypeSafe on X ↗ Use case: LLM tool routing →
Jev 1.14 is live on LLM Gateway
LLM Gateway announced Jev 1.14 on its open-source API gateway, served as jev-1.14.0 alongside Jev 1.13 through its /v1/decisions route.
TechCrunch: OpenAI opens up its Jev competitor as decision-model competition heats up
TechCrunch reports OpenAI made its Decisions API model widely available, entering the category Jev created; it notes TypeSafe CEO Diogo Almeida distrusts benchmarks and urges teams to trial models on their own work, and that TypeSafe's models are currently the most widely queried decision models on OpenRouter.
TypeSafe connector brings Jev to UiPath workflows
UiPath announced a TypeSafe connector in UiPath Integration Service for bringing Jev models into automation workflows; TypeSafe's official account shared it, saying Jev is about an order of magnitude cheaper and faster for judgment steps.
Base Labs open-sources mammoth-1 and elephant-1 decision models
Base Labs released open weights for two post-trained decision models, claiming elephant-1 is the best model under 15B and beats Jev, with a technical report.
A 4B open judge trained on Jev datasets beats Jev 1.14 on held-out calls
Maziyar Panahi reports an open 4B judge trained in 22 minutes on one H200 ($1.87) scored 89.4% vs 82.2% for Jev 1.14 on 4,035 held-out RAG and safety calls; Jev still won on unseen PubMed answers.
Jev 1.14 is available on Databricks
Databricks announced Jev 1.14 for its customers in partnership with TypeSafe; accounts enable it to try it out.
OpenRouter launches Armor, a Jev-powered tool-call guard, in beta
OpenRouter launched OpenRouter Armor in beta, using Jev to score each agent tool call against a set of rules plus a slice of the transcript and flag out-of-policy calls the user never asked for.
LiteLLM adds day-0 support for Jev 1.14
LiteLLM's blog announced day-0 support for jev-1.14.0 on both decision routes (/v1/systemone and /v1/decisions), including image input, the new box question type and long-context mode via forwarded TypeSafe-Danger-Zone headers.
Jev 1.14 is available on Vercel AI Gateway
Vercel's changelog announced Jev 1.14 on AI Gateway as typesafe-ai/jev-1.14.0, including image state and the experimental long-context mode, with per-request Zero Data Retention and No Training options.
Jev 1.14 is available on OpenRouter (typesafe/jev-1.14)
OpenRouter announced Jev 1.14 on its platform the day it shipped, at unchanged pricing ($0.042 per 1M input tokens, output free); it runs on OpenRouter's Decisions API rather than chat completions.
Jev 1.14 released (jev-1.14.0)
TypeSafe's model docs list Jev 1.14 as the newest release, available from October 14, 2026, at the same price as Jev 1.13 ($0.042 per Mtok input, output free) and the same 64k per-request / 32k per-question context. The docs now list jev-latest as pointing to jev-1.14.0.
Workday's Sana builds its UI on the fly with Jev
Workday made Sana the home screen of its HR, finance and IT apps, using Jev to pick prebuilt design-system components as the user types. Chief AI officer Joel Hellermark told press a classification model lets them "literally do it instantly" instead of generating the whole interface.
TypeSafe raises an $870M Series A led by a16z
TypeSafe announced an $870 million Series A at a $7.5B valuation, led by Andreessen Horowitz with Sequoia and DCVC, with Martin Casado joining the board; TechCrunch covered the round weeks after Jev's September 15 launch.
InfoQ covers the Jev launch and early adoption
InfoQ summarises the launch, Vercel AI Gateway and Netlify support, a LangChain TypeSafeClassifier integration, and advice to pin jev-1.13.0 and read the jaggedness page before replacing LLM classifiers.
Independent benchmark of jev-1.13.0 on 37 datasets
An arXiv paper evaluates jev-1.13.0 zero-shot on 37 datasets (346,009 requests for under US$10), reporting it beats Qwen3.8-27B on 27 of 37 and Gemma-4-E4B on all 37, with well-calibrated choice probabilities. Code and raw responses are released.
Jev Router lands on OpenRouter (typesafe/jev-router)
OpenRouter lists TypeSafe: Jev Router, released September 25, 2026: it uses Jev to pick the model and reasoning effort per request behind one OpenAI-compatible endpoint, billed at the price of the serving model.
TypeSafe introduces System One models and Jev
Founder Diogo Almeida announced Jev in early access: a model that returns typed, probabilistic decisions instead of text, trained with Reinforcement Learning for Calibrated Decisions (RLCD).
Earlier
Undated background notes.
Jev 1.13 on Flatkey
Flatkey serves typesafe/jev-1.13 through POST /api/alpha/decisions at $0.042 per 1M input tokens.
Flatkey model page ↗Open reimplementations are multiplying
jevlike, OpenJev, decider, NanoJev, Laya and SemIf explore different parts of the same decision shape.
Browse the directory →Official model and docs
TypeSafe's docs cover model aliases, API shape, limits and documented failure modes.
Read official information →From generated text to typed decisions
Open work explores decision heads, logit readout, KV-cache broadcasting and parallel decoding.
Read the comparison →Return to the primary source
Prices, latency, accuracy and performance claims keep their original links and evidence labels.
Read our standard →