Typed Decision Models: The Roster and the Benchmark Fine Print

Three weeks after TypeSafe shipped the first hosted one, typed decision models are a crowded category with a de facto standard interface. OpenAI, Cloudflare, Perplexity, Fastino and Torchcast all return the same three output types. That convergence will outlast this week’s leaderboard order.

The idea is narrow on purpose. Rather than generating prose, these systems answer a constrained question with a typed result: one of N options, true or false, or a number in a range, usually with calibrated probabilities. TypeSafe’s launch post calls this “System One” inference and claims 70-500ms responses, 40x-200x faster than frontier LLMs on comparable tasks. Those are vendor numbers with no cross-provider testing behind them.

The roster, by what you can deploy

Availability splits three ways: hosted APIs you rent, open weights you can run, and implementations of someone else’s schema.

OpenAI’s Decisions API reached public beta today via POST /v1/decisions, powered by gpt-6-luna, taking text and images and returning predicate, choice or score. OpenAI’s “~10x faster” figure compares the endpoint to OpenAI’s own Responses API, so it says nothing about Jev or Clef. It is a proprietary endpoint, and an open API shape is a separate thing from open weights.

Cloudflare’s Clef and Clef-Flash are the most flexible option: Apache-2.0 weights post-trained on Qwen3.8-27B and Qwen3.5-9B, with a vision encoder and 64k context, plus Workers AI endpoints if you would rather rent. Perplexity’s Decider v1.1 tops the index at roughly $0.02 per million input tokens, posted today. Torchcast Decision 27B claims 33ms single-question latency on an H100, then inherits CC BY-NC 4.0 from its base model, which rules it out for commercial work. Fastino’s GLiDE scores well, but both of its public pages describe an API with no statement either way on weights, so plan around the hosted endpoint.

Two open projects earn a mention for different reasons. Kev is a LoRA-adapter family from 0.8B to 27B that implements Jev’s schema. Laya is a 421M ModernBERT-large encoder with a decision head and a 322M multilingual variant covering 100-plus languages; its 4.43 composite reflects breadth it was never built for. OpenJev is an MIT-licensed implementation of the pattern with several independent forks, so treat it as a framework rather than a checkpoint.

No market-share or customer data is published for any of these, so “most popular” in this category means most visible and most available.

The Decision Index composite is 80% private

The v0.3 snapshot generated 2026-10-06 17:39 UTC covers 112 models across 110,201 requests, built from 37 public benchmarks plus private extensions. The composite weights public tasks at 20%, private tasks testing the same skills at 50%, and private new-domain tasks at 30%. That design buys resistance to leaderboard training, at the price of independent reproduction for four fifths of the score.

Two consequences follow. The composite measures accuracy on nothing you can name, and the index’s own tie rules mean the 0.10-point gap between GLiDE and Jev carries no information. Launch comparisons from Cloudflare and Fastino used Decision Index v0.2.1, a different methodology version, so Fastino’s claimed 64.81 against Jev’s 57.91 cannot sit beside the numbers above.

The arXiv preprint on Jev 1.13.0 posted 7 days ago covers 37 datasets in 7 categories and reports Jev beating Qwen3.8-27B on 27 of 37 and Gemma-4-E4B on all 37. It has not been peer reviewed, and its scope excludes Clef, Kev, Laya and OpenAI Decisions.

Red Hat and Cisco found conventional classifiers competitive

The useful evidence came from people with a reason to check. Red Hat’s NeMo Guardrails comparison put Jev, Laya, DiffusionGemma, general LLM judges and small classifiers on prompt-injection and content-safety sets. On the injection set, Jev scored 86.35% while a tuned DeBERTa-v3 classifier of roughly 200M parameters hit 89.01%, at 54.1ms median latency against 348-360ms for Jev.

Read the confounds before you spend that number. Small models ran on a MacBook M1 and larger ones on AWS and ROSA GPU clusters, with a transatlantic hop adding about 56ms to the hosted calls. Prompt wording moved results by double digits in both directions: Laya’s safety score climbed from 57.87% to 75.20% with a tuned policy, while the same tuning pass cost Jev 3.67 points.

Cisco’s 0.5% false-positive-budget study is the sharper result for safety work. Training on Cisco’s own taxonomy caught 61% of harmful examples on its multi-turn set, against 34% for Jev and near zero for Laya. On eight public benchmarks Jev’s AUC beat a prompted Gemma 4 31B judge on 7 of 8 with small gaps, so read the finding as a statement about strict thresholds and Cisco’s policy wording rather than general classification skill.

The counter-case to both studies is fair. Their tasks may be simple enough that the results fail to generalize, and trading generality for speed has always been the deal with specialized classifiers. They still agree where it counts: rankings reverse by task, and tuning swings scores further than the gaps between models.

What to do with this

If you are routing requests or classifying short inputs at volume, these APIs delete a pile of parsing and retry code, and Clef or Kev will run on your own hardware under Apache-2.0. Validate on your own labeled data before trusting any index rank, and measure calibration, abstention on out-of-distribution inputs, and latency on the hardware you will deploy to. Where you have labels and a tight false-positive budget, train the small classifier too and make the decision model earn the slot.

Sources: Introducing System One Models & Jev • OpenAI Decisions API docs • Cloudflare: Introducing Clef • Jev Decision Index • Red Hat guardrails benchmark • Cisco: Decision Models on Trial • arXiv 2609.37647v1

Coverage: OpenAI Decisions beta announcement • Perplexity Decider v1.1 • HN: Red Hat benchmark thread • HN: System One launch thread

Related on this blog: [AUTO] Code Mode’s Unsafe Foundation • The ‘Tens of Thousands’ Agent-Incident Claim Has No Public Breakdown • OpenAI Tripled Its ARC-AGI-3 Score by Fixing Its Own Plumbing