Decision models

On Sep 15, 2026 TypeSafe launches Jev. You send a piece of text and typed questions; it returns a probability for every answer you allow, at $0.042 per million input tokens, with no output to bill.³⁴

By Oct 4 we count 109 decision models. OpenAI, Cloudflare, Perplexity and Liquid each announce one between Sep 29 and Oct 1.^1,46,38

Three things change the economics. Perplexity's open weights top the independent board (Oct 5 to 7), OpenAI prices its Decisions API (Oct 6), and Liquid ships open 3B weights for edge devices (Oct 7).^11,9,15,18

Our own runs on Oct 9 add two results. An open 27B matches Jev on our email job, and a new form-triage study shows where every model still fails.^8,7

Jev sets the anchor: $0.042 per million input tokens, nothing for output (Sep 15).³⁴

By Oct 4 the hosted field runs from $0.04 to $0.30 per million input tokens, every one of them input-only.^1,39,40

Since then the floor dropped. Perplexity charges $0.02 for v1.1 (Oct 5), and OpenAI enters at $0.10 with no cached-input discount on that endpoint (Oct 6).^12,15,16

The open bases under most top decision models (Qwen3.8-27B, Gemma 4 31B, Gemma 4 26B-A4B) are already on serverless chat endpoints, at $0.13 to $0.24 per million input tokens on the one provider charted (Parasail, Oct 9).³⁶ Through a chat endpoint each answer also pays for output and asks one question per call. A decision endpoint bills input only and answers every question in one pass.

A decision model takes a piece of text, the state. Here it is a support ticket about a checkout outage.

You ask typed questions: pick one of these teams, yes or no is it urgent, and where it sits on a four-level severity scale. All three are read in the same pass.

It returns a probability for every option and writes no prose, so there is nothing to parse. Jev answers this ticket in 432 ms (our run, Oct 5).

The answer is always one of your options, and it can still be wrong. On a force-push to main, Jev puts 0.89 on "destructive" and Clef 27B 0.94; an untrained 0.4B model (Laya) puts 0.31 on it (our runs, Oct 5 and Oct 9).

A chat request has two phases: prefill reads the prompt in parallel, then decode writes the answer one token at a time.³⁷

A decision request stops after prefill and reads one probability per option. No decode, no output tokens, no cache kept after the request.

Every question about the same state rides on one prefill. TypeSafe reports 13 questions in one call as 12.2 times cheaper and 10 times faster than 13 calls.³⁵

On one GPU a 27B decision model answers in about 100 ms per request (Decision Index board, Oct 7). Hosted round trips we measured run 240 to 641 ms, so the network sets the floor, and nobody has published batched throughput yet.^9,7,8

We ran real agency work through one method: rules first, then the model, then a confidence cutoff fitted on half the items and scored on the other half. Only the keyword set has human labels; the email and intent keys come from a frontier model working blind.^7,2

On Oct 5 we added Cloudflare's 9B Clef Flash and trained the small model on email and intent. It matched Jev's intent spreads and failed on email.

On Oct 9 we added Clef 27B and a fifth job: sorting 1,290 real contact-form submissions into seven kinds.^8,7

A small model trained on 1,314 of our labels wins the narrow keyword sort (0.893, 98% handled alone). Jev and Clef 27B lead the context-heavy email calls; two general LLMs match them on accuracy at 5 to 200 times the cost (our runs, Oct 1 to 9).^7,2

Of 246 fresh emails, 47 need a reply, so answering "no" every time already scores 0.809.

Asked plainly, "does this need me?", no model clearly beats that line. Jev scores 0.650, below it; Clef 27B comes closest at 0.854.

A precise question that names what counts and what doesn't takes Jev to 0.878 on the same emails (McNemar p ≈ 7×10⁻¹⁰), pre-registered and scored once.^7,8

One line of client context, added to help, made the spam gate worse for Jev, GPT-4.1-mini and Haiku on 1,290 form submissions (Oct 9, p < 0.001). The question is part of the product; test any change to it on labels first.⁷

Each model acts alone only above a cutoff fitted to reach 0.90 accuracy. Jev clears it on 91% of the emails for about 2 cents per 1,000.⁷

Clef 27B, open weights hosted on Cloudflare, clears it on 89%, catches more of the real asks (recall 0.79 against 0.68) and is better calibrated (ECE 0.029 against 0.044), at about 7 cents per 1,000 (Oct 9).^8,1

GPT-4.1-mini and Claude Haiku score 0.829 and 0.825 on the same question and settle three in four, at up to 200 times Jev's cost.

Accuracy at the top is close; what differs is whether a 0.9 means 90%. That decides how much work leaves the queue, and it is a property of the model, the question and the cutoff together. (A stub, What a confidence of 0.9 means, will take this further.)

The strongest open decision models are fine-tunes or readouts on a few open bases. Three carry most of the board: Qwen3.8-27B, Gemma 4 31B and Gemma 4 26B-A4B, all on serverless chat endpoints today (one provider shown, Oct 9).³⁶

Seven of the top nine entries on the Decision Index (Oct 7) sit on those three bases. Two need no decision training at all: a stock Gemma 4 31B with a prompt and a readout ranks fifth and ninth.⁹

Perplexity's open v1.1 now scores above hosted Jev, 62.75 to 60.11. The index blends public and private tests, so read the rank as one signal among several: how people are evaluating Jev-like models.⁹

On our own email job, Clef 27B on the same Qwen base matched Jev within noise, 0.870 to 0.878, and the whole run across five jobs cost $0.27 (Oct 9). For the first time the open option is a real one.⁸

Serve a native /v1/systemone on an open base you already host. Seven of the top nine board entries sit on three common open bases, and SGLang and llama.cpp already speak the format.^9,21,20

Decision APIs bill input only, from $0.02 to $0.10 per million tokens (Oct 9), and a chat readout on the same weights pays output and loses fan-out.^12,14

Every public latency figure is one request at a time; the first batched number becomes the one others are measured against.

Label 200 to 500 of your own decisions, test the wording before the model, and gate on a cutoff you fit in shadow mode. The leader's board score has read 59.5, 51.67, 57.91 and 60.11 in three weeks; your own labels do not move.^42,41,9