On 15 September, a start-up released an AI model that can’t write a single sentence. Five weeks later, it has USD 870 million in fresh funding, an OpenAI copy and more than 30 rivals. The model is Jev, from TypeSafe AI, co-founded by Diogo Almeida, a co-author of OpenAI’s InstructGPT paper [1][2]. Here’s what this new category - decision models - is, why it matters and what happened since, plus a durian battle to test it.

What Is It, Exactly?

A decision model answers yes/no, pick-one-of-many and score-on-a-scale questions about a text, each with a probability and a confidence. It reads; it doesn’t write. TypeSafe calls it a “System One” model, after Kahneman’s fast thinking. Ask it about a customer email saying “I was charged twice”: is this a refund request, and which team should handle it? One call returns “yes, 93%” and “billing, high confidence”.

The napkin version: an LLM writes token by token, one trip through the network per token (the loop from my autocomplete post). A decision model reads the input once and lets a tiny head score the options you offered - nothing to generate, so no output to bill, and an answer in a tenth to half a second [1].

A durian opinion with four questions: an LLM writes JSON token by token, Jev returns a probability per option

Figure 1: Four of the 18 questions about one durian opinion. An LLM writes its answer token by token; a decision model reads once and returns a probability for every option (Jev’s real answers).

made with Claude Design · opinion: sixthseal.com (2012)

Inside, the open versions are mostly not an encoder with a classifier, but a pretrained LLM (usually Alibaba’s Qwen) as the “torso”, with small LoRA adapters and a scoring head of about a million parameters (in plain words: an existing language model does the reading, a light add-on tunes it for the job, and a tiny final layer turns what it read into a probability for each option) [3][4][5]. Encoder-only versions exist, but they trail on knowledge [6]. Jev’s own design is unpublished: “not an LLM and not autoregressive, but it is transformer-based” [7].

Why Does It Matter?

It doesn’t decide - it estimates, and tells you how sure it is. So you can accept the confident answers automatically and send only the doubtful ones to a person or a bigger model [8]. In other words, it knows when to ask for help. And it’s cheap enough to use almost anywhere: USD 0.042 per million input tokens, output free [1].

Where does it pay off? Wherever you ask the same few questions about a lot of text:

  • Comment coding. Sort comments into categories, from Google Maps reviews to employee feedback.
  • Risk screening. Spot risks in news articles - easier and cheaper than with an LLM, and a big help for early-warning systems. In one of ours, which checks every article for some twenty risk triggers, a decision model would do the job for about a hundredth of the cost of a mid-size LLM (our estimate at list prices; the cheapest LLMs come much closer).
  • Question routing. Send each question in any chatbot to the right agent or tool.

What Happened Since?

In five weeks, Jev became a category. As of 24 October 2026:

  • Money. USD 870 million at a USD 7.5 billion valuation (9 October), from Andreessen Horowitz (lead), Sequoia, DCVC (which also led the seed) and angels [9][10].
  • OpenAI. Its Decisions API - remember the name - puts GPT-6 Luna behind a typed interface for USD 0.10 per million input tokens; generally available since 22 October, no fine-tuning yet [11][12].
  • Rivals. More than 30 [13], from Microsoft (on Alibaba’s Qwen, at Jev’s exact price), Cloudflare and AWS [14][3][4]. On Artificial Analysis’ independent Decision Index, Liquid AI’s d1-ultra (65.2), Subquadratic’s SubQ Snap 1.1 and Baseten’s Elephant 9B lead; Jev 1.14 is fourth (57.1), but among the fastest at about 80 ms per task - while Elephant scores higher at about a fifth of Jev’s price [15].
  • Open source. Kev (by Jared Palmer, 0.8B to 27B) [5] and OpenJev (a diffusion model underneath) [16] speak Jev’s API, and Ollama and Pydantic AI support such models [14][17].
  • Fine-tuning. Jev and OpenAI: none - only instructions, option descriptions and thresholds. Open weights: fully - Kev-4B went from 80% to 90% on a complaints task after one pass over 5,219 labels, for about USD 1 of GPU time [5].

Does It Work? A Durian Battle

Nobody is neutral about durians, so I let the models judge 83 real opinions about Musang King (also sold as Mao Shan Wang, D197), Black Thorn (D200) and Red Prawn (D175) from English, Chinese and Malay blogs, forums and news - 18 questions per opinion, in one call: which aspects it mentions and how it feels about them, the overall sentiment, whether it compares durians. Fair warning: a fun experiment, not a rigorous benchmark - 83 short quotes, one run, the expected answers drafted with Claude and checked by me. The data, the MIT-licensed code and a tutorial notebook are on GitHub.

Who Reads Durians Best?

System Accuracy Seconds per opinion USD per 1M opinions Confidence to route on?
Claude Opus 5.5 (LLM, reference) 0.95 5.1 ~8,900** no
GPT-6 Luna (LLM) 0.93 3.1 88** no
GPT-5.6 Terra (LLM) 0.92 2.3 2,109** no
Jev (decision model) 0.87 0.3 72 yes
OpenAI Decisions (GPT-6 Luna) 0.85 0.5 318 yes
Clef Flash 9B (decision model, local) 0.84 31.1 free* yes
Kev 4B (decision model, local) 0.79 7.5 free* yes

Accuracy: the mean of four scores against the expected answers. List prices of 24 October 2026; local models ran via Ollama on my MacBook Air.

* No API bill. On a rented AWS GPU (g6.xlarge, one NVIDIA L4, USD 0.80 an hour), we estimate about 0.2 seconds per opinion and USD 50 per million opinions for Kev 4B, and 0.9 seconds and USD 200 for Clef Flash 9B (compute only, without the network trip) - so a self-hosted Kev would be about as fast as Jev and a bit cheaper, though less accurate [18].

** With prompt caching: the instructions and the 18 questions repeat for every opinion, so about 98% of the input costs a tenth of the price - a twentieth for Claude [19][20][21][22]. OpenAI caches automatically; Claude’s caching is opt-in and was off in our run, so the Opus figure is our estimate. Without caching: Luna USD 231, Terra USD 4,974, Opus USD 21,469. The decision APIs can’t do this: Jev has no prompt caching, and OpenAI’s Decisions API caches only the shared text, not the questions [23][12].

  • On raw accuracy, the LLM wins: GPT-6 Luna 0.93, Jev 0.87.
  • With its confidence, Jev wins. Send its least confident 15% of answers to review, and 97% of the rest are right (confidence cut: 80%). Luna gets 94.5% of individual answers right, but can’t tell you which 5.5% are wrong (Figure 2).
  • Ten times faster, a bit cheaper: 0.3 against 3.1 seconds per opinion; USD 72 against 88 per million opinions.
  • OpenAI’s version works, but Jev beat it on every count, confidence included: when OpenAI’s model said it was 80-90% sure, it was right only about half the time. Set the confidence cut for each model separately [17].
  • Open models hold up: on my laptop, Clef Flash and Kev-4B scored 0.84 and 0.79 - slow, but private. A local 8B LLM managed only 0.50.
  • Malay was the hardest language for everyone.
Accuracy of the kept answers rises as more go to review; Jev passes the LLMs at about 9% sent to review

Figure 2: Accept or review. Send the least confident answers to a person, and the rest are right more often - Jev stays on top at every share. The LLMs give no usable confidence, so they can’t play this game.

made with Claude Design · data: jev-on-durians

The Durian Verdict

The people’s Durian Elo - from direct comparisons, writers who tried two durians and every opinion’s score - reads Red Prawn 1613, Black Thorn 1580, Musang King 1430, and Jev, Kev, the Decisions API and the three cloud LLMs crowned the same king. On taste, it’s a near tie (0.68, 0.66 and 0.65 on a 0-1 scale). So why did the king come last? Not the price: drop every opinion about price or value, and it’s still last. It’s the yardstick - all seven direct comparisons in the quotes were against Musang King, and it lost every one - and too much of a good thing: five of its 13 negative or mixed reviews say too sweet, too rich or too heavy. Its fans love exactly that [24]:

Its custardy, luscious, in-your-face swagger hits all the right notes hard and fast, like loud orchestral fanfare.

Christopher Tan, The Peak

The crown goes to Red Prawn, the cheaper everyday alternative (S$12-18 a kilo in Singapore this June, against about S$20 for Musang King) [25]. Next time, more durians: there are only about 30 species, nine of them edible, but hundreds of named varieties [26]; D24 (Sultan), XO and Black Gold are first in line [27].

Durian Elo 2009-2026: Red Prawn ends first, Black Thorn second, Musang King last; six of nine AI systems agree

Figure 3: The Durian Elo, 2009-2026. Red Prawn wins the people’s vote; six of the nine AI systems crowned the same king.

made with Claude Design · data: jev-on-durians

My Take

I’d rename them. “Decisions API” suggests the model decides - it doesn’t. It estimates - a risk, a category, a feeling - and says how sure it is. The decision stays with us: accept the confident answers automatically, send the doubtful ones to review. The durians showed exactly that: a bit less accurate than a cheap LLM on its own, Jev was more accurate on the answers it kept - and told us which ones to check. That’s why I prefer TypeSafe’s own label, System One: fast, cheap and a bit shallow, built for routing and pre-filtering. That’s where it will replace LLMs first: question routing in agentic apps, then guards, validators and tool choice. System One models will save us a lot of money and time: a simple model wherever it’s enough, and the LLM only for the hard cases.

One more lesson from the durians: some decisions you can’t leave to the crowd. The best durian is Musang King, no doubt - and it still came last.

References