Judgment is no longer a luxury

TypeSafe AI, a San Francisco company that had spent two years out of sight, released on 15 September 2026 a model called Jev that cannot write ^[1]. Jev takes a text and a list of questions whose possible answers are defined in advance, and returns a choice, a score or the probability of a yes, each with its degree of confidence ^[2]. Its founder, Diogo Almeida, describes it as a function call: «unstructured state in, typed probabilistic decisions out» ^[1]. The model charges 0.042 dollars per million input tokens and nothing for output ^[1]. Sixteen days later, on 1 October, Cloudflare and Amazon published open models of the same class, and OpenAI had announced its own two days earlier ^{[3] [4] [5]}.

The launch is best read as an instance of something larger. Artificial intelligence is moving from the conversation into the conditional: it stops producing text for a person who reads it and starts returning a value that another program executes without anyone looking at it. The buyer of intelligence stops being someone who reads and becomes a line of code. What a reader used to correct, by noticing when an answer sounded off, is now left to a number: the confidence threshold above which the program acts alone. That number does not come with the model. Whoever installs it writes it.

A model that cannot write

The company calls the class «System One models», after Daniel Kahneman's distinction between fast, intuitive thinking and slow, deliberate thinking ^[1]. The model's name recalls William Stanley Jevons, the economist who observed that more efficient coal multiplied its consumption ^[1]. Jev accepts text only, allows up to 255 options per question and answers every question in a request in parallel ^[2]. It cannot be downloaded or fine-tuned: it is served through an API and, according to its documentation, «the same weights serve every account» ^[2]. TypeSafe promises responses in 70 to 500 milliseconds and says Jev is not a language model, though it has not published its architecture ^[1]. Its input price is 238 times lower than that of the flagship models of OpenAI and Anthropic, which charge 10 dollars per million tokens ^[6].

Almeida spent four years at OpenAI improving ChatGPT's responses and left in 2024 ^[7]. The firm DCVC led a 40-million-dollar seed round ^[8], which valued the company at 200 million according to a single source cited by Forbes ^[7]. His bet fits in a sentence he gave that investor: in an automated economy, more than 99 % of calls to artificial intelligence will be made by software and not by people ^[8]. Vercel reported that within 24 hours Jev reached nearly 13 % of the paid teams on its model gateway, twice the share of the GPT-5.6 family ^[9]. Almeida told the Wall Street Journal that a quarter of Fortune 500 companies already use it and that the model was processing a trillion tokens a day ^[10]. Neither figure has any source other than him.

Sixteen days

On 1 October Cloudflare published Clef and Clef-flash, built on the Chinese open models Qwen, released under an Apache 2.0 licence and compatible with Jev's interface ^[3]. The company charges 0.24 and 0.09 dollars per million tokens for them and maintains, on its own measurements, that they respond in a median of 209 and 39 milliseconds against 524 for Jev ^[3]. Its announcement says two things that weigh more than the figures. The market for these models, it writes, «is getting increasingly saturated», and a human no longer necessarily needs to be in the loop of an agent's decisions ^[3].

Amazon published Strands Decider 2B the same day, a model of 2 billion parameters that runs on a laptop with a median of 153 milliseconds, together with its training data and scripts ^[4]. Its authors concede that the class is «significantly worse» than reasoning models when the problem is complex. The reason that matters little is also theirs: a decision this cheap can sit in a path of the program where a call to a language model never could ^[4]. OpenAI announced a Decisions API on 29 September in limited preview, with no published price ^[5].

What the type does not guarantee

TypeSafe wrote that Jev «can't hallucinate» ^[1]. The phrase is true in a narrow sense that the company itself explains: the model never returns anything outside the list of answers, and its figure of zero hallucinations «is not empirical» but a guarantee of format ^[1]. Its website admits it plainly: Jev cannot invent a category, but it can choose the wrong one ^[11]. A Hacker News reader put it in one line: type safety is not factual correctness ^[12]. Almeida replied in that thread that the model is probabilistic and that it can also be confidently wrong ^[12].

Third-party tests measure how much. One of them asked it to classify 2,000 synthetic emails and Jev got 62.6 % right, against 81.3 % for Claude Haiku 4.5 and 91.6 % for a hand-written rule, although it cost twelve times less than Haiku ^[13]. The consultancy AY Automate found that Jev got 5 of 112 answers wrong when its confidence was 0.90 or more, all of them because of two labels that overlapped. Its conclusion is that no confidence score warns of a badly defined catalogue ^[14]. Another trial asked for the priority of 300 requests whose rule was not in the text: the model was right 44.7 % of the time and gave its answers a mean probability of 0.74 ^[15]. A cake recipe was classified as a technical issue with a confidence of 0.94 ^[16]. An academic preprint changed the names of two options without touching the criterion and Jev's ability to separate the cases fell from 0.81 to 0.58, with zero type errors ^[17].

Armin Ronacher, chief technology officer of Earendil, gave TechCrunch the most precise description: the design «delegates the hallucination problem a little bit to the user» ^[18]. Whoever uses the model decides whether a 50 % probability is a coin toss and whether a 95 % one is an order.

The factory language

Jev's documentation carries a warning that almost nobody quoted. English is the primary language of its training and where it is most accurate, other languages are handled «but not equally well», and it is advisable to test on your own content before trusting it with a workload that is not in English ^[2]. Amazon trained its model on eighteen English datasets and fitted its calibration only on examples in that language ^[19]. Cloudflare and OpenAI say nothing about languages ^{[3] [5]}.

A developer compared Jev on 3,200 cases written in English and in Spanish, and found that in Spanish it loses between 3.0 and 6.4 points of accuracy and doubles its calibration error on the two hardest tasks ^[20]. Automating what the model answers with a confidence of 0.9 or more covers 72.2 % of the hardest task in English and 63.4 % in Spanish, and what gets automated in Spanish is less accurate ^[20]. Laya, an open model of the same class, sent a caller to sales with a confidence of 1.00 when the caller said «No me pase con ventas, solo quiero saber el horario» — do not put me through to sales, I only want the opening hours ^[21]. No paired test measures Jev, Clef or Strands Decider in Portuguese, and almost all the Spanish measured is Spain's.

Three weeks in the region

Latin American developers did not wait for those measurements. A Brazilian desktop program uses Jev to label a business's WhatsApp conversations —quote, booking, operation— fifteen seconds after the last message ^[22]. Another developer set five systems to serve the patients of a fictional dental clinic: OpenAI's most expensive model resolved 88 % of 26 cases and Jev 81 % ^[22]. A third classifies Pix scam messages on simulated data ^[22]. A tool that runs offline on Laya scored 93 % on news in English and between 68 and 70 % on Brazilian reviews, two different tasks ^[22]. All of them are demonstrations, and none documents a customer.

The region adopted artificial intelligence, and it adopted it by hand. Some 52 % of Brazilian small businesses had used it in the two weeks before Sebrae's survey, more than twice the share of their United States peers ^[27]. What is missing is intelligence wired into a process: among firms of ten or more people, only 15 % of the small ones use an application of that kind, against 50 % of the large ones ^[25]. A decision model is that wiring, because it does not converse with the owner but works inside the program.

The price changes which decisions are worth making. A thousand decisions on a short message cost two cents at list price, so classifying for a year every conversation of a shop that receives 300 messages a day costs less than seven dollars. The WhatsApp inbox, through which 79 % of the Brazilian firms that sell online do their selling, is manual work today: separating the quote from the complaint, the booking from the scam and the urgent from what can wait ^[25]. The first pass costs about a thousandth of what a minimum-wage reviewer earns while resolving one case a minute, and the reviewer is kept for what falls under the threshold.

The queues of the State admit the same arithmetic: Brazil's courts received 40.9 million cases in 2025 and closed the year with 75.5 million pending ^[26]. Its tribunals rate three obstacles to using artificial intelligence as «very high»: the lack of specialised staff, of budget and of hardware ^[28]. An open model that runs on a laptop stops demanding all three, and a first classification of every new case, at 5,000 tokens per file, would cost about 8,600 dollars.

That judgment on every operation was the preserve of the large. Nubank took five years to build the platform that evaluates 450 million events a day ^[29], and Mercado Libre reviewed 1,084 million listings in a semester ^[30]. A text model does not replace those systems, which work on transaction data, but it gives whoever lacks them a judgment over messages and documents. The one who will deliver it is not the laboratory but the software vendor that already serves the shop, the clinic or the law office, because four in five firms buy their artificial intelligence ready-made ^[25].

The condition is the one the facts already showed: the threshold is set by whoever installs it, and the confidence is unmeasured in our languages. Each new class of model is born calibrated in English and reopens the gap, as happened to OpenAI's moderation between 2022 and 2024 ^[31]. A model fitted in English nearly doubles its calibration error in other languages, and a few local examples correct much of it ^[32]. A Spanish speaker notices the bad Spanish of a chat. Nobody notices that of a conditional. The two tests available already recommend writing the criteria in English even when the customer speaks Spanish, and that is the trend to watch: business rules in one language and customers in another ^{[20] [21]}.

The counterpoint has to be conceded in full. Classifying text is not new, and one of the first objections was a question: «So... a classifier model?» ^[12]. Queues move with rules before they move with models: Brazil's court backlog fell because a policy on tax-enforcement suits removed ten million cases in two years ^[26]. The law draws a boundary as well, because labelling a message affects nobody and refusing a loan falls under the right to review of automated decisions ^[33]. Price has stopped being the constraint, and the usual ones remain: labelled data, integration and language.

THE PLAY

1. Start with the WhatsApp inbox. Quote, booking, complaint, scam and «put me through to a person» are five labels, and the sixth is «none of these», because without it one trial flagged none of 30 off-topic messages ^[16].
2. Sell it inside the software they already buy. The model costs cents, so the margin is in integrating it into the point of sale, the appointment book or the accounts.
3. Measure in your language and keep the record. Two hundred real, labelled sentences tell you whether the confidence means anything, and Laya went from 0.396 to 0.912 accuracy with half an hour of fine-tuning in Spanish ^[21]. Put a name and a date on the threshold, and keep every decision with its outcome, which is the only thing that cannot be downloaded.