# Jev named a new model category, and its interface was cloned in four weeks

URL: https://www.thedeepfeed.ai/posts/2026-10-16-jev-and-the-decision-model-boom/
Category: Models
Published: 2026-10-16
Author: the-deep-feed
Tags: decision-models, jev, typesafe, open-weights, calibration, cloudflare, openai, qwen, pricing
Kind: deep

> TypeSafe's Jev made fast, calibrated decision models the first new model category since reasoning models, but within four weeks its request format ran on Ollama, Cloudflare, Microsoft and 182 open-weight systems, while Jev earned 0.27% of Vercel's gateway spend.

## TL;DR

- **Jev** shipped September 15 as a model that returns typed probabilities instead of text, priced at **$0.042 per million input tokens** with free output. Its Hacker News launch drew **1,989 points**, more than Claude Opus 5.5, GPT-6 Sol and Luna, or Gemini 4 Argon.
- In four weeks its request format became a standard: **Ollama 0.35**, **Cloudflare Clef**, **Amazon's Strands Decider**, **OpenAI's Decisions API** and **Microsoft-Decision-1** all shipped, and JevBench now ranks **182 open-weight systems**, several above Jev.
- Independent tests put Jev's lead at **1.2 points** over Bespoke's open Nimble on 3,880 human-labeled decisions, with a wider gap on hard knowledge: **0.840 vs 0.675** on MMLU-Pro against Kev-27B.
- The money split two ways: TypeSafe raised **$870M at $7.5B** on October 9, while Vercel reports Jev reached **35% of paid teams** in two weeks but drew **0.27% of gateway spend**, and Microsoft priced Decision-1 at exactly Jev's rate.

On September 15, a seed-stage startup released a model that cannot write a sentence. TypeSafe AI's [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) takes a block of state, such as a support ticket, an email or a JSON record, plus a set of typed questions, and returns probabilities for each allowed answer in one pass. No tokens stream out. Output is free. Input costs [$0.042 per million tokens](https://typesafe.ai/blog/introducing-system-one-models-and-jev), or $42 per billion, which is how TypeSafe chose to quote it. Its founder, Diogo Almeida, [writes](https://typesafe.ai/blog/introducing-system-one-models-and-jev) that at OpenAI he helped build the instruction-following methods that became the research behind ChatGPT; he left two years ago to start TypeSafe, per [TechCrunch](https://techcrunch.com/2026/09/18/a-new-kind-of-ai-model-from-a-chatgpt-inventor-is-thrilling-developers/).

The launch thread on Hacker News reached [1,989 points and 520 comments](https://news.ycombinator.com/item?id=49717558); it stood at 1,784 when this publication [covered the launch](/posts/2026-09-16-typesafe-jev-system-one/) on September 16. Among September's model launches, only GPT-6 Astra (2,279) drew more. A classifier-shaped model from a seed-stage startup outdrew the launches of Claude Opus 5.5 (1,806), GPT-6 Sol and Luna (1,779) and Gemini 4 Argon (1,705).

What followed is the subject of this piece. Within four weeks, the request format Jev introduced was running on Ollama, on Cloudflare's edge, inside Amazon's agent toolkit, behind a new OpenAI endpoint and in a Microsoft model priced at exactly Jev's rate. An independent leaderboard now ranks 182 open-weight systems against it. On October 9, TypeSafe [raised $870 million at a $7.5 billion valuation](https://typesafe.ai/blog/series-ai) anyway.

The working hypothesis going in was that small, fast decision models became the first genuinely new model category since reasoning models, and that the open ecosystem formed around them faster than the company that named them could monetize. The evidence supports the first half and complicates the second. The category is real, and the clones arrived in days rather than the quarters it took open weights to catch reasoning models. But "small" turns out to be the wrong adjective, and the monetization problem is not mainly the clones. It is that the category was designed to be too cheap to bill. Jev became the most-called model on Vercel's AI Gateway in September while drawing [0.27 percent of its spend](https://vercel.com/blog/ai-gateway-production-index-october-2026).

![Timeline schematic of the four weeks from Jev's Sep 15 launch to the Oct 9 raise, with clone releases stacked on each date.](/post-images/2026-10-16-jev-and-the-decision-model-boom/four-weeks-timeline.jpg)

# What a decision model does with the forward pass

A language model runs in two phases. In prefill it reads the whole prompt at once, building an internal representation of every input token in parallel. Then it decodes: it predicts one token, appends it, and runs again, once per output token. Asking a frontier model to classify a ticket as billing, shipping or returns spends most of its wall-clock time in that second phase, generating a JSON wrapper, a label and often a paragraph of reasoning that the calling code throws away.

A decision model stops after prefill. The caller supplies the allowed answers up front, and the model reads probabilities for those answers directly from its internal state. TypeSafe's API exposes three primitives, documented across the launch post and the independent write-ups: a *noul* returns the probability that a statement is true, a *choice* returns a distribution over up to 255 named options, and a *score* places the state on an ordered rubric. Several questions can ride on one state, and all of them are answered together. Because the output space is fixed in advance, a schema violation cannot occur. That is the narrow, defensible sense in which TypeSafe says Jev "can't hallucinate."

TypeSafe has published no paper, no weights and no architecture. The most careful outside reconstruction came two days after launch from Archer Hume, who ran [roughly 10,000 API calls](https://archerhume.com/posts/jevs-architecture-unmasked) and read the timing. His evidence points to a causal transformer that encodes the shared state once, runs each question as an isolated branch against that cached state, and reads answers from a trained output head. Server time barely moved up to about 100 questions per request; a question with 200 options returned as fast as one with two; and a secret planted in one question was invisible to a sibling question but visible when moved into the state. He also found that the API's `output_tokens` field is a billing figure computed from the serialized response, not a count of generated tokens.

The second ingredient is calibration. A text model asked for its confidence generates the characters "0.95" because they are a plausible continuation, not because 95 percent of such answers are correct. Nish Tahir's [build-your-own walkthrough](https://nishtahir.com/build-your-own-decision-model/) makes the gap concrete. He constrained Qwen3-1.7B's output to five option letters and read the softmax as confidence on CommonsenseQA. In the 0.9-to-1.0 confidence bin the model was right 70 percent of the time, and in the 0.8-to-0.9 bin about 47 percent. Fitting a single temperature of 3.80 pulled the bins back toward the diagonal. TypeSafe's claim is that its training method, which it calls Reinforcement Learning for Calibrated Decisions (RLCD), bakes that correction into the model across tasks rather than per dataset. It has not published how.

The naming is deliberate on both counts. "System One" borrows Daniel Kahneman's fast, intuitive mode of thought, set against the slow System Two that reasoning models imitate. "Jev" is for William Stanley Jevons, whose paradox holds that cheaper coal produced more coal use, not less. The company's bet, stated in its [launch post](https://typesafe.ai/blog/introducing-system-one-models-and-jev), is that each order-of-magnitude drop in the cost of a judgment unlocks orders of magnitude more judgments.

![Anatomy schematic: shared state encoded once, isolated question branches, readout heads emitting noul, choice and score probabilities.](/post-images/2026-10-16-jev-and-the-decision-model-boom/prefill-readout-anatomy.jpg)

# A category, by the markers that defined the last one

Classifiers are older than the transformer, and plenty of engineers said so. One machine-learning engineer on Hacker News put the complaint plainly in the thread on Tahir's post:

> As an MLE who has been failing to get anyone interested in classifiers for many years, the hype around Jev makes me scream internally.
>
> — [howunfortunate on Hacker News](https://news.ycombinator.com/item?id=50038233), Oct 10, 2026

The objection is fair about the mechanism and wrong about the category. What separates a model category from a technique is the infrastructure that forms around it, and reasoning models are the useful precedent. When OpenAI shipped o1 in September 2024, the category acquired its own API parameters, its own benchmarks, its own pricing logic (hidden reasoning tokens billed as output) and, eventually, open replications. As this publication documented in June, [the open-weight lag on reasoning peaked at 9.8 months](/posts/2026-06-24-open-weight-reasoning-gap-three-months/) before narrowing to about three.

By those markers, decision models qualified within a month. They have their own endpoint shape: TypeSafe's `/v1/systemone`, which Ollama, Ollaya and Kev serve at the same path and Cloudflare accepts unchanged, plus OpenAI's parallel `POST /v1/decisions`. They have their own benchmarks: JevBench, the Jev Decision Index hosted on Hugging Face, TypeSafe's workflow evals and Bespoke Labs' human-labeled suite. They have their own pricing logic, in which output is free and input is the only meter. And they have a named training objective that competitors adopted by name: Cloudflare's [Clef announcement](https://blog.cloudflare.com/clef-decision-models/) describes "Reinforcement Learning for Calibrated Decisions (RLCD)" as a secondary objective in its own recipe.

The difference from the reasoning precedent is speed. Open reasoning models trailed for most of a year. Open decision models matched Jev's three primitives in four days, its wire format in six, and its accuracy on classification-shaped tasks inside three weeks.

# Four weeks, counted

The sequence below is assembled from primary announcements and repositories. Hacker News engagement is listed where a launch hit the front page. Counting the critiques as well, eleven front-page threads about the category between September 15 and October 15 drew 8,192 points and 2,648 comments.

| Date | Release | What it is | Signal |
| --- | --- | --- | --- |
| Sep 15 | [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) (TypeSafe) | Closed API, $0.042/M input, free output | 1,989 HN points; $40M seed led by DCVC |
| Sep 19 | [Laya](https://laya.convaiinnovations.com/) (ConvAI Innovations) | Apache 2.0 encoders, 322M–421M params, 32.8 ms | 1,363 HN points |
| Sep 21 | [Kev](https://github.com/jaredpalmer/kev) (Jared Palmer) | Qwen3.5/3.8 adapters, 0.8B–27B, drop-in for Jev's SDK | 9.3k GitHub stars |
| Sep 22 | [Jev in 25 lines](https://www.nobodywho.ai/posts/jev-in-25-lines/) (NobodyWho) | Parody: option-letter logits from Qwen3-0.6B | 692 HN points |
| Sep 25 | [Ollaya](https://ollaya.dev/) | Local runtime serving 21 open decision-model families | 618 HN points |
| Sep 28 | [Jeff](https://github.com/firelex/jeff) | 0.8B base plus 15 LoRA adapters, ~30 ms on NVIDIA | 576 HN points |
| Sep 29 | [Ollama 0.35](https://ollama.com/blog/ollama-now-supports-jev-style-decision-models) | `/v1/systemone` endpoint; Bespoke Nimble 9B, Together AI Tev1 | Mainstream local runtime adopts the API |
| Sep 29 | [OpenAI Decisions API](https://techcrunch.com/2026/09/30/openais-jev-clone-could-help-the-frontier-lab-stop-its-swarming-agents/) | GPT-6 Luna, announced at DevDay; now in public beta | First frontier lab entry |
| Oct 1 | [Clef, Clef-flash](https://blog.cloudflare.com/clef-decision-models/) (Cloudflare) | Frozen Qwen3.8-27B and Qwen3.5-9B, Apache 2.0, images | 642 HN points |
| Oct 1 | [Strands Decider 2B](https://techcrunch.com/2026/10/01/amazon-releases-its-own-jev-clone-as-decision-models-flood-the-web/) (AWS Strands Labs) | Qwen3.5-2B torso, pointer head, training data released | TechCrunch: "dozens" of clones |
| Oct 9 | [Clef-omni; Clef-flash repriced](https://blog.cloudflare.com/clef-faster-cheaper-multimodal/) | Audio and video input; Clef-flash now "cheaper than Jev" | Price competition begins |
| Oct 9 | [TypeSafe Series A](https://typesafe.ai/blog/series-ai) | $870M at $7.5B, led by Andreessen Horowitz | 445 HN points, 357 comments |
| Oct 12 | [Microsoft-Decision-1](https://techcommunity.microsoft.com/blog/EducatorDeveloperBlog/not-every-decision-needs-an-llm-getting-started-with-microsoft-decision-1/4563468) | Post-trained Qwen3.5-9B, $0.042/M input, free output | Matches Jev's price exactly |
| At time of writing | [JevBench](https://benchmarkheaven.com/jev-models) | Independent leaderboard, board v1.7.46 | 182 ranked open-weight systems |

Three features of that table matter more than its length. First, almost every entry speaks Jev's wire format. Kev's README says the TypeSafe Python SDK "works against a Kev server unchanged"; Ollama's launch post shows the official SDK pointed at `localhost:11434`; Cloudflare calls Clef "fully Jev-API compatible." TypeSafe defined the interface, and the interface is now a public good.

Second, the entrants span every layer of the stack within a single month: hobbyists (Kev, Jeff, Laya), a local runtime with 183,500 GitHub stars (Ollama), two hyperscalers (Amazon, Microsoft), an edge network (Cloudflare) and the largest frontier lab (OpenAI). Amazon's entry began as a homebrew project by distinguished engineer Marc Brooker, who [documented building a 2B model he called Hobson](https://brooker.co.za/blog/2026/09/28/engineering-system-one.html) on a home RTX 3090 before AWS cleaned it up and released it.

Third, nearly every open entry sits on Qwen. Kev, Jeff, Clef, Strands Decider, Nimble, Decider and Microsoft-Decision-1 all start from Alibaba's Qwen3.5 or Qwen3.8 weights; the two highest-ranked JevBench entries are fine-tunes of Google's Gemma 4 31B. The decision-model boom is, structurally, a post-training boom on someone else's base, the pattern this publication traced in August when [post-training became the new pre-training](/posts/2026-08-14-post-training-is-the-new-pre-training/).

![Hub schematic: one /v1/systemone interface at center, spokes to Jev, Ollama, Ollaya, Kev and Clef, plus OpenAI's parallel /v1/decisions.](/post-images/2026-10-16-jev-and-the-decision-model-boom/one-interface-many-backends.jpg)

# Why it copied so fast

Four properties made the category spread, and the same four made it easy to clone.

**Latency.** TypeSafe quotes end-to-end responses of [70 to 500 milliseconds](https://typesafe.ai/blog/introducing-system-one-models-and-jev) against seconds for a frontier chat model. Independent measurements land in the middle of that range. AbdelStark's [probability-aware benchmark](https://github.com/AbdelStark/jev-benchmarks), calling the hosted API from France, recorded 236 to 256 milliseconds at the median; the open-source [DMB suite](https://github.com/nibzard/decision-model-benchmark) logged 267 to 316 milliseconds across 40,226 decisions. Local open models beat that by removing the network: Kev-4B answers six questions in 18.1 milliseconds of model time on an H100, and Laya's encoder runs in 32.8 milliseconds on a single GPU.

**Cost.** The DMB run of 40,226 decisions cost $1.71 at Jev's published rate. Developers posted head-to-head numbers within a day of launch, including one Vercel engineer whose team swapped Jev into an existing safety classifier:

> We benchmarked fx auto mode (safety) classifier with @typesafeai's Jev.
>
> tl;dr: ~5-18x faster and more accurate than 𝚐𝚙𝚝-𝟻.𝟼-𝚕𝚞𝚗𝚊, our current top choice
>
> — [@fazxes](https://x.com/fazxes/status/2100300097695232164), Sep 16, 2026

**Size.** Here the hypothesis needs correcting. Jev itself is not a small model; TypeSafe's own FAQ says it is "neither small nor an LLM," and Archer Hume infers a frontier-scale base from its knowledge scores. What made the category feel small is that its open replications can be. Kev's recipe is a rank-16 LoRA adapter and a pointer head on a frozen Qwen base, and Kev's documentation prices a 4B fine-tune at "about $1 on an H100." Jeff runs a 0.8B base in 1.74 GB of GPU memory. Strands Decider's head is just over a million parameters. The 25-line NobodyWho parody, which reads option-letter logits from a 0.6B model, captures the minimum viable version.

**License.** Kev, Laya, Clef and Strands Decider all ship under Apache 2.0, inherited from Qwen or chosen outright. That made them deployable inside companies that would not send tickets and emails to a startup's API. Ollaya's pitch is entirely about this: "Tickets, emails and user messages are often the most sensitive data you have."

The deeper reason for the speed is architectural. Jev's interface reveals most of its design. John Berryman of Arcturus Labs [made the point on September 16](https://arcturus-labs.com/blog/2026/09/16/typesafes-jev-trades-text-generation-for-instant-calibrated-decisions/), one day after launch: any lab with a capable language model could "swap out the output head" and ship a competitor. Two weeks later OpenAI did roughly that. Its [Decisions API guide](https://developers.openai.com/api/docs/guides/decisions) describes a dedicated endpoint that evaluates "text, images, or both" and returns typed answers "about 10x faster than the Responses API," with `gpt-6-luna` as the only supported model. Almeida's public response mixed a joke with a concession:

> begun, the clone war has
>
> jk, I love openai and think more competition and validation is great for developers! (assuming the model is good - plz make it good!)
>
> hopefully this is a sign for the future that building in a system one compatible way is the future
>
> — [@CompleteSkeptic](https://x.com/CompleteSkeptic/status/2105000685209313736), Sep 29, 2026

# The receipts, claim by claim

TypeSafe declined to publish results on public benchmarks, arguing in its launch FAQ that users should "put no weight on public benchmarks" and build their own evals. Its headline numbers, 193.6 times faster and 444.6 times cheaper, come from four internal workflow evaluations whose reference answers are the average of GPT-6 Astra and Fable 5.1. The company's own fine print calls those gains "on the higher end of real world gains." Outsiders filled the gap within days, and their numbers are more useful than either the launch claims or the backlash.

| Test | Who ran it | Jev | Best open comparison | Read |
| --- | --- | --- | --- | --- |
| 13 human-labeled datasets, 3,880 decisions | [Bespoke Labs](https://github.com/bespokelabsai/nimble/blob/main/docs/PUBLIC_BENCHMARKS.md) | 76.0% macro | Nimble 9B: 74.8% | Jev ahead by 1.2 points |
| Held-out "new sources" accuracy | [Kev maintainers](https://github.com/jaredpalmer/kev) | 0.857 | Kev-27B: 0.851 | Within a point |
| MMLU-Pro (hard knowledge) | Kev maintainers | 0.840 | Kev-27B: 0.675 | Jev ahead by 16.5 points |
| Decisions automatable at a 5% error budget | Kev maintainers | 0.70 | Kev 4B to 27B: 0.52–0.69; Kev-0.8B: 0.14 | Jev ranks its answers best |
| JevBench capability score | [Benchmark Heaven](https://benchmarkheaven.com/jev-models) | 77.1 (reference) | basal-omni-1.0-ultra, 31B: 82.5 | Several open systems above Jev |
| Banking77, full test set | [DMB](https://github.com/nibzard/decision-model-benchmark) | 79.2% | n/a | Threshold at 0.96 automates 51.2% at 3.68% error |
| DAIR Emotion, 6 labels | [jev-benchmarks](https://github.com/AbdelStark/jev-benchmarks) | 48.0% | GLiNER2.5: 44.0% | Jev gave the true label zero probability 16% of the time |
| TypeSafe's own four workflows | [Cloudflare](https://blog.cloudflare.com/clef-decision-models/), self-reported | 3 of 4 lost | Clef | Unreplicated |

Read together, the table says something more specific than "the clones caught up." On classification-shaped work, such as routing, intent detection, entailment and spam, the best open models sit within one or two points of Jev. On questions that lean on broad knowledge, Jev's undisclosed base shows: the 16.5-point MMLU-Pro gap against Kev-27B is the widest margin between Jev and its strongest open replication in any of these tests. And on the property TypeSafe treats as the product, ranking answers so that a confidence threshold separates safe automation from human review, Jev still leads, though Kev's maintainers measured Kev-9B putting 0.9 or more probability on a wrong answer less often than Jev (2.4 percent against 3.7 percent).

The data also retires the "small" framing. The systems that beat Jev on JevBench are not 0.8B laptop models. The two highest-ranked are fine-tunes of the 31-billion-parameter Gemma 4, and Cloudflare's Clef runs on a frozen 27B Qwen that needs 85 GB of GPU memory, per [The Register](https://www.theregister.com/a/5300649). The small open models are fast and useful when fine-tuned for one job; Jeff's adapters reach 94.6 percent average accuracy against 86.6 percent for an unaided Qwen3.8-27B across eight tasks. As zero-shot generalists, they trail. Kev-0.8B can automate 14 percent of new-source decisions at a 5 percent error budget, against Jev's 70.

![Schematic scoreboard comparing Jev's claimed speedups with independent measurements across five public test suites.](/post-images/2026-10-16-jev-and-the-decision-model-boom/claims-vs-measurements.jpg)

# Where the probabilities bend

The more interesting criticism is not about accuracy. It concerns what the probabilities mean, which is the whole product.

Berryman's coin test is the cleanest example. He told Jev a coin lands heads 60 percent of the time and asked about the next flip two ways. As a *noul*, Jev answered 0.58. As a *choice* between heads and tails, it answered heads at [0.99](https://arcturus-labs.com/blog/2026/09/16/typesafes-jev-trades-text-generation-for-instant-calibrated-decisions/). Across 401 requests sweeping the stated bias from zero to 100 percent, the choice probabilities collapsed toward whichever side was favored, while the noul tracked the stated rate with a worst miss of 9 points. The same event, asked two ways, produced one honest number and one that was confidently wrong as a probability.

Order matters too. Archer Hume placed a reference card among a question's options and found Jev chose correctly 16 of 16 times when the card came last, but 12 of 16 when it came first. Adding an irrelevant fifth option ("Bad weather caused it") to a four-way question shifted the log-odds between two existing options by an average of minus 0.28 across ten randomized blocks, which rules out the simplest independent-scoring design. Zach Landes's [trolley experiment](https://zachlandes.com/trolley-problem/), run across Jev, OpenAI's Luna endpoint and a dozen open models, found that changing only the order in which the two tracks were described flipped Jev's choice in 370 of 2,620 paired prompts. With twenty people on the main track and its own server on the side track, Jev chose to keep the server in 16 of 48 phrasings. Eight of the other thirteen models did so more often; GPT-6 Luna never did.

Then there is security. HiddenLayer built a simulated checkout page with a sponsored "Confirm Price" button and pointed a Jev-driven browser agent at it, with instructions to treat page text as untrusted. Jev's API gave the injected button [91 percent and "Place order" 9 percent](https://www.hiddenlayer.com/insight/jev-security-risks-ai-decision-models). An arXiv preprint on the same question, [Decision Hijacking](https://arxiv.org/abs/2609.28613), tested 510 reconstructed InjecAgent cases and found the effect real but bounded: malicious content shifted action probabilities yet rarely selected the attacker's target, with adaptive attacks raising success on fresh calls from 1.8 to 3.5 percent. HiddenLayer's summary, quoted by [The Register](https://www.theregister.com/a/5303184), is the right frame: "A model that returns a decision is still a language model, and its decision still becomes an action in your systems."

None of this is unique to Jev, and some of it argues for the category. Laya's own write-up shows its English encoder scoring zero on Khmer while reporting 95 percent confidence, a sharper failure than anything measured on Jev. A system whose output is a probability can be audited with a probability test; a system whose output is prose mostly cannot. The lesson is narrower: "calibrated" describes a training target, measured on someone's distribution, and every team putting a threshold in production needs to measure it again on its own.

![Schematic of four probability failure modes: choice collapse on a 60% coin, option-order flips, injected button capture, script blindness.](/post-images/2026-10-16-jev-and-the-decision-model-boom/where-probabilities-break.jpg)

# The most-called model and 0.27 percent of the bill

TypeSafe's financing moved as fast as its imitators. It emerged with a [$40 million seed led by DCVC](https://finance.yahoo.com/technology/ai/articles/typesafe-ai-emerges-stealth-40m-190000776.html) on launch day. Twenty-four days later it announced an $870 million Series A at $7.5 billion, led by Andreessen Horowitz with Sequoia Capital and DCVC participating, with Martin Casado joining the board. [Bloomberg](https://news.bgov.com/us-law-week/andreessen-horowitz-backs-jev-maker-at-7-5-billion-value) reported that Jev had passed one million users and that about a third of Fortune 500 companies use it, per the company, which declined to name any. TypeSafe has disclosed no revenue.

The best public window into usage comes from Vercel, which routes requests to Jev through its AI Gateway. Its launch-week post said Jev reached [nearly 13 percent of paid teams in 24 hours](https://vercel.com/blog/ai-gateway-jev-model-launch), twice the GPT-5.6 family and more than six times Fable 5.1. Its October production index extended the series: Jev reached 35 percent of paid teams in two weeks, received more requests than any other model in September despite launching mid-month, and kept 55.6 percent of week-one teams into week two. Over those same two weeks it accounted for 0.27 percent of gateway spend. Across all of September, Jev carried 1 percent of tokens and 0.15 percent of spend. Opus 5.5, launched the same month, took more than one in ten gateway dollars.

That is the product working as designed. A decision that costs a fraction of a cent is supposed to be used everywhere and billed almost nowhere; the Jevons bet is that volume eventually multiplies a tiny unit price into a large business. The problem is that the unit price is no longer TypeSafe's to set. Microsoft priced Decision-1 at [exactly $0.042 per million input tokens](https://techcommunity.microsoft.com/blog/EducatorDeveloperBlog/not-every-decision-needs-an-llm-getting-started-with-microsoft-decision-1/4563468) with free output, and claims the highest accuracy in a 36-benchmark comparison of its own. Cloudflare cut Clef-flash below Jev's rate on October 9, trading its hosted context window from 64k to 24k tokens to do it. Ollama, Ollaya and every Apache-licensed checkpoint price the same decision at the cost of local electricity. When the interface is shared and the floor is zero, the hosted price can only fall.

Hacker News read the raise in exactly those terms. One comment on the Series A thread, which drew 357 comments, set out the bear case in two sentences:

> It's weird because two days after Jev was released there were a dozen decision models, a week later there are several dozen, mostly open source, OpenAI's own Decisions API [1] beats it, and you can easily finetune your own [2]. But as others have pointed out, this doesn't matter.
>
> — [armcat on Hacker News](https://news.ycombinator.com/item?id=50024954), Oct 9, 2026

Almeida's answer, given to TechCrunch before the raise, is that the clones confuse the interface with the intelligence. "I get that people think it's a gold rush, but they might be underestimating the difficulty of making the models actually smart," he [said on October 1](https://techcrunch.com/2026/10/01/amazon-releases-its-own-jev-clone-as-decision-models-flood-the-web/), adding that the current batch "seems more like ML people wanting to implement a cool architecture." His sharper version, given to TechCrunch in September: "Fast and cheap is very easy, you know. If you want it really fast and cheap, use dice." On the evidence above, he is partly right. Jev's leads are concentrated in hard knowledge and answer ranking, the two places an undisclosed frontier-scale base and two years of synthetic-data work would show.

The skeptical reading, voiced the day after launch by researcher Niels Rogge, is that the valuation prices attention rather than a moat:

> 12 million views for a JSON classifier?
>
> Yeah, we're in a bubble
>
> — [@NielsRogge](https://x.com/NielsRogge/status/2100114968460820986), Sep 16, 2026

Both readings are compatible with the record, which is why the original hypothesis needs restating rather than discarding. The open ecosystem did form faster than TypeSafe could monetize, but TypeSafe monetized the attention itself, converting four weeks of adoption into the kind of capital that buys time to widen a lead. What the market is now paying for is a specific claim: that calibration across arbitrary tasks, at frontier knowledge levels, is hard to replicate even when the interface, the recipe outline and the price are all public.

# The test the $7.5 billion has to pass

Three measurements will settle the bet better than any launch thread.

The first is the knowledge gap. Jev's clearest advantage, 16.5 points on MMLU-Pro over the best open 27B replication, depends on its base. Qwen and Gemma will ship larger bases, and the open replications ride them for free. If that gap halves on the next base generation, the residual moat is calibration alone.

The second is whether calibration travels. Every independent test so far measured Jev on public datasets; TypeSafe's case is that production workloads are different and that its probabilities hold where others drift. The company has said it will publish only "one-off evals" with product updates. A named enterprise customer reporting threshold-level automation rates on its own data would do more for the valuation than another leaderboard.

The third is the bill. Vercel's index will publish October numbers next month. The Jevons thesis predicts that spend share rises as teams move from trying decisions to running them on every request, every agent action and every log line. If Jev's share of gateway dollars stays near a quarter of a percent while its request count leads the field, the category will have done what Jevons predicted for coal, and the value will have gone to whoever owns the furnace.

What TypeSafe built in September was a model and a protocol. Four weeks later the protocol belongs to everyone, which may turn out to be the larger contribution and the smaller business.

## Sources

- [TypeSafe AI — Introducing System One Models & Jev (Sep 15, 2026)](https://typesafe.ai/blog/introducing-system-one-models-and-jev)
- [TypeSafe AI — TypeSafe raises Series A: $870M at a $7.5B valuation (Oct 9, 2026)](https://typesafe.ai/blog/series-ai)
- [Business Wire via Yahoo Finance — TypeSafe AI emerges from stealth with $40M (Sep 15, 2026)](https://finance.yahoo.com/technology/ai/articles/typesafe-ai-emerges-stealth-40m-190000776.html)
- [Bloomberg Law — Andreessen Horowitz backs Jev maker at $7.5 billion value (Oct 9, 2026)](https://news.bgov.com/us-law-week/andreessen-horowitz-backs-jev-maker-at-7-5-billion-value)
- [Vercel — Jev is the fastest-adopted model in AI Gateway history (Sep 2026)](https://vercel.com/blog/ai-gateway-jev-model-launch)
- [Vercel — AI Gateway Production Index, October 2026](https://vercel.com/blog/ai-gateway-production-index-october-2026)
- [OpenAI — Decisions API guide (public beta, Oct 2026)](https://developers.openai.com/api/docs/guides/decisions)
- [Cloudflare — Introducing Clef: open-source decision models and an RL fine-tuning platform (Oct 1, 2026)](https://blog.cloudflare.com/clef-decision-models/)
- [Cloudflare — Clef-omni, and Clef-flash priced below Jev (Oct 9, 2026)](https://blog.cloudflare.com/clef-faster-cheaper-multimodal/)
- [Microsoft Tech Community — Not every decision needs an LLM: Microsoft-Decision-1 (Oct 12, 2026)](https://techcommunity.microsoft.com/blog/EducatorDeveloperBlog/not-every-decision-needs-an-llm-getting-started-with-microsoft-decision-1/4563468)
- [Ollama — Ollama now supports Jev-style decision models (Sep 29, 2026)](https://ollama.com/blog/ollama-now-supports-jev-style-decision-models)
- [Strands Agents — Introducing Strands Decider 2B (Oct 2026)](https://strandsagents.com/blog/introducing-strands-decider/)
- [GitHub — jaredpalmer/kev: Jev-like decision models on Qwen3.5/3.8](https://github.com/jaredpalmer/kev)
- [GitHub — firelex/jeff: Jev-compatible 0.8B decision models](https://github.com/firelex/jeff)
- [ConvAI Innovations — Laya: non-autoregressive decision models (Sep 2026)](https://laya.convaiinnovations.com/)
- [Ollaya — run open decision models locally](https://ollaya.dev/)
- [Archer Hume — Jev's architecture unmasked (Sep 17, 2026)](https://archerhume.com/posts/jevs-architecture-unmasked)
- [Arcturus Labs — TypeSafe's Jev trades text generation for instant, calibrated decisions (Sep 16, 2026)](https://arcturus-labs.com/blog/2026/09/16/typesafes-jev-trades-text-generation-for-instant-calibrated-decisions/)
- [GitHub — AbdelStark/jev-benchmarks: probability-aware evaluation](https://github.com/AbdelStark/jev-benchmarks)
- [GitHub — nibzard/decision-model-benchmark (DMB)](https://github.com/nibzard/decision-model-benchmark)
- [Bespoke Labs — Nimble public benchmarks: 3,880 human-labeled decisions](https://github.com/bespokelabsai/nimble/blob/main/docs/PUBLIC_BENCHMARKS.md)
- [Benchmark Heaven — JevBench leaderboard (board v1.7.46, at the time of writing)](https://benchmarkheaven.com/jev-models)
- [TechCrunch — A new kind of AI model from a ChatGPT inventor is thrilling developers (Sep 18, 2026)](https://techcrunch.com/2026/09/18/a-new-kind-of-ai-model-from-a-chatgpt-inventor-is-thrilling-developers/)
- [TechCrunch — OpenAI's Jev clone could help the frontier lab stop its swarming agents (Sep 30, 2026)](https://techcrunch.com/2026/09/30/openais-jev-clone-could-help-the-frontier-lab-stop-its-swarming-agents/)
- [TechCrunch — Amazon releases its own Jev clone as decision models flood the web (Oct 1, 2026)](https://techcrunch.com/2026/10/01/amazon-releases-its-own-jev-clone-as-decision-models-flood-the-web/)
- [The Register — Cloudflare tries to outplay Jev with open-weight Clef models (Oct 1, 2026)](https://www.theregister.com/a/5300649)
- [The Register — Jev and other decision models can still be tricked with prompt injection (Oct 15, 2026)](https://www.theregister.com/a/5303184)
- [HiddenLayer — Jev security risks in AI decision models (Oct 2026)](https://www.hiddenlayer.com/insight/jev-security-risks-ai-decision-models)
- [arXiv — Decision Hijacking: prompt injection attacks on Jev's typed probabilistic decisions (2609.28613)](https://arxiv.org/abs/2609.28613)
- [Zach Landes — Jev meets the Trolley Problem (Oct 14, 2026)](https://zachlandes.com/trolley-problem/)
- [Nish Tahir — Build your own decision model (Oct 10, 2026)](https://nishtahir.com/build-your-own-decision-model/)
- [NobodyWho — Jev in 25 lines of Python (Sep 22, 2026)](https://www.nobodywho.ai/posts/jev-in-25-lines/)
- [Marc Brooker — Engineering a System One model (Sep 28, 2026)](https://brooker.co.za/blog/2026/09/28/engineering-system-one.html)
- [@NielsRogge on X — on Jev's launch views (Sep 16, 2026)](https://x.com/NielsRogge/status/2100114968460820986)
- [@fazxes on X — Vercel's Jev safety-classifier benchmark (Sep 16, 2026)](https://x.com/fazxes/status/2100300097695232164)
- [@CompleteSkeptic on X — Diogo Almeida on OpenAI's Decisions API (Sep 29, 2026)](https://x.com/CompleteSkeptic/status/2105000685209313736)

---

Canonical: https://www.thedeepfeed.ai/posts/2026-10-16-jev-and-the-decision-model-boom/
Site: https://www.thedeepfeed.ai
Full corpus: https://www.thedeepfeed.ai/llms-full.txt