Jev Cannot Hallucinate. It Can Still Be Wrong.

What TypeSafe Shipped

TypeSafe AI emerged from stealth on September 15, 2026 with $40 million in seed funding led by DCVC. The company was founded in 2024 in San Francisco by Diogo Almeida, Erik Gafni, and Sasha Sheng. Almeida was a researcher at OpenAI who worked on reinforcement learning from human feedback, InstructGPT, ChatGPT, and GPT-4. Forbes reported a $200 million valuation, citing a person familiar with the transaction.

The product is Jev, which TypeSafe calls a System One Model. The name comes from Daniel Kahneman's split between fast intuitive judgment and slow deliberate reasoning. Jev handles the first kind.

Jev does not generate text. It takes program state plus a set of typed questions and returns typed answers with probability distributions, all in one parallel pass. The training method is Reinforcement Learning for Calibrated Decisions, which TypeSafe positions against RLHF and RLVR: instead of optimizing for human preference or verifiable rewards, it optimizes for probabilities that are honest about their own uncertainty.

The pitch is narrow on purpose. Almeida told TechCrunch that models today are optimized for human language rather than automation, and rejected the frontier lab label: "the main product of frontier labs is fear or hype." The company's own press release opens by calling TypeSafe "a frontier AI lab."

The API Surface

Jev exposes three primitives, confirmed in both TypeSafe's documentation and Cloudflare's model listing:

Primitive	Returns	Notes
Noul	Probability a yes/no statement is true, 0 to 1	Boolean judgment
Choice	One option plus a probability distribution and confidence	Up to 255 options
Score	An ordinal level plus distribution and confidence	Can land between named levels

Current model ID is jev-1.13.0. Published limits are 250,000 tokens per second and 1,200 requests per minute, marked subject to change. The context budget is 64K tokens per request, with 32K allocated to the state plus the longest individual question. Cloudflare's listing shows 32,000, which matches the state-plus-question budget rather than the request total. Jev accepts text only - strings, JSON objects, or text arrays - and does not accept images, audio, or video.

Pricing is $0.042 per million input tokens, or $42 per billion. Output tokens are free. For comparison, Claude Sonnet 4.6 charges $3 per million input tokens and $15 per million output tokens, as covered in the Sonnet 4.6 release. TypeSafe's homepage claims a 238x lower input price than Claude Fable 5.1.

TypeSafe derives confidence from the concentration of the probability distribution; the documentation's interactive demo approximates it for three options as (3 x largest probability - 1) / 2. The documentation is direct that correct thresholds depend on the domain and advises starting conservative and testing against real data.

The Benchmark Problem

TypeSafe's headline numbers are 193.6x faster and 444.6x cheaper, against a claimed end-to-end response time of 70ms to 500ms. Both multipliers come from an evaluation format the company designed itself. The figures are not internally consistent: the funding press release claims up to 100 times faster and cheaper, roughly half the homepage number.

The workflow evals cover four tasks: Security Incidents, Agent Trace Observability, Invoice Processing, and Customer Service. Every model gets the same workflow. The scoring target is not ground truth. It is the average of GPT-6 Astra and Claude Fable 5.1, "both at high thinking, answering every question in the harness."

That design choice matters. The reported figures measure agreement with two specific competitor models, not correctness. A model could agree with the reference on every case and still be wrong on every case. TypeSafe is upfront about the tradeoff: rather than debating the correctness of the harness and labels, it assumes the code is correct and measures against the current smartest large models.

There is a second asymmetry. The reference models run at high thinking, but "all other models are evaluated using the provider's default reasoning settings." The speed and cost multipliers are therefore measured against LLMs with reasoning off, while the accuracy target is set by LLMs with reasoning turned up. TypeSafe makes the same point about one of its demos: "The LLMs look much worse at this task than with reasoning enabled."

TypeSafe flags at least four limitations in its announcement. The workflows were built by its own model capabilities team. The reference models bias results toward OpenAI and Anthropic. The competing LLMs "use our System One LLM wrapper, which constrains LLMs to output structured decisions" - the company calls this "the most accurate way to get decisions from LLMs," but it also means the comparison runs on TypeSafe's harness. And on the headline multipliers themselves, the company writes that "these are on the higher end of real world gains."

TypeSafe has published no results on standard public benchmarks, no architecture paper, and no calibration curves or expected calibration error figures. That last gap matters most, because calibration is the central selling point. No major lab or publication has measured it independently.

Early Adoption

The adoption signal is real and comes from a third party. Vercel reported that Jev became the fastest-adopted model in AI Gateway history. It reached a tenth of paid teams within 18 hours and nearly 13% by hour 24 - twice the first-day share of the GPT-5.6 family and more than six times that of Fable 5.1. Vercel's own framing was measured: the next test is whether that early adoption lasts.

Developer reports collected by TechCrunch are mixed. Vercel's Pranit Sharma said replacing OpenAI's Luna with Jev for safety classification returned results 5 to 18 times more quickly and with greater accuracy. Nikhil Mudholkar of Bryo AI tested email classification and found Gemini "slightly more accurate, but 10 to 20 times more expensive," while Jev was the only model that returned a real probability.

Armin Ronacher, CTO of Earendil, put the sharpest version of the catch. Jev "delegates the hallucination problem a little bit to the user," he said: the developer has to decide that a 50% result is a coin toss to discard while a 95% result is actionable. The model reports its uncertainty rather than resolving it, and choosing the threshold is unglamorous domain work that the benchmark numbers do not capture.

Price per token also does not settle price per decision. Cost depends on how much state each call carries, and a workflow that ships a large state blob to answer one small question can cost more than a cheap LLM answering from a short prompt.

What the Claims Actually Guarantee

Four claims carry the launch: type safety, zero hallucinations, calibrated confidence, and the 193.6x/444.6x multipliers. The first is verifiable and holds. The rest are weaker than they read.

Jev cannot emit a type error because the output space is fixed in advance. It is also not novel. OpenAI's Structured Outputs has guaranteed 100% JSON Schema conformance through constrained decoding since August 2024. Type safety is table stakes, and TypeSafe says as much.

"Cannot hallucinate" is the weaker claim. The Register's Thomas Claburn noted that the absence of hallucination does not preclude being incorrect. Constraining output to a fixed option set limits what can be fabricated; it does not make the chosen option right. The underlying technique is also long-established. Scoring a model's probabilities across a predefined option set rather than letting it generate free text is how MMLU has been evaluated since 2020, and fine-tuned encoder classifiers have done comparable work for years. What Jev adds is doing it zero-shot across domains, where a fine-tuned classifier covers one.

TypeSafe's own documentation is more candid than its marketing. The model jaggedness notes state that Jev does not count reliably, underperforms on numeric formats such as hex or RGB values, cannot reliably judge whether two values are near each other, and reads dates as text rather than ordered quantities. Accuracy falls as the state grows with content unrelated to the decision, and questions requiring multiple hops or double negatives degrade. Separately, the model page notes that English is the primary training language and that other languages, including CJK scripts, are "handled but not equally well."

One documented weakness cuts directly at the calibration pitch: there is no guaranteed mathematical relationship between semantically related outputs. The probability of a statement and the probability of its negation need not sum to one.

Almeida told TechCrunch the company generates all of its own training data, calling it "one of the best bets I've ever made in my life." Direct API access remains waitlisted, though Jev is already reachable without it through Vercel AI Gateway and Cloudflare Workers AI - so the absence of independent evaluation reflects how new the model is, not a lack of access.

What TypeSafe Shipped

The API Surface

The Benchmark Problem

Early Adoption

What the Claims Actually Guarantee

Sources

Enjoyed this article?

More from the blog

Let's Talk