TypeSafe's Jev Sells Decisions, Not Words: What a Decision-Only AI Model Changes

Last Updated: 16 September 2026

A new kind of AI model launched this week that never writes a word. Jev, the first release from a startup called TypeSafe AI, reads a piece of text and returns structured decisions (a choice from a list, a score, or a yes/no probability) in under 100 milliseconds, with output tokens billed at zero. If the published pricing holds, the cheapest part of automating a business workflow stops being the writing and starts being the deciding, and that inverts which automations are worth doing first.

What Jev actually is

Jev is not a chatbot with a formatting constraint. According to TypeSafe, whose launch materials describe a model trained with a method the company calls Reinforcement Learning for Calibrated Decisions, every call takes some state (an email, a support ticket, a log line) plus a list of typed questions, and returns answers drawn from a fixed set: pick one option from a list you supplied, place the state on a scale you defined, or return the probability that a yes/no claim is true. Each answer carries a confidence score. The model cannot return anything outside the structure you gave it, because text generation is not part of what it does.

[Answer: Jev is a decision-only AI model. It reads one piece of state, answers only the typed questions you gave it, and returns a choice, a score or a probability with a confidence value attached. It cannot write prose, code or conversation. That narrowness is the product: nothing to parse, nothing to format-check, and a number that tells your software when to act and when to defer to a person.]

The people behind it

The company behind it is credentialed in exactly the area that matters here. According to the launch announcement carried by Businesswire, TypeSafe emerged from stealth on 15 September with $40 million in seed funding led by DCVC, founded in 2024 by Diogo Almeida, a former OpenAI researcher credited as a co-inventor of RLHF, the training method behind ChatGPT. Forbes, covering the launch the same day, reported Almeida's argument that AI models have been optimised to please humans while staying silent about their own uncertainty, and that Jev is built to attach a probability to every answer so software can tell a 98% call from a 60% one.

That confidence score is the part worth more than the speed. A chat model hands you a paragraph and lets you guess how much of it to trust. A model that returns "billing, 0.94" or "complaint escalation, 0.51" lets a program act on the first and route the second to a person. The governance rule writes itself: software acts when confidence clears a threshold you set, and a human sees everything below it.

Why the output price is zero

[Answer: Jev's output is free because it generates no output tokens. A chat model pays for every word it writes; Jev scores every answer in one parallel pass, so there is nothing to meter. The published price is $0.042 per million input tokens. The company's own claim of up to 100 times faster and cheaper is not independently verified, and TypeSafe itself says it cannot yet prove the price is not subsidised.]

Jev's output is free because it generates no output tokens at all. A chat model pays for every word it writes, one token at a time. Jev scores every answer in the schema in a single parallel pass, so there is nothing to meter. TypeSafe's published price is $0.042 per million input tokens, and the same launch materials claim the model runs up to 100 times faster and less expensive than comparable frontier models, at under 100 milliseconds of latency. The price is a published number you can check on the company's price sheet. The performance multiples are the company's own claim, and no independent replication exists yet.

To make the pricing concrete, consider an illustrative case (the arithmetic is ours, from the published rates, not a benchmark): a service business routing 40,000 items a month, each around 800 tokens of input, consumes about 32 million input tokens monthly. At $0.042 per million, that is roughly $1.34 for the month's decision layer. The same volume on a frontier chat model at a few dollars per million input tokens, plus generated output on every call, lands in a different category of spending entirely. Whether the price is sustainable is another question: TypeSafe itself has said publicly that it cannot yet prove the pricing is not subsidised. A launch price this low is a strategy statement as much as an economics statement, and the model is named after the economist who observed that when a resource gets radically cheaper, consumption explodes.

What the launch does and does not prove

[Answer: the launch proves a product exists at a published price, not that its accuracy claims hold. The pricing is public and checkable. The accuracy comparisons are the company's own evaluations, built by its own team against reference answers averaged from rival models, and its own dashboard shows Jev losing the accuracy column. No independent verification accompanied the release.]

The honest reading splits the claims into three piles. The first pile is checkable today: the pricing, the request format, the latency band the company publishes. The second pile is the company's own homework. TypeSafe's launch materials acknowledge that its accuracy evaluations were built by its own team, with reference answers derived from averaging the judgments of two rival frontier models, and flag that as a possible bias. On the company's own dashboard, Jev wins the cost and latency columns and loses the accuracy column. The third pile does not exist yet: independent verification. No architecture paper, no public benchmark entries, no third-party accuracy audit accompanied the launch.

Zero hallucinations, narrowly

"Zero hallucinations" belongs in the second pile, and it is the claim most likely to be repeated badly. It is true in a narrow, structural sense: the model cannot invent an answer outside the list you gave it. It can still confidently pick the wrong option from that list, and a ticket tagged "billing" that belongs with tech support passes the format check either way. Structure is guaranteed; correctness is a calibration claim that only testing against your own data can settle.

There is a second-100-days point hiding in this. Any founder who has bought a business knows the pattern: a tool was bought on the vendor's evaluation, the workflow was wired around it, and nobody ever measured whether it was right. The artefact that prevents this is not sophisticated. It is a small labelled sample of your own real cases, agreed before deployment, against which any model's accuracy is measured, with the threshold and the deferral rule written next to it.

What an owner can usefully take from it

[Answer: the useful move is an inventory, not a purchase. List the judgment points in your workflows, decide for each whether software may act alone or must defer, and write down the confidence rule and the human path. That artefact costs nothing, works with today's models, and turns a launch like this into a substitution you can evaluate on your own data rather than the vendor's.]

The practical takeaway is not "buy this model." It is early access, waitlisted, speaking its own request format, and unproven on accuracy. The takeaway is about the shape of the work.

In discovery workshops, owners rarely arrive with a list of what to automate. The list gets built in the conversation, walking through a normal week task by task, and the candidates that separate themselves are rarely prose work. They are decision points: which queue does this go to, is this invoice ready to pay, does this message need the owner or can it wait, is this lead worth a call. Those judgments are often cheap for a person and expensive per call for a chat model, so they stay manual long after the writing around them has been automated. A decision layer priced like this, if the calibration is real, is precisely what those waiting tasks have been waiting for.

The discipline that makes any of this safe is the same one the first 100 days should already be running. Name the judgment points in a workflow, decide for each one whether software may act alone or must defer, write the confidence rule down, and keep a human path for everything the model is unsure about. That inventory costs nothing to produce, works with today's chat models as well as with whatever comes next, and turns a vendor launch from a gamble into a substitution decision you can actually evaluate.

Sources

• The Register, "TypeSafe AI debuts model for machines that plays Doom", Thomas Claburn, 16 September 2026. https://www.theregister.com/ai-and-ml/2026/09/16/typesafe-ai-debuts-model-for-machines-that-plays-doom/5296711
• Businesswire, "TypeSafe AI Emerges From Stealth With $40M in Funding With New Model for Composable AI", 15 September 2026. https://www.businesswire.com/news/home/20260915525333/en/TypeSafe-AI-Emerges-From-Stealth-With-%2440M-in-Funding-With-New-Model-for-Composable-AI
• Forbes, "This $200 Million Startup Wants To Fix AI's Overconfidence Problem", 15 September 2026. https://www.forbes.com/sites/the-prompt/2026/09/15/this-200-million-startup-wants-to-fix-ais-overconfidence-problem/
• TypeSafe AI launch announcement (subject company's own statement about itself): https://typesafe.ai/blog/introducing-system-one-models-and-jev

Pricing performance claims in this piece are TypeSafe's own, attributed as such; no independent accuracy verification existed at publication. The worked pricing example is illustrative arithmetic from the published rates, not a benchmark.