I am blown away by the release of a new class of AI yesterday. If you're interested in AI, read this

I think a small San Francisco lab has just published the most important rethink of how software uses AI since the first LLM APIs shipped. TypeSafe AI was founded in 2024 by people with OpenAI and FAIR backgrounds. Yesterday it released documentation for a model called Jev, the first of what it calls System One models. It does not chat, it does not write, and it does not explain itself. It makes decisions, fast and cheaply, with probabilities attached. I think this is truly revolutionary, and I don’t use the word much.

Anyone who has created a serious application on top of a large language model has hit the same wall. The model is astonishingly capable and astonishingly awkward to build with. You send it a prompt, it sends back prose, and your application has to parse, validate, retry, and hope. Every call costs tokens in both directions, most of the output is explanation you did not ask for, and the same input can come back differently tomorrow. It’s inconsistent, and the clever models are very expensive to use.

I have been building Aria, a virtual actuary that works as a member of an actuarial team rather than as a chatbot, and it is currently with a dozen or so beta testers. Almost every hard engineering problem I have hit is a version of that wall, and until I read TypeSafe’s documentation I assumed the wall was simply the price of using AI at all.

What a System One model does

This gets a little “techie”, but please give it a go.

A System One model takes the same natural-language input an LLM takes. It does not write anything back. You define the questions in advance, each with a fixed answer space, and the model returns a typed answer, a probability distribution over the options, and a confidence figure. There are three question types: a Choice (which of these options), a Score (where on this ordered scale), and a yes/no question TypeSafe calls a Noul, which returns the probability that the answer is yes. The name is borrowed from Professor Kahneman: fast, intuitive judgment rather than slow deliberation.

This has three consequences, and they fundamentally change the economics of building with AI.

First, it is fast. TypeSafe says most queries complete in about 100 milliseconds, which means it can sit alongside a person while they work rather than run overnight, and it can respond in real time to changing processes or unexpected events.

Second, it is cheap in a structural way, not a discounted way. The AI you’re used to charges for input tokens, and usually about 5x that for output tokens. The output tokens dominate the cost, because a reasoning model generates thousands of them before it answers. A System One model generates essentially nothing. You can also ask dozens of independent questions about the same state in one call, evaluated in parallel. TypeSafe’s own cookbook runs a 13-question briefing over a single document and reports the batched version at 11.5x cheaper and 9.6x faster than asking one question at a time. The company’s stated target is more than a 100x improvement in intelligence per unit of speed and cost over conventional models, and the pricing it has posted on X suggests it means it: $0.042 per million input tokens, and no output charge at all, because there are no output tokens to charge for. To calibrate, Claude Fable, the frontier model I use for Aria’s hardest reasoning, is $10 per million input tokens and $50 per million output. That is not the fair comparison, though, because nobody would run a classifier on a frontier model. The fair comparison is OpenAI’s GPT-5.6 Luna, the cheapest model in its current line-up after a much-discussed 80% price cut in July, at $0.20 input and $1.20 output. Jev undercuts it by roughly 5x on input, and Luna still has to write its answer out before your code can use it. Even at a fraction of the claimed gains, this is a different category.

Third, the probabilities are calibrated. The model is trained so that the answers it gives at 80% are right about 80% of the time across a population of decisions. Calibration is a group property, and TypeSafe are careful to say so: it does not make any individual answer correct. It does mean your code can route on confidence, act when the model is sure, and escalate to a person or a reasoning model when it is not. This is exactly what decision support models, where humans must retain control, should do.

Sounds subtle, it’s actually very important

The big idea in TypeSafe’s design guide is that you build a normal software workflow and insert AI only where judgment is needed. Code owns the control flow. Deterministic rules stay in code. The model handles the narrow common-sense calls over unstructured data, one atomic question at a time, and the code composes the answers.

That is the opposite of how most “AI-native” products are built today, where an agent reads the instructions, decides its own next step, and loops. The agent architecture works well when a person is watching. Every additional loop is another chance to wander off. Huge amounts of software are then written to try and keep the LLM on course.

Software built around decisions is a different discipline from software built around text.

The reason this excites me is that it is how people in real processes actually work. A claims handler does not compose an essay about a claim; she decides whether it is covered, whether the reserve is adequate, and whether it needs a senior. An underwriter looking at a submission is making a small number of bounded calls: accept, decline, refer, and on what terms. Very little of professional judgment is free-form generation. Most of it is choice under uncertainty from a menu the institution has already defined. The LLM era forced us to route those decisions through a machine that wants to write paragraphs. System One routes them through a machine that wants to choose.

What this means for Aria (my virtual actuary)

An actuary reviewing a reserving run makes a sequence of judgments that look nothing like a chat. Has the development pattern in this triangle changed? Is this large loss an outlier or a signal? Does this line justify a Bornhuetter-Ferguson approach, or is the chain ladder fine? Should the a priori loss ratio move? Does the movement warrant a note to the CFO?

Today Aria answers those questions by reasoning with an LLM, which means long prompts, long answers, and a real cost per decision. Decomposed into typed questions, each becomes a Choice or a Noul with a probability attached, evaluated in a single call, at a cost and latency that would let Aria assess every line and every accident year on every run rather than only the ones a human has time to flag. The reasoning model is not gone. It is reserved for the cases where confidence is low or the decision is unusual, which is exactly where a human actuary would slow down and think too.

That is faster and cheaper, and aligned to what is needed. Actuaries do not trust a colleague who narrates. They trust a colleague whose calls are consistent, whose confidence tracks reality, and who says “I am not sure, look at this one” at the right moments. A calibrated probability is a much better fit for that relationship than a well-written paragraph. Underwriting, claims triage, and compliance review have the same shape.

The caveats

Jev is a new model from a lab that was in stealth until recently, and I have no independent benchmarks beyond the company’s own cookbooks. Calibration on customer-support tickets does not guarantee calibration on reserving judgments; that has to be tested on labelled data. The model gives no explanation of its answer, which regulators in my industry will notice. My view is that decomposition supplies the explanation: a decision assembled from ten named, inspectable probabilities is more auditable than a narrative generated after the fact. That argument still needs to be made to regulators, and I think it is a stronger one than most explainability claims being made today.

There is also the dependency question I raised earlier in this series. If you rebuild your decision layer around one lab’s model, you have concentrated a new risk. I would want the question format to become a standard that more than one provider can serve before betting a regulated process on it.

Where I am betting

I could be wrong about TypeSafe as a company. Model labs are fragile, and a better-funded competitor could ship the same primitive next quarter. I do not think I am wrong about the primitive. Typed, calibrated, parallel decisions at 100 milliseconds are a more natural unit of enterprise AI than generated prose, and I expect that within two or three years most production AI in insurance will be built this way, with reasoning models pulled in only at the edges. If that is right, the cost of intelligence inside a decision process falls by orders of magnitude, and the interesting question stops being “can we afford to run AI on this” and becomes “which decisions are we still making by hand, and why?”

Tell me where this is wrong. I would rather hear it now than after I have rebuilt Aria around it!