TL;DR
On 19 September a startup released a model that will not write you a sentence. You ask a typed question, it returns a value and a probability, and its own benchmark claims up to 194 times faster and 445 times cheaper than a frontier model. The speed is the boring part. The interesting part is that the answer arrives with a number saying how much to trust it, which is the first thing in three years of enterprise AI you can actually test on your own data.
On 19 September, TypeSafe AI released a model that will not write you a paragraph. Ask it something and it answers with a value and a probability. Nothing else. No prose, no explanation, no apology.
The company calls it a decision model. Its founder, Diogo Almeida, worked on ChatGPT's training at OpenAI, so this is not someone who missed the language era. He built its opposite on purpose.
What actually shipped
Three question types. Choice picks one option from a set and returns the distribution across the rest. Score places an input on an ordered scale. Noul answers a yes or no proposition with a probability between zero and one. You give the model a state, you ask several questions about that state in parallel, and your own code decides what happens next.
TypeSafe's benchmark reports 0.114 seconds and $0.000081 per call, against 8.566 seconds and $0.013880 for a frontier model on the same workflow. Roughly eight cents for a thousand decisions. Input costs $0.042 per million tokens. Output is free, because there is no output to speak of.
A frontier model hands you a confident paragraph. A calibrated model hands you a number and tells you how much to trust it.
The interesting number is not 194x
Speed and price are the headline, and they are the least surprising part of the story. A model that emits one token does less work than a model that emits four hundred. That is arithmetic, not innovation.
The part worth your attention is the probability. TypeSafe trained the model with something it calls Reinforcement Learning for Calibrated Decisions. Calibration means that when the model says 70%, it is right about 70 percent of the time. Not confident. Correct about its own confidence.
Text output never gave anyone that. A wrong paragraph and a right paragraph look identical: both fluent, both well formatted, both fine in a demo. A calibrated number behaves differently. You can measure it per segment, set a threshold, and route on it. Above 0.9 it goes straight through. Below 0.6 a human looks. In between you log everything and review the cost of being wrong once a month.
Meteorology worked this out in 1950, when Glenn Brier published a score for grading probabilistic forecasts. Weather forecasters have been marked on their humility for seventy six years. Software has not.
What you cannot buy yet
Now the cold water. The benchmark is vendor run, on workflows the vendor picked. There are no published weights and no self hosting, only a hosted API behind a waitlist. The context window stops at 64,000 tokens. The model is weak at arithmetic, counting, date comparisons and anything adversarial, it cannot explain its own answer, and TypeSafe says plainly that it is not a security boundary.
So calibration is a claim until you measure it on your own labelled data. Which is, pleasingly, exactly what the model is asking you to do.
Why this matters for your business
Most of what companies bought as AI in the last three years was language wrapped around a judgement call. The judgement was the valuable part. The language was the invoice.
If a category of cheap, typed, calibrated decisions holds up, the question you put to a vendor changes shape. Not "what can your model do", but: show me the calibration curve on my data, segment by segment, with the threshold you recommend and the review load it implies.
Ask it this quarter. The answer sorts the room faster than any demo.
Pan