Key takeaways
- ✓Jev does not return text, but a decision in a fixed format, with a confidence score
- ✓On the first 240 test cases Jev gave the most correct answers: 91.7% against 88.3% for the Decisions API
- ✓Jev was the fastest API, and the Decisions API cost about 23% more on average
- ✓Each system has its strong side, so we choose per task
A model for software, not for people
Today's language models are designed for people. That has two drawbacks. As people, we more readily believe a model that hallucinates with confidence. And as people, we are slow to get to the point, which leads to far too long answers. Together, that makes AI in software less reliable than it should be, slower than needed and expensive in tokens.
Jev, the model TypeSafe AI presented in September 2026, takes a different approach. You send a focused question along with the allowed answers, and you get one back with a confidence score. According to TypeSafe, that makes it 200 times faster and 444 times cheaper than GPT 5.6 Terra. With a score that is right, the software can decide for itself when to continue and when to hand over to a person.
Why this matters for our clients
Correct classification is essential to build good context in a company. And because the volumes are high, cost matters too. We immediately tested Jev on the tasks that come back most often at our clients: classifying emails automatically, linking them to the right file and extracting data from documents. One of our engineers first compared it with four other models.
How we set up the test
When OpenAI launched the Decisions API shortly after, we extended the test. We compared four setups: Jev from TypeSafe AI, the Decisions API, and GLiNER2.5 Decide from Fastino Labs, both through the API and locally on a Mac.
We used 240 test cases and 120 extra Dutch examples, with the correct answers fixed in advance. We measured three things: is the answer right, how fast does it come back, and what does the request cost?
Per task
Classifying short commands, such as recognising that someone wants to play music or set an alarm: on the 40 English commands, Jev and GLiNER locally both scored 40 out of 40. On the Dutch commands, Jev and the Decisions API tied at 39 out of 40.
Applying rules and conditions to a described situation: Jev scored best, with 45 out of 60 against 38 out of 60 for the Decisions API. Understanding a Dutch text and picking the right answer out of four options: Jev scored 57 out of 60, the Decisions API 53 out of 60.
Comparing two Dutch sentences, deciding whether the second follows from the first, contradicts it or lacks information: here the Decisions API won, with 53 out of 60 against 51 out of 60 for Jev. Recognising spam in text messages: the Decisions API scored 19 out of 20, Jev and GLiNER locally both 18 out of 20.
Across the whole test
On the first 240 cases Jev gave the most correct answers: 91.7%, against 88.3% for the Decisions API. On the extra Dutch set they were closer: 108 out of 120 against 106 out of 120.
Jev was also the fastest API, at 0.41 seconds on average against 0.52 seconds. The Decisions API cost 1.23 times as much as Jev on average, about 23% more: per 1,000 similar requests around $0.0230 against $0.0187. The GLiNER API was the cheapest API. GLiNER locally was the fastest, at 0.17 seconds and with no API cost per request, but scored 71.3% correct on this test mix.
What we choose
Our choice remains Jev for applying rules and understanding texts. The Decisions API won at comparing sentences and recognising spam. GLiNER locally was strongest on short English commands. Systems like these are fast and cheaper, and a confidence score that is right lets you tune the decisions in your processes to it.
Which model fits your process?
Tell us how the work runs today.
