The jev-latest alias now points to Jev 1.14: what to check in your integration

The jev-latest alias, which TypeSafe AI lists as the stable option, no longer points to jev-1.13.0 but to jev-1.14.0. The models page now describes it as the most recent stable release, which is also the default in the client SDKs. S01 As recently as mid-October, version 1.14 was available only as a preview, as described in our article on Jev 1.14. Anyone calling Jev through the alias or leaving the model choice to the SDK now receives answers from the new version without any change to their own code. S01

What exactly changed

The models page lists two current versions: jev-1.14.0, available from 14 October 2026, and the older jev-1.13.0. S01 Both cost 0.042 USD per million input tokens, output tokens are free, and both accept only text as input, not audio or video. S01 The jev-preview alias currently points to the same model as jev-latest, because no preview build is available. It will move ahead of jev-latest again when TypeSafe ships the next preview. S01

The documentation does not say exactly when the alias moved. Neither the October entries in the Python SDK changelog (versions 0.7.3 to 0.7.5) nor the JavaScript SDK 0.7.0 entry mention a change of the default model. S63S64 The JavaScript SDK even still states that Box questions are meant for the jev-preview model. S64 The models page still does not list the differences between versions 1.13 and 1.14. S01

Why it matters: thresholds and language

The documentation itself warns that an alias can change answers without any code change, and recommends specifying a concrete version in the request, especially if you have tuned confidence thresholds against it. S01 This is a common situation: an integration is set up so that above a certain confidence a customer message is classified automatically, and below it the message goes to a human. Such a threshold was derived from the probability distribution of the old model.

At the same time, the confidence page does not claim that the confidence value corresponds to the share of correct answers. It describes it as one possible summary of the probability distribution, and says the right thresholds depend on the domain and on how the model performs in the specific use case. It advises starting conservatively, testing with your own data and adjusting thresholds based on the results. S65 The same applies to a new model version: a threshold that fit 1.13 may not fit 1.14. The confidence and calibration guide covers this in more detail.

Language adds to this. According to the models page, Jev achieves its best accuracy in English, and other languages should be tested on your own content. S01 For Slovak texts, there is therefore no vendor figure that would show whether the new version behaves better or worse.

What the documentation does not yet cover

The page on known weak spots exists only for version 1.13. The vendor last reviewed it on 2 October 2026 and it does not mention version 1.14; it only says that many of the described problems are to be fixed in later versions, without saying which ones. S66 No similar page for 1.14 appears in the documentation index either. S30 Until one is added, it is reasonable to treat the advice for 1.13 as a starting point: leave counting, dates and arithmetic to code, send only the state that is needed, and with Choice questions check that the answer does not depend on the order of the options. S66

The models page also describes a long-context mode. By default, the limit is 64,000 tokens per request and 32,000 tokens per question; with the TypeSafe-Danger-Zone-Long-Context: true header, both limits rise to 262,000 tokens. The price per token is the same, but the vendor writes that answers are slower and less accurate, and recommends this mode only when it is really needed. S01

What to do in a Slovak online shop

First, find out which model is actually answering you. Every response contains a model field with the concrete version, so checking your logs is enough. S01 If you now see jev-1.14.0 there and have tuned thresholds, you have two options. You can specify jev-1.13.0 in the request, which is still listed as current, and keep the previous behaviour. S01 Or you can keep the new version and re-check the thresholds on a sample of your own messages, for example on a few hundred complaints that have already been classified by hand. The second approach is a JEV.sk suggestion, not a vendor recommendation.

If you pin a version, expect to have to change it by hand later. Costs do not change with the switch, because both versions have the same price. S01 The cost calculator offers an estimate for your own volume. For new projects, it makes sense to put the version number in the configuration from the start, so that every model change is a deliberate decision rather than a surprise in production.