Jev from TypeSafe AI: confirmed pricing and concerns about opacity and hidden bias
TypeSafe AI clarified the price of Jev (0.042 USD/million input tokens, output free), and commentators point to the risk of hidden bias and the absence of explanations for decisions in this classification model with no text output.
The latest news about Jev from TypeSafe AI adds pricing details and warnings about risks to earlier reports on speed and low costs. The company confirmed a price of 0.042 USD per million input tokens; output tokens are free. Commentator Simon Willison points out that Jev is even more of a “black box” than conventional language models – it returns no explanation for its decisions, only a number, so sensitive uses (e.g. evaluating job applicants) carry a risk of hidden bias that is difficult to detect. As an example, he cites his own informal test in which the model rated cities in the San Francisco Bay Area based on the question “good city?”, with results suggesting possible bias. Other uses mentioned include spam filtering and reranking search results (search reranking); the company places the model in the “System One models” category.
Jev is a transformer model that returns numbers instead of text – categories, yes/no answers, ratings and confidence levels (calibrated probabilities). Because the outputs are predefined by the user, the company says the model cannot hallucinate in the usual sense of the word, although it can still choose an incorrect option within the allowed answers. According to the company, an answer arrives in 70 to 500 milliseconds, and adding more questions to a query barely increases the response time.
The model was introduced by founder and CEO Diogo Almeida, formerly a researcher at OpenAI who contributed to the development of InstructGPT and the RLHF method. According to developers who tried the model, it delivered practical benefits: Vercel reports a 5 to 18-fold speedup and higher accuracy compared with its previous solution using ChatGPT Luna, while Bryo AI reports a 10 to 20-fold lower cost compared with Gemini for email classification, although Gemini was slightly more accurate in its test.
The source The Decoder adds that the published comparative tests were conducted by TypeSafe AI itself, use answers from other models as a reference (rather than independently verified correct solutions), and do not include GPT-6 Astra. Access to Jev is currently through a waiting list.
Why it matters
For developers, it is a cheap and fast alternative to LLMs for tasks such as sorting requests, detecting intent or checking consistency – independent references (Vercel, Bryo AI) report significant speedups and lower costs compared with conventional models. At the same time, however, the model returns only a number without an explanation, so decisions affecting people (e.g. evaluating applicants) carry a risk of undetectable bias, and companies should conduct their own evaluation before deployment rather than relying solely on figures from the provider.
Release card
Jev
TypeSafe AI
- Price
- 0.042 USD per million input tokens; output tokens are free
- Availability
- Access via API, currently only through a waiting list (waitlist)
- Doba odezvy API 70–500 ms According to TypeSafe AI, this is how long it takes to return an answer to a query; adding more questions barely increases the time.
- Klasifikace bezpečnosti příkazů u firmy Vercel (náhrada modelu ChatGPT Luna 5.6) 5 až 18× rychlejší, vyšší přesnost According to a developer at Vercel, this was the result of the comparison with the classifier previously used.
- Klasifikace e-mailů u firmy Bryo AI (vs. model Gemini) 10 až 20× nižší cena, model Gemini mírně přesnější According to the CTO of Bryo AI, this was the result of the cost and accuracy comparison when sorting company emails.
- Text classification and sorting, for example spam, labels or request priority
- Reranking search results (search reranking) after candidates have been shortlisted by a cheaper algorithm
- Checking outputs from other AI agents, for example detecting jailbreaks or inconsistencies with the conversation
- Returns no explanation for its decisions, only a number - sensitive decisions (e.g. evaluating job applicants) carry a risk of hidden bias that is difficult to detect
- Cannot hallucinate beyond the allowed options, but can still select a factually incorrect answer within them
TypeSafe AI places Jev in the “System One models” category and contrasts it with conventional language models, which generate text for classification tasks and are slower and more expensive according to the cited comparisons; in the deployments mentioned, it replaced or bypassed ChatGPT Luna 5.6 and Gemini.
The card summarizes information from the article and any dated corrections, with a link to the original source. It is not our assessment of the model. It does not yet have a dedicated editorial profile. Model selection and other announcements →
What was added since the original report
Verified updates
-
Model price: $0.042 per million input tokens, output free; Risk of hidden bias in the classification model; The model operates as a black box without explaining its decisions; Applicable to spam filtering and search reranking; TypeSafe AI categorizes it as 'System One models'
- Model price: $0.042 per million input tokens, output free
- Risk of hidden bias in the classification model
- The model operates as a black box without explaining its decisions
- Applicable to spam filtering and search reranking
- TypeSafe AI categorizes it as 'System One models'
-
Jev is a transformer with predefined outputs, so it cannot hallucinate or make up answers; Output tokens are free; input tokens are measured in billions instead of millions; Vercel achieved a 5–18× speedup and better accuracy than ChatGPT Luna using Jev; Bryo AI found that Jev is 10–20× cheaper than Gemini for email classification; Diogo Almeida actively worked on developing RLHF at OpenAI and addressed the problem of hallucination in LLMs
- Jev is a transformer with predefined outputs, so it cannot hallucinate or make up answers
- Output tokens are free; input tokens are measured in billions instead of millions
- Vercel achieved a 5–18× speedup and better accuracy than ChatGPT Luna using Jev
- Bryo AI found that Jev is 10–20× cheaper than Gemini for email classification
- Diogo Almeida actively worked on developing RLHF at OpenAI and addressed the problem of hallucination in LLMs
Two audiences, two different impacts
What this means
For individuals
Developers integrating classification into software (query routing, intent detection, consistency checks) gain a cheap and fast alternative to conventional LLMs, but must account for the fact that the model returns only a number without an explanation and requires their own evaluation before deployment.
For a business
Companies deploying classification and automation (request routing, spam filtering, result ranking) may, according to the cited cases (Vercel, Bryo AI), reduce costs and speed up processing, but the model does not explain its decisions and carries a risk of hidden bias, which is a problem especially for decisions affecting people.
DevelopmentCheck the original
Event sources
confirmed by 3 independent sources · 3 publishers, 3 independent. We count feeds from the same owner only once.