The San Francisco start-up, founded by former OpenAI researcher Diogo Almeida, introduced Jev as its first “System One” model after two years in stealth. Almeida helped develop research behind ChatGPT and reinforcement learning from human feedback, a training approach that helped make large language models better at following human instructions.
Jev takes a different route. Instead of producing strings of text, it returns typed decisions and probability estimates that software can use directly. TypeSafe says the model is aimed at tasks such as classification, routing, scoring, verification and guardrails, where an application needs a bounded judgement rather than a written response.
The company says Jev can complete such decisions in roughly 70 to 500 milliseconds and charges $42 per billion input tokens, with no charge for output tokens. Those performance and cost claims are based on TypeSafe’s own evaluations and have not yet been independently validated across a broad range of production workloads.
Developer testing has nevertheless generated attention. Vercel has made Jev available through its AI Gateway, describing it as a probabilistic decision model that can return Choice, Score and Boolean answers in parallel. The platform said the structure removes unnecessary text generation and can let software automate high-confidence cases while sending uncertain ones for review.
TypeSafe also briefly lost the ability to serve users through its API after launch demand exceeded available capacity, underscoring the initial developer interest.
Pranit Sharma, a software engineer at Vercel, tested Jev on a classifier used to review commands for safety. In that test, replacing an OpenAI model with Jev produced results between five and 18 times faster while improving accuracy, according to figures shared publicly about the experiment.
Nikhil Mudholkar, chief technology officer at Bryo AI, separately compared Jev with Google’s Gemini for classifying business emails. Gemini was slightly more accurate in his test, but Jev was substantially cheaper. Mudholkar highlighted the probability scores returned with each decision, which can help developers set thresholds before software acts automatically.
That distinction is central to TypeSafe’s pitch. Conventional large language models can be prompted to return structured formats, but they still generate tokens sequentially and may produce malformed or unreliable responses. Jev instead limits the possible outputs in advance and attaches probabilities to them, allowing surrounding code to decide whether to proceed, retry or seek human review.
TypeSafe describes the training method behind Jev as Reinforcement Learning for Calibrated Decisions, or RLCD. Almeida says the company relies heavily on synthetic data and has designed the model around machine-facing automation rather than human conversation. The company has not disclosed full architectural details or released the model weights.
The model is intended to complement, as well as replace, language models in some narrowly defined tasks. Developers could use Jev to select which model should handle a request, route work between software agents, assess whether an action appears safe, or monitor another model’s output for possible jailbreaks or policy violations.
Armin Ronacher, chief technology officer of Earendil and a developer of the open-source Pi model harness, has said Jev’s probability scores could be useful for model routing and other real-time decisions. He also noted that the approach shifts responsibility to developers to decide what confidence level is sufficient for an automated action.
That caveat matters because a probability is not a guarantee of correctness. A model can still make the wrong bounded decision, even when its output is well formed, and production systems must determine when errors are tolerable. TypeSafe’s claim that Jev cannot hallucinate refers to its inability to invent open-ended text outside predefined outputs, not to an inability to make mistakes.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
