TypeSafe AI Unveils Jev, a Fast AI Model Built for Software Automation
Friday, 18 September 2026, 23:01
A former OpenAI contributor says the next automation breakthrough may not speak back at all, but quietly choose the right action.
Diogo Almeida, who helped create ChatGPT at OpenAI, believes that language models are already good at processing human language, but are still not particularly effective at automation. That is why he left the company two years ago and founded TypeSafe AI.
We have lightning in our hands, but it is of no use to us. We have become excellent at working with human language, but that does not work for automation because computers communicate in a different language.
– Diogo Almeida in an interview with TechCrunch
A model that does not generate text
This week, TypeSafe AI introduced Jev – a transformer-based system that does not belong to the class of large language models. Instead of generating textual responses, it calculates probabilities for a predefined set of options. The company calls these outputs calibrated decisions.
The developer defines the list of possible responses, so Jev does not generate arbitrary text and is less prone to hallucinations. The lack of generation also reduces the system’s latency and operating cost: output tokens are not billed, while the volume of input data is measured in billions rather than millions of tokens.
Demand for the new model proved so strong that TypeSafe AI was briefly unable to serve all users through its API. For now, software process automation is considered Jev’s most promising application.
How Jev performed in testing
Vercel engineer Pranit Sharma said that his team had been using ChatGPT Luna 5.6 to check commands for safety. After switching to Jev, responses arrived 5–18 times faster, while classification accuracy improved.
Nikhil Mudholkar, CTO of Bryo AI, tested Jev and Gemini while classifying business correspondence. Gemini was slightly more accurate, but using it cost 10–20 times more. At the same time, Jev returned a calibrated confidence probability, which is especially important for automatically triggering workflows.
This is the only system that returns a true probability, which makes it ideally suited to workflow automation.
Jev can not only perform individual tasks in place of language models, but also check their work. Potential use cases include analyzing software agents’ activity logs, detecting dangerous behavior, and blocking attempts to bypass established restrictions.
Armin Ronacher, CTO of Earendil, noted that the probability score helps distribute responsibility between the system and the user. A result with 50% confidence can be rejected, while a score of 95% may be sufficient grounds for taking an automated action.
Why the model could become a mass-market tool
Another potential application for Jev is model routing – the automatic selection of the system best suited to a particular task. Using a large language model for such predictions is expensive, while Jev’s speed and low cost could enable it to make similar decisions in real time.
The name Jev refers to British economist William Stanley Jevons, whose name is associated with a paradox: making a resource cheaper often leads to increased consumption. Diogo Almeida hopes that affordable “intelligence” will accelerate the spread of software capable of making decisions autonomously.
We believe intelligent software will appear everywhere – distributed and almost invisible, resembling the early internet more than the large, universal applications people are trying to build today.
TypeSafe AI has not yet disclosed Jev’s architecture. Outside observers speculate that the system may be built on a language model with open weights. The company itself describes Jev as a “System One” model focused on intuitive decision-making and carrying out specific tasks.
The system is trained exclusively on synthetic data using an approach that TypeSafe AI calls reinforcement learning based on calibrated decisions. The team plans to release new versions for working with other types of data. Armin Ronacher expects competing solutions to emerge once Jev demonstrates its practical value in the market.
If the model maintains its claimed speed, accuracy, and low cost, it could become a separate class of tools for software automation – without textual responses, but with decisions ready for immediate use.
We recommend paying attention to:
- Warp has launched Warp Factories, a platform that helps engineering teams coordinate AI agents for coding, testing, review, and software development workflows.
- Ramp has launched Router, an AI model routing service that lets businesses connect to multiple providers through one API while managing costs, performance, and usage.
- AI-generated legal errors have appeared in at least 1,395 US court cases, despite sanctions and repeated warnings that lawyers must verify submissions.
