TypeSafe AI Launches Non-Text Transformer Model to Cut Software Automation Costs

TypeSafe AI, a startup founded by former OpenAI researcher Almeida, has released Jev, a transformer-based model designed specifically for software automation rather than natural language generation. Almeida, who helped build ChatGPT and co-invented reinforcement learning from human feedback (RLHF), launched the company after concluding that optimizing AI systems for natural language creates unnecessary overhead when automating backend software, as first reported by TechCrunch AI (https://techcrunch.com/2026/09/18/a-new-kind-of-ai-model-from-a-chatgpt-inventor-is-thrilling-developers/).

Instead of outputting text, Jev returns numerical probabilities—what the company terms calibrated decisions. Because users predefine the exact output parameters, TypeSafe says the model cannot hallucinate. The company offers free output tokens and meters input tokens by the billion rather than the million, cutting runtime and operational costs compared to conventional large language models. High initial developer interest briefly overwhelmed the company's API serving capacity at launch.

Early developer implementations highlight significant efficiency gains on classification and validation tasks. Pranit Sharma, a software engineer at agentic infrastructure startup Vercel, reported that swapping OpenAI's ChatGPT Luna 5.6 model for Jev to classify command safety resulted in higher accuracy while running 5 to 18 times faster. At Bryo AI, CTO Nikhil Mudholkar tested Jev against Google's Gemini on email classification; while Gemini was marginally more accurate, it proved 10 to 20 times more expensive, with Mudholkar noting Jev's discrete confidence scores as ideal for automated logic.

Beyond replacing text models, Jev can serve as a lightweight supervisor and routing layer. Armin Ronacher, CTO of Earendil, pointed out that Jev allows developers to programmatically set confidence thresholds to filter low-probability outputs, while its speed and low price point enable real-time model routing without the prohibitive expense of using a full LLM.

Almeida named the model after 19th-century economist William Stanley Jevons, whose paradox describes how increasing efficiency in resource use lowers costs and expands total consumption. TypeSafe classifies Jev as an intuitive 'System One' model trained entirely on synthetic data through a proprietary method dubbed reinforcement learning from calibrated decisions. While external observers suspect the architecture builds on an open-weight model, TypeSafe has kept the foundational architecture private and plans to expand the approach to additional modalities.