Full Breakdown
Jev: A Non-Text AI Model Aims to Redefine Software Automation
By Drooid · · How we work
Core Event – Launch of Jev, a Decision-Only Model
On September 15, TypeSafe AI released Jev, a transformer-based model that outputs calibrated probabilities instead of text. Marketed as a “System One model” focused on intuition rather than reasoning, Jev’s input tokens cost $0.042 per million while output tokens are free, a price point roughly 240 times lower than Anthropic’s Fable 5.1, which charges $10 per million input tokens. The launch caused a brief API outage due to unexpectedly high demand.
Background & Context – From RLHF to Calibrated Decisions
Diogo Almeida, a former OpenAI researcher who helped invent reinforcement learning from human feedback (RLHF), left OpenAI two years ago to address what he calls the “language-only” optimization problem. Almeida’s new technique, “reinforcement learning from calibrated decisions,” trains Jev exclusively on synthetic data, sidestepping the need for massive text corpora.
Data & Statistics – Speed, Cost, and Early Benchmarks
- Response time: 70 – 500 ms per query.
- Pricing: $0.042 per million input tokens (free output tokens).
- Internal benchmark: Jev ties Sonnet 5 at 67.8 % on TypeSafe’s workflow evaluation, using averaged answers from GPT-6 Astra and Claude Fable 5.1 as reference. Independent third-party benchmarks are still limited.
Official Statements & Responses – Views from Developers and Partners
- Vercel: Software engineer Pranit Sharma reported that swapping OpenAI’s Luna classifier for Jev produced results 5 to 18 times faster with greater accuracy.
- Bryo AI: CTO Nikhil Mudholkar found Gemini slightly more accurate but 10 to 20 times more expensive than Jev.
- Almeida: Emphasized that calibrated confidence scores let developers automate high-confidence decisions and route uncertain cases to human reviewers, a capability he says is essential for reliable automation.
Verbatim Quotes
- “We have lightning in a bottle, and yet it is not useful,” — Diogo Almeida
- “At the end of the day, it delegates the hallucination problem a little bit to the user,” — Armin Ronacher, the CTO of Earendil
- “We think that there’s just going to be smart software all over the place in a way that’s emergent and distributed … much more like the early internet than you know like the the mega apps that people are trying to build right now,” — Diogo Almeida
- “If a model can do a task 95% of the time but doesn’t say when it’s in the 5%, it can’t automate that task,” — Diogo Almeida
Why It Matters – Potential Shift in AI-Powered Automation
By eliminating free-form text generation, Jev avoids the hallucinations that require downstream parsing and validation. Its low latency and cost aim to make decision-level AI viable for real-time model routing, LLM safety checks, and large-scale software automation. Almeida likens the anticipated diffusion to the early internet, suggesting that cheap “intelligence” could become embedded across many applications rather than concentrated in a few “mega apps.”
What’s Next – Expansion and Competition
TypeSafe AI plans to build additional Jev versions in new modalities. Observers, including Ronacher, expect competitors to launch similar decision-only models now that the concept has proven utility. The broader ecosystem may see Jev-style models used as inexpensive guards for more expensive LLM agents and as routers that direct tasks to the most appropriate foundation model.