TypeSafe AI released Jev 1.13.0 in mid-September, and three days later Convai Innovations posted Laya, an open-weight counterpart built on ModernBERT. The releases are close enough in timing and intent that they should be read together: a new model class aimed squarely at the classification layer that has been quietly draining budgets around every production LLM stack.

Neither model writes prose. Both return typed, calibrated decisions, the kind of output most teams currently extract from a frontier model by way of prompt engineering and hope.

The price disclosure is what forces the category into view. TypeSafe meters Jev at $42 per billion input tokens, or $0.042 per million, against a frontier band it cites at $0.20 to $10 per million. TypeSafe's own demo claims Jev is 193.6x faster and 444.6x cheaper than the average of two frontier reference models, though MarkTechPost notes the reference answers and the workflows themselves came from TypeSafe's capabilities team. Reported end-to-end response times sit between 70ms and 500ms. Weights, parameter counts, and self-hosting aren't on offer.

Laya fills the gap TypeSafe won't. Convai posted three checkpoints under Apache 2.0: a 421M English base on ModernBERT, a 322M multilingual variant on mmBERT-base covering 100+ languages, and a typed-decisions checkpoint. The multilingual model records 32.8ms p50 latency on a Tesla T4. On a shared typed-decisions benchmark, Laya's fine-tuned checkpoint scores 0.766 accuracy against Jev 1.13.0's 0.727, with the caveat that Laya was fine-tuned on that benchmark's own training split. Zero-shot, Laya lands at 0.362, barely above the 0.318 random baseline. Calibration tells a similar story: 0.466 mean ECE out of the box, 0.081 after refitting one temperature per question type.

Developer receipts are already circulating. Vercel engineer Pranit Sharma clocked Jev running a command-safety classifier five to 18 times faster than OpenAI's Luna 5.6. Bryo AI CTO Nikhil Mudholkar told TechCrunch that Gemini was slightly more accurate on email classification but 10 to 20 times more expensive.

KDnuggets contributor Abid Ali Awan framed the significance directly: classification, intent detection, zero-shot classification, and calibrated probabilities are not new, he wrote, but TypeSafe AI "appears to have done is rethink the architecture, training, inference, calibration, and developer experience around these familiar problems."

That's the honest read. The industry has spent two years paying autoregressive tax on decisions that never needed language generation, a pattern this pipeline shares with the broader cost overhang in agentic architectures and with the gap between frontier coding-agent benchmarks and real enterprise work. Independent benchmarks are still owed. The pricing case, for now, is doing the arguing.

Jev and Laya define a new model class: typed-decision heads that replace LLM calls for routing

Sources