Jev, a model that outputs probabilities instead of text, runs 25x faster and ~600x cheaper than Fable

On September 16, Dan Shipper of Every revealed that his team spent nearly a week testing Jev, a new kind of foundation model from typesafeai. Unlike conventional language models, it outputs probabilities rather than text — Shipper likened it to "a code linter for knowledge work" and called it something people will find indispensable in 6-12 months.

Confirmed

• The model, named Jev, was developed by typesafeai; Every collaborated on roughly a week of testing, which Shipper described as striking
• Core trait: it generates no text, only probabilities, making it efficient as a judge model
• In comparable judging tasks it was 25x faster and 600x cheaper than Fable
• Example use cases: checking code against standards, detecting AI-sounding writing, rating whether tweets are worth reading

Why it matters

Outputting probabilities directly bypasses text generation, enabling extremely low latency and cost in workflows built on judging, scoring and filtering. This could reshape how AI review and quality-control pipelines are architected. Note that the claims come from Shipper's own testing and await broader independent verification.

2026-09-16 ~ 2026-09-16 · 7 related posts

• [source] A foundation model that outputs probabilities, not words: 25x faster, ~600x cheaper — danshipper · 2026-09-16
• Every tests a probability-output foundation model: 25x faster, 600x cheaper than Fable as a judge — danshipper · 2026-09-16
• TypeSafe's probability-output model: 25x faster, 600x cheaper than Fable as a judge — danshipper · 2026-09-16
• A new type of LM that outputs probabilities: 25x faster and 600x cheaper as an LLM judge — danshipper · 2026-09-16
• Jev outputs probabilities, not words: 25x faster and 600x cheaper than Fable as an LLM judge — danshipper · 2026-09-16
• Probability-only model tested as LLM judge: 25x faster, 600x cheaper than Fable-class — danshipper · 2026-09-16

1 near-duplicate retellings: danshipper