TypeSafe AI Jev: An AI model that doesn't generate text – and is 200x faster

On September 15, 2026, TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, launched a model that does not generate text. Jev – named after the Apollo 11 lunar module – produces only typed decisions: Choices, Scores, and Yes/No probabilities. The result is 20 to 200 times faster and 40 to 400 times cheaper than classic LLMs.

System One: The fast way of thinking for machines

TypeSafe AI calls Jev a “System One Model” – inspired by Daniel Kahneman’s Thinking, Fast and Slow. System 1 is the fast, intuitive judgment: “Is this email spam?” – without thinking for long. System 2 is the slow, conscious analysis.

Large language models like GPT, Claude, or Gemini are built as System 2 – they reason step by step. Jev is pure System 1: It makes fast, calibrated decisions without generating a single word.

The three primitives: Choice, Score, Noul

Jev knows exactly three output forms – nothing more:

All three can be combined in a single API call – and in parallel. That means: You ask 20 questions about the same input, and Jev answers them all simultaneously in one forward pass. No token-by-token generation, no autoregressive loop. That is the source of the speed.

The numbers: What really impresses

The independent editorial team at Every was the first to test Jev with their own data:

• 777 decisions across 37 documents in under 0.7 seconds – for about a quarter of a cent
• Median 0.35 seconds per run versus 8.83 seconds for Claude Fable 5.1 – about 25x faster
• Approximately 1/580 the cost of Claude Fable 5.1

TypeSafe’s internal benchmarks are even more impressive: 193.6x faster and 444.6x cheaper than comparable LLMs in the best case.

That output is free is not a marketing trick – it is mechanical: Jev has no autoregressive decoding, hence no tokens to bill.

Can Jev really not hallucinate?

That is the most discussed question. The answer is nuanced:

Yes, structurally: Jev cannot return a value outside the defined output schema. Choice will always be one of the up to 255 options. Score will always be within the defined range. Schema violations are mathematically excluded.

But: A schema-conforming value can still be factually wrong – just as a classifier with high confidence can be mistaken. TypeSafe CEO Diogo Almeida himself acknowledged this in the Hacker News discussion: “It is possible to be wrong with high confidence – and all future models will be smarter and still have this possibility.”

The decisive advance is not zero hallucination, but calibrated confidence: Jev provides a probability distribution for every decision. A model that says “70% confidence” should be right in 70% of cases. That is a completely different property from LLMs, whose confidence is notoriously uncalibrated.

RLCD: The training makes the difference

Jev is trained with Reinforcement Learning for Calibrated Decisions (RLCD) – a method that Diogo Almeida himself distinguishes from his work on RLHF (which shaped ChatGPT):

• RLHF optimizes for human preferences
• RLVR optimizes for verifiable correctness
• RLCD optimizes for calibration – the confidence should be right

This difference is the entire research thesis of TypeSafe: A model that knows when it does not know something is more valuable for automated decisions than one that always has an eloquent answer ready.

What Jev is intended for (and what not)

TypeSafe is clear: Jev does not replace GPT or Claude. It has no chat interface, cannot write code, cannot see images, and cannot generate text.

Use cases:

• Routing logic in AI agent pipelines
• Content moderation (“Spam or not?”)
• Document classification
• Quality scoring of texts
• Decision trees in production systems
• Yes/No checks (PII detection, policy compliance)

Not suitable for:

• Text generation, summaries, code writing
• Complex reasoning tasks
• Creative work
• Conversational interfaces

Early criticism – and what is justified

The Hacker News community discussed Jev intensively – 256 comments on launch day. The main objections:

1. “Frontier Model” is exaggerated. Jev cannot hold a conversation, cannot write text. It is a niche model, not a general-purpose LLM.
2. The Doom demo lacks context. Jev plays Doom – but based on structured JSON positions, not visual input. Impressive for decision latency, but not general game intelligence.
3. Speed comparisons are not always fair. Jev against an LLM in fully autoregressive mode with long output – of course the model without generation wins.
4. Benchmarks are missing. TypeSafe has not published any public leaderboards – their internal workflow evals compare against the average of two models, not against ground truth.

Despite this criticism, the overall tone was surprisingly positive: The core idea – a model specialized in decisions, extremely fast and cheap – was welcomed by many as long overdue.

What this means in practice

Jev is not the model that replaces GPT or Claude – it is the model that complements them:

“The real choice is not Jev versus Claude. It is to define the decision logic interactively with an LLM – and then have that logic executed in production by Jev, at a fraction of the cost and latency.” – Hacker News commenter

In a typical architecture:

1. An LLM (GPT, Claude) defines the decision rules
2. Jev executes them in production – thousands of times per second
3. At low confidence, Jev escalates back to the LLM

Availability

Jev has been available in Early Access via a waitlist since September 15, 2026. Public API access is in preparation. OpenRouter listing, model ID: code>jev-latest/code>.

Conclusion

TypeSafe AI has delivered a proof of concept with Jev that addresses a real gap. The prices and speeds are impressive even in independent validation. The model is not a chat rival – but for automated decisions in production pipelines, it could be the first true specialist in a field previously occupied by generalists.

The coming months will show whether TypeSafe can keep its promises at larger scale and whether the RLCD research leads to a proper scientific publication.

Sources

• TypeSafe AI – Introducing System One Models & Jev
• explainx.ai: TyypeSafe AI Launches Jev (2026)
• Jev: TypeSafe’s Decision Model, Speed and Cost Explained – OrcaRouter
• Hacker News Launch Discussion (256 Kommentare)
• TypeSafe AI – Homepage