AI DevelopmentNew Release8 min readPublished September 16, 2026

No text output · typed values · probabilities · vendor figures only

TypeSafe Jev: A Model That Returns Decisions, Not Text

TypeSafe AI's Jev returns typed values with probabilities, not text, and claims 70 to 500 ms responses. What the vendor figures mean and how to test them.

DA
Digital Applied Team
Research and practical guidance
Editorial dateSeptember 16, 2026
SourceTypeSafe AI announcement

On September 15, 2026 a startup called TypeSafe AI announced Jev, a model that cannot write a sentence. You give it a block of text describing a situation and a list of questions with the allowed answers. It returns one of those answers for each question, with a probability attached. TypeSafe calls this class a System One model and says Jev responds in 70 to 500 milliseconds.

This post is for developers and founders who pay a language model to make thousands of small decisions a day: label this ticket, route this lead, pull this field, approve or escalate. It reads TypeSafe's announcement, written by founder Diogo Almeida, and separates what the design guarantees from what the vendor measured. Every number below is TypeSafe's own. No independent benchmark of Jev existed when we read the post on September 17, and access is through a waitlist.

Key takeaways
  1. 01
    Jev returns values, not text.Possible outputs are declared in advance and produced in one parallel step. TypeSafe says the model gives up string generation entirely.
  2. 02
    Zero type errors is a design property.TypeSafe says its 0% figure is not empirical. An output that must come from a declared list cannot have the wrong type. It can still be the wrong choice.
  3. 03
    The speed and price claims are vendor-run.70 to 500 ms per call and $0.042 per million input tokens are TypeSafe's figures, measured from laptops near its US West Coast service.
  4. 04
    Calibrated confidence is the real pitch.A decision that arrives with an honest probability lets your code send the uncertain cases to a person. Test that claim before any other.

01The releaseWhat TypeSafe announced

Almeida writes that he worked at OpenAI on the instruction-following methods behind ChatGPT, and that TypeSafe spent two years in stealth. His argument is that chat models are good at talking to people and awkward inside software, because software needs a value it can branch on, not a paragraph it has to parse.

The name borrows from Daniel Kahneman's Thinking, Fast and Slow, where System 1 is fast, intuitive judgement and System 2 is slow reasoning. TypeSafe's one-line description of Jev is "a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out".

Three parts are named and not explained further: a new model architecture, a parallel sampler, and a training method called Reinforcement Learning for Calibrated Decisions, or RLCD. The post gives no parameter count, no licence and no information about weights. Jev is a hosted service in early access, and TypeSafe says it is bringing developers off the waitlist as fast as it can. A waitlist is not availability, so plan accordingly.

02The tradeWhat Jev gives up and gains

An ordinary language model writes one token at a time, and anything can come out: an answer, a refusal, malformed JSON. Jev's outputs are fixed before the call. The table sets the two side by side, as TypeSafe describes them.

Source: TypeSafe AI announcement, September 15, 2026. Descriptions are the vendor's, paraphrased.
AspectGives upGains
OutputAll free text: no drafting, summaries, code or explanationsValues that always match the declared type, so no parsing or validation step
GenerationStep-by-step reasoning written out before the answerAll answers produced in a single parallel query, which is where the speed comes from
ConfidenceNothingA probability with every answer, which TypeSafe says is calibrated: higher confidence means higher accuracy
Choice sizeOpen-ended answers; a single choice is capped at 255 optionsA two-stage score-then-choose step for larger option sets, which TypeSafe says is sometimes slower

Two claims in the post need careful reading. TypeSafe says Jev "can't hallucinate". In context that means it cannot emit invented text, because it emits no text. It does not mean its decisions are always right. A spam label can be wrong without being a hallucination.

The same applies to type errors. TypeSafe writes that its 0% figure "is not empirical" and that schema matching is guaranteed by construction. That is a useful property, and it is a property of declaring outputs in advance, not a measure of accuracy. Structured output modes on mainstream models offer a version of the same guarantee, which we covered in our guide to OpenAI structured outputs.

03The figuresThe vendor's numbers

These are the figures in the announcement. Each is TypeSafe's own measurement or list price. None has been reproduced by a third party.

End-to-end response time, JevVendor-run, from laptops on the US West Coast, where the service is hosted
70–500 ms
End-to-end response time, frontier LLMsTypeSafe's comparison range
3–329 s
Stated speed-up on System One queriesVendor claim for the same level of intelligence
40x–200x
Input priceList price; the post also writes it as $42 per billion tokens
$0.042 / MTok
Output priceTypeSafe: too cheap to meter
Free

TypeSafe attaches its own caveats, and they are unusually frank. The headline figures on its home page, 193.6x faster and 444.6x cheaper, come from four workflow tests that it expects "are on the higher end of real world gains". The workflows were written by its own model capabilities team, "so some bias could exist". On pricing, it says it cannot prove the price is not subsidised.

The accuracy method matters too. There is no ground-truth answer key. TypeSafe scores every model against the average of GPT-6 Astra and Fable 5.1 running the same workflow, and says Jev reaches similar results. That measures agreement with two large models, not correctness. The comparison models also ran through TypeSafe's own wrapper that forces probability outputs, which the post says makes them slower and more expensive than a plain call.

Location is the last caveat. A 70 ms response measured next door to the server will not be 70 ms from Frankfurt or Sydney. If your use case depends on real-time speed, measure from where your code runs.

04FitWhich jobs fit

The test is whether the answer you need can be written as a list of allowed values before the call. TypeSafe's phrase for the fitting jobs is "smart if-statements". The grouping below is ours, based on the constraints the post states: no strings, text input only, and a fixed output shape.

Editorial grouping from the limits stated in TypeSafe's announcement of September 15, 2026.
JobFitWhy
Routing a ticket, lead or documentYesA short fixed list of destinations, and volume makes latency and price matter
Classification and scoringYesLabels, yes or no, a number on a scale: all declared in advance
Approve or escalate gatesYesThe probability decides which cases go to a person
Checking another model's outputYesTypeSafe lists judging, guardrails and jailbreak detection as intended uses
Extracting a name, address or free-text fieldNoThe value is a string, and Jev does not generate strings
Drafting, summarising, explainingNoText output
Open-ended agent planningNoThe next step is not a choice from a known list
Anything that reads imagesNot yetTypeSafe says its demos run on text state, "not on images (yet…)"

Most real automations mix both kinds of step. An agent that drafts a reply needs a text model. The step before it, deciding whether this message deserves a reply at all, is a decision. Our guide to choosing between an agent and a fixed workflow covers how to split a process that way. A decision model is a candidate only for the fixed-workflow half.

05The argumentWhy confidence matters more than accuracy

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, founder, TypeSafe AI, September 15, 2026

This sentence is the strongest part of the announcement, and it holds whatever model you use. Average accuracy tells you how many errors to expect. It does not tell you where they are. If every answer looks equally sure, a person has to review all of them, and nothing has been automated.

A calibrated probability changes the arithmetic. Calibrated means that of all the answers given at 90% confidence, about 90% are right. With that, your code can act on the confident answers and queue the rest for review. The review queue shrinks to the hard cases.

TypeSafe says mainstream models asked for a confidence score "tend to be overconfident and inconsistent". It is also why many production pipelines ignore a model's self-reported confidence. Whether Jev's probabilities are better is the claim to test, because it is the one that decides how much human review you can remove. Our post on structured output reliability in production covers the validation layer you would keep either way.

06MethodHow to test the claims on your workload

Do not move a production step on a vendor's benchmark. You can prepare the test before you have access, because most of the work is on your side.

  1. Pick one decision step you already run at volume. Save 500 to 1,000 real inputs with the outcome a person would accept as correct. This is the answer key TypeSafe's own tests lack.
  2. Run them through your current model in its structured output mode. Record accuracy against your key, latency from your own servers, and cost.
  3. Run the same inputs through the candidate. Record the same three numbers, plus the probability on every answer.
  4. Sort the candidate's answers by confidence and check accuracy in each band. If the 90% band is right about 90% of the time, the confidence is usable. If the low-confidence band is no worse than the high one, it is decoration.
  5. Read the disagreements by hand. Fifty rows will tell you whether the errors are ambiguous inputs or plain mistakes.
A cost sanity check

TypeSafe's own demo gives a useful scale. It says a bot making 10 queries a second costs about $7 an hour. At that rate a decision step is cheap per call and not free at volume. Multiply by your real call rate before you compare it with a small mainstream model on a batch discount.

We run this kind of side-by-side test for clients through our AI transformation practice. The method is the same for any new model class: your inputs, your answer key, your network.

07Next stepA narrow model with one claim worth testing

Put it into practice

Build the answer key now, join the waitlist second

List the steps in your automations where a model picks from a fixed set of answers. Choose the highest-volume one and save a labelled sample of real inputs. That sample will tell you whether your current model's confidence can be trusted, and it is ready for Jev or any later decision model when access arrives. Treat the speed and price figures as vendor claims until your own numbers replace them.

Digital Applied

Test a new model on your own decisions before you switch.

We build labelled test sets from your real workload and measure accuracy, latency and confidence side by side, so a model change is a measured decision.

Labelled test setsLatency from your serversCost models
Your next project

Start with one high-volume decision

  • Save 500 real inputs with outcomes
  • Measure your current model first
  • Check accuracy by confidence band
Questions and answers

Applying this post

Only through early access. TypeSafe's announcement of September 15, 2026 says Jev is available in early access and that developers are being brought off a waitlist. The post gives no general-availability date, no licence terms and no information about model weights.
Digital Applied newsletter

Deep dives on AI, marketing and development.

Practical guides and fresh insights by email. No recycled takes.

Related dispatches

Continue reading

AI Development

Paper2Agent in Nature: Research Papers Become Working Tools

A Stanford method published in Nature turns a paper and its code into tested tools an AI agent can call. The reported results, and the pattern for business.

September 16, 2026 · 8 minRead
AI Development

AI Agents Left a Public Trail: Logs to Keep for Next Time

SentinelLABS matched public Hugging Face commits to OpenAI's May agent incident, to the second. What the report shows, and the logs a platform should keep.

September 16, 2026 · 8 minRead
AI Development

Claude Chat and Cowork Are Now One App: What Teams Should Do

Anthropic merged Claude chat and Cowork on September 16 and added Docs and Slides in beta. The rollout by plan, and the one approval setting to decide.

September 16, 2026 · 6 minRead
AI Development

Eight Worlds of AI Agents Faced Three Attacks: None Passed

Emergence AI ran eight worlds of ten agents for up to 21 days, then staged three attacks. No world passed all three. The scores, and three fixes for builders.

September 16, 2026 · 8 minRead
AI Development

State of AI Agents 2026: 200+ Data Points Compiled

The definitive State of AI Agents 2026 — 247 data points across adoption, ROI, autonomy, and governance, sourced from McKinsey, Stanford HAI, and Gartner.

May 22, 2026 · 16 minRead
AI Development

AI Video Generation 2026: Omni vs Sora vs Veo 3 Compared

Gemini Omni, OpenAI Sora 2, and Google Veo 3.1 compared for video — quality, per-second cost spread of 17x, and the September 24 Sora API sunset clock.

May 22, 2026 · 15 minRead