Key takeaways
- Jev separates judgement from generation, returning bounded choices, scores and yes-or-no probabilities that software can use directly.
- Its speed and cost could improve startup unit economics and make it economical to classify, route and verify every item or agent action rather than sampling a small subset.
- Jev is not an LLM replacement or a guarantee of correctness. Its strongest role is as a decision and control layer around generative models and conventional code.
For the past few years, the dominant assumption in AI has been that intelligence should look like a conversation.
We type a prompt. A model writes an answer. Even when software rather than a person consumes that answer, we often ask a large language model to generate text or JSON one token at a time, then parse and validate it before anything useful can happen.
Jev, a new model from TypeSafe AI, starts with a different premise: much of what software needs from AI is not a paragraph. It is a decision.
Should this support ticket go to billing or engineering? Is this transaction suspicious? Which ad uses a scarcity hook? Is this tool call destructive? How relevant is this document? Should an agent proceed, stop or ask for human approval?
Jev is designed to make those judgements directly, quickly and at very low cost. If the approach works at scale, it could create a new layer in the AI stack: probabilistic decision infrastructure sitting between generative models and traditional software. For startups, that creates both an opportunity and a threat. It could materially improve the speed, reliability and unit economics of AI products, while commoditising features built primarily around classification, scoring and routing.
Before we go even further, just check out this video of Jev in action to understand the power of this tool:
What is Jev?
Jev is the first product from TypeSafe AI, a San Francisco AI lab founded by Diogo Almeida, a co-author of the InstructGPT paper that helped lay the groundwork for ChatGPT. The company emerged from stealth in September 2026 with $40 million in seed funding led by DCVC.
TypeSafe describes Jev as a “System One” model, borrowing Daniel Kahneman’s distinction between fast, intuitive judgement and slow, deliberate reasoning.
Instead of asking Jev to generate an explanation, a developer sends it two things:
- A state, such as a customer message, document, agent plan, product record or block of JSON.
- One or more typed questions, defining exactly what decisions the application needs.
Jev supports three basic output types:
- Choice: select from a predefined set of options, with a probability for each.
- Score: rate the state against a defined scale or rubric.
- Noul: return the probability that a yes-or-no statement is true. The name is derived from Bernoulli.
All the questions are evaluated in parallel against the same state. The response is structured and ready for software to use. There is no prose to interpret and no JSON generated token by token.
TypeSafe currently lists Jev at $0.042 per million input tokens, with output free, and reports end-to-end latency of roughly 70 to 500 milliseconds. The company says its architecture reads the state once and produces all the requested decisions in parallel.
That is the core idea: unstructured state in, typed probabilistic decisions out. The name Jev is a reference to economist William Stanley Jevons and the Jevons paradox: when a resource becomes more efficient and cheaper to use, total consumption can increase. TypeSafe is making the same bet about machine intelligence.
Why this is potentially disruptive
1. It separates intelligence from language generation
LLMs have become the default tool for almost every AI workload, including many that do not need language generation at all. We routinely pay a general-purpose model to reason, compose a response and format it as JSON when the application only needs to know whether something is relevant, risky or urgent.
Jev suggests that generation and judgement may become separate model categories.
A frontier LLM can still research, reason, write or code. Jev can sit beside it, deciding what deserves attention, which model should handle a task, whether an output meets a standard and what the software should do next. For founders, this creates an opportunity to redesign the model stack rather than defaulting to the most capable and expensive model for every step.
This is less a replacement for ChatGPT or Claude than an unbundling of the work we currently ask them to do.
2. It changes the economics of applying intelligence at scale
At $0.042 per million input tokens, the marginal cost of a decision approaches ordinary software economics. That matters because the interesting use cases are not one-off questions. They are millions of small judgements embedded throughout a product or workflow.
In Matthew Berman’s demonstration, Jev classified 724 live ads from 37 brands across dimensions including hook, format, offer, call to action and awareness stage. The reported run took around 40 seconds and cost nine cents. The capability is expected to be incorporated into StealAds.
The important point is not the nine cents itself. It is what happens when analysing every ad, support ticket, sales call, product listing or agent action becomes cheap enough to do by default rather than by exception. Startups that currently analyse a sample could continuously structure a customer’s entire dataset. That changes the product from an occasional assistant into a persistent intelligence layer across the business.
It could also improve AI unit economics. Many startups currently use frontier models for simple intermediate decisions because those models are already integrated into the product. Moving classification, routing and verification to a cheaper decision model could reduce inference costs without weakening the generative experience that customers actually see.
3. It produces an interface software can act on
Generative models are designed primarily for human-readable output. Production software needs bounded values, predictable schemas and clear thresholds.
Jev cannot invent a fourth option when the developer supplied three. It always returns a value of the expected type, with probabilities that can be used to set operating rules. An application could automatically approve high-confidence decisions, send ambiguous cases to a stronger model and escalate low-confidence or high-risk cases to a person.
That makes it possible to design systems around uncertainty instead of pretending every model answer is equally reliable. A startup could automatically accept high-confidence decisions, send ambiguous cases to a stronger model and reserve human review for genuinely risky cases. This is not only an infrastructure improvement; it can change the operating margin and service model of the business.
4. It could become a control layer for agents
Agents need to make many small decisions before, during and after taking action. Which tool should I call? Is this command consistent with the user’s request? Is this action reversible? Does the result satisfy the goal? Should I retry or escalate?
Using a frontier model for every check increases both cost and latency. Hard-coded rules are fast but brittle. A fast probabilistic model could fill the gap, acting as a policy, routing and verification layer around more capable agents.
This may be Jev’s most important role. The generative model does the work; the decision model helps determine whether, where and how that work should proceed. Faster and cheaper control loops could make agentic products commercially viable in categories where repeated frontier-model calls currently make them too slow or expensive.
There is a defensive implication too. Startups whose main feature is using an LLM to classify, tag, moderate or score information may see that capability become much easier to reproduce. “We apply AI to categorise your data” is unlikely to be a durable moat. The defensibility must come from proprietary data, workflow ownership, distribution, integrations, feedback loops or the actions the product takes after reaching a decision.
Core use cases
| Use case | What Jev does | Examples |
|---|---|---|
| Classification and routing | Evaluates an item across several dimensions and routes it according to the returned choices, scores and probabilities. | Send a support ticket to billing or engineering; qualify an inbound lead; prioritise a claim; classify an email by urgency, topic and risk. |
| Search, retrieval and reranking | Scores whether retrieved content is genuinely relevant before it enters an LLM’s context. TypeSafe’s published reranking example reports higher retrieval accuracy at very low classification cost. | Rerank legal precedents; select the most relevant product documentation; filter research papers; choose which CRM records an agent should read. |
| Agent guardrails and tool approval | Assesses a proposed action before execution and gives application code a fast checkpoint. It should complement, rather than replace, deterministic security controls. | Flag a destructive shell command; detect an off-task action; require approval for irreversible changes; restrict an agent operating outside its permitted scope. |
| Output verification | Checks whether an AI output satisfies defined requirements, making it affordable to evaluate far more responses. | Verify that a citation supports a claim; check whether instructions were followed; detect prohibited content; decide whether an agent completed its task. |
| Model routing | Estimates a request’s difficulty, risk or modality and routes it to the cheapest model capable of handling it. Jev is already available through Vercel AI Gateway. | Send simple extraction to a small model; reserve a frontier reasoning model for complex tasks; escalate high-risk prompts; choose a specialist coding or vision model. |
| Large-scale data enrichment | Converts large collections of unstructured material into consistent, queryable features across multiple dimensions. | Structure product catalogues; enrich CRM records; tag sales transcripts; classify legal documents, survey responses or research archives. |
| Real-time decisions | Makes probabilistic judgements quickly enough for interactive products where multi-second model calls would undermine the experience. | Adapt a game’s behaviour; flag a live fraud signal; moderate content as it is posted; update recommendations; guide an agent during an active session. |
| Advertising and competitive intelligence | Turns creative libraries into structured datasets that can be compared across brands, campaigns and funnel stages. | Identify hooks, formats, offers and calls to action; map ads to awareness stages; detect landing-page mismatches; track how competitors change their messaging. |
What Jev does not solve
The excitement needs a few important caveats.
First, “cannot hallucinate” is too broad. Jev cannot return a malformed type or invent an option outside the schema, but it can still choose the wrong valid answer. Type safety is not the same as truth.
Second, probabilities are valuable only if they are well calibrated for the actual workload. Developers will still need labelled examples, evaluation sets and thresholds. A model that is calibrated in aggregate may perform differently across domains, languages or edge cases.
Third, Jev works best on bounded, decomposable questions. TypeSafe’s own jaggedness documentation warns about weaknesses involving arithmetic, date comparison, multi-step indirection, noisy context and adversarial content. It is not a replacement for a reasoning model, calculator or deterministic security control.
Fourth, defining the choices and criteria is real product work. The flexibility of a chat prompt allows teams to begin with an ambiguous problem. Jev requires them to decide what the system is actually judging and what actions should follow. That constraint may improve production systems, but it is not free.
Finally, many of the most striking speed and cost comparisons come from TypeSafe’s own evaluations. The company acknowledges that its published gains are likely at the high end of real-world results. Independent benchmarking across real production workloads will matter.
A new primitive rather than a new chatbot
The past three years of AI have been dominated by models that speak. Jev is interesting because it asks whether the next wave of automation will depend just as much on models that decide.
If intelligence becomes cheap enough to embed inside every branch of a workflow, AI stops being a destination and becomes part of the application’s control flow: a smarter if statement with uncertainty attached.
That could change where value accrues. Model routers, moderation providers, classification APIs, data-labelling workflows, evaluation tools and some vertical AI features may all face a new low-cost primitive. At the same time, startups can use it to make their own products faster, cheaper and more reliable.
The larger lesson for founders is that model intelligence is becoming cheaper and more modular. Value will increasingly sit in knowing which questions to ask, how to handle uncertainty, what action should follow each answer and how outcomes improve the system. The defensible asset may not be the model call. It may be the decision architecture surrounding it.
The most useful mental model is not “Jev versus ChatGPT.” It is Jev alongside generative models: one model creates, reasons and communicates; another classifies, scores, routes and checks.
The question for founders is simple: how many expensive LLM calls in your product are producing words that nobody ever needed to read?
He is a former General Partner at Google Ventures (GV) in Europe, former head of Google for Entrepreneurs in Europe, and founding head of Campus London, Google's first startup hub. Eze writes on Israeli tech, venture capital, artificial intelligence, and founder strategy.
He is also the founder of Techbikers, a nonprofit that brings together the startup ecosystem on cycling challenges in support of Room to Read.
- Jev and the missing decision layer for AI - September 22, 2026
- Weekly Firgun Newsletter – September 18 2026 - September 18, 2026
- Your AI infrastructure provider wants your customers - September 17, 2026
