# Jev vs GPT and Haiku: the software that decides without writing

> Jev, TypeSafe's System One model, decides in milliseconds for $0.042 per million without writing text. Price and speed comparison against GPT and Haiku.
> By José Maldonado · Artificial Intelligence · 2026-10-09

TypeSafe AI, a startup emerging from two years of stealth, presented on September 15 its first System One called Jev, a model that does not generate text but structured decisions with calibrated probability, according to the [company's official announcement](https://typesafe.ai/blog/introducing-system-one-models-and-jev). Its demo video surpassed 30 million views in days and developers began replacing frontier models in their classification pipelines.

The success of this idea can be seen clearly in two figures. Jev responds in 70 to 500 milliseconds and costs $0.042 per million input tokens, with output free because it is too cheap to measure, always according to the manufacturer's own figures. Against the $0.20 to $10 per million charged by large language models, plus output that usually costs five times more, the promise is breakneck speed at bargain price.

The key is that Jev gives up words to keep judgment. It receives a situation described in text or JSON and some questions with allowed answers, and returns labels, scores or yeses and nos with their associated confidence. It does not write or reason aloud, because it cannot make a type error in its output.

## A model that does not write

The idea draws on Kahneman's psychology and his fast and slow thinking. System One models (like Jev) decide by trained intuition, while System Two models (like Gemini, Claude or GPT Astra) deliberate step by step like a current reasoning model. TypeSafe transfers that distinction to software architecture, with one model to decide and a different one to write when prose is needed.

The contract with the programmer breaks with everything known. Instead of messages, a state is sent with typed questions and each answer comes with the chosen option and its probability. Those shapes cover typical automatable work, from classifying tickets to scoring risks.

The format's limits are as explicit as its virtues. Each request allows 64,000 tokens of context and the state plus the long question cannot exceed 32,000. It only accepts text as a string, JSON object or list, with no image, audio or video, and the model does not adapt with customer data or train on the requests it receives.

After all, the breakthrough is not accuracy but the mechanical guarantee. A language model with constrained JSON output can break the schema under pressure and everything must be validated. Jev returns native types by construction, so the subsequent verification layer is reduced to checking the business logic and not the format.

## Jev vs the big players in figures

An honest comparison requires laying all the cards on the table. Frontier model prices vary by provider and date, and Jev's figures come from the seller itself without independent audit, as TypeSafe itself acknowledges. With that caution, the table summarizes the order of magnitude of each alternative as of September 2026.

Model

Typical latency

Input price

Output price

Its turf

Jev 1.13

70 to 500 ms

$0.042/M

Free

Classify, route, score

GPT-5 Nano

Seconds

$0.05/M

Paid

Light tasks with text

GPT-5.6 Luna

3 to 30 s

$0.20/M ref.

5x approx.

General chat and code

Haiku 4.5

2 s in demo

Mid-range

Paid

Fast agents with text

The nuance that almost all summaries omit is in the fine print of the prices. The GPT-5.6 discounted on APIMaster is priced at $0.0178 per million input, less than half of Jev, although it charges output at $0.1067. Since Jev does not generate output tokens, the fair comparison depends on how much prose your task needs and not just on the input price.

**[OpenAI swarms hacking databases in March](https://observatorioblockchain.com/en/cybersecurity/enjambres-openai-hackeando-bases-datos-marzo/)**

## The price under the microscope

The manufacturer's math is eloquent at scale. A thousand tokens cost about $0.00004, so classifying the thirty turns of a long call comes to a tenth of a cent. The company itself illustrates the scale with a robot that runs ten queries per second for about $7 an hour, a figure that must be multiplied by your real traffic before signing anything.

It is worth noting that the price may be subsidized in this early access phase. TypeSafe admits this with unusual frankness and warns that it will need the long term to demonstrate the sustainability of the rate. The 444.6 times cheaper on its page comes from four workflows written by its own team, which the company places at the optimistic end of real gains.

The cost per call also hides the toll of the split architecture. Jev decides but does not write, so any task that needs text also requires paying for a generative model. The total bill of a hybrid flow adds both pieces and the savings are concentrated in the pure decision steps, which are many but not all.

My reading is that the savings exist and are substantial in the right niche. Thousands of daily microdecisions for routing or moderation burn budget on large models that charge dearly for every word generated. Moving those steps to a cheap decider cuts the bill without touching the quality of the final text.

## Speed matters

Response time follows the same order-of-magnitude logic. TypeSafe announces 70 to 500 milliseconds per call against ranges of 3 to 329 seconds in frontier models for equivalent tasks. In the robot demo, Jev posted a median of 233 milliseconds against 1,971 for Haiku 4.5 doing the same decision work.

The difference shows in the product and not just in the metrics dashboard. A classifier that responds in tenths of a second allows validating every turn of a conversation in real time. The same step with a large model introduces perceptible pauses or requires parallelizing with more infrastructure and operational complexity.

Even so, the gains narrow outside its ideal turf. In tasks with heavy chained reasoning, where you must deliberate before deciding, the fast model does not replace the deliberative one. If your current classifier is already well tuned and solves with few calls, the improvement may be modest against the cost of migrating.

The infrastructure comes with generous limits for serious production. The documentation declares 250,000 tokens per second and 1,200 requests per minute on a dynamic basis. These are highway figures designed for massive classification traffic and not for laboratory experiments.

## What Jev cannot do

The catalog of absences decides as many projects as the catalog of virtues. Jev does not extract free fields or write summaries, does not execute multi-step reasoning and does not process images. Any of those needs sends the project back to a classic language model or a combination of both worlds.

Access also imposes its conditions in this early phase. The model lives behind a waiting list as a hosted API, with no open weights or self-deployment in the customer's data center. Organizations with strict data residency requirements will have to wait for an installable offering that does not exist today.

The training method introduces its own unknown. TypeSafe describes a new architecture with a parallel sampler and an algorithm called reinforcement learning for calibrated decisions. No independent laboratory has yet reproduced the results and the architecture remains closed to external scrutiny.

**[Nvidia shields AI agents with Openshell and Sentry](https://observatorioblockchain.com/en/ai/nvidia-seguridad-agentes-openshell-sentry-openai/)**

## Verification pending

The skepticism here is clear. Every relevant figure in the launch comes from the seller's measurements, taken from laptops next to its infrastructure on the US West Coast. The company itself publishes its caveats and asks that they be read, a rare gesture worth appreciating without granting it probative value.

Independent observers confirm the interest and ask for the evidence. Simon Willison published his analysis on September 21 and the next day released a plugin to use Jev from his command-line tool. DigitalOcean added it to its model catalog on September 24 with the manufacturer's figures and the warning that the mandatory third-party tests are still to come.

The comparison with frontier models and the agents the sector expects requires additional caution. Frontier models compete by generating text and reasoning, a turf where Jev does not even play. Measuring them by the same yardstick would be like comparing a scale with a tape measure because both give numbers.

The sector's track record advises waiting six months before passing judgment. Launch prices rarely survive contact with scale and laboratory latencies degrade with real queues. What is solid for now is the concept and the type guarantee, not the record table.

## When each one suits you

The decision framework for your team fits in three questions. How many microdecisions you pay a large model for per day, how much latency your product tolerates and how much prose each step needs. High volume with little prose points to Jev, while open text and reasoning call for a classic model.

The winning architecture for most will be hybrid and without dogmas. A cheap decider that routes and filters, plus a generator that writes only when needed, combines the best of both prices. That pipeline already exists in APIMaster with the identifier jev-latest alongside conventional text models.

We will have to wait for independent tests and price stability to close the debate. And the question that 2027 will answer will remain: whether deciders become a category of their own or just another function of large models. For now, deciding without writing already has an official price and it is four cents per million.

**[Spec driven development and why it changes software](https://observatorioblockchain.com/en/ai/spec-driven-development-que-es-y-por-que-cambia-software/)**
