A new kind of AI model from a ChatGPT inventor is thrilling developers

Now I have solid facts. Let me write the article directly.

TypeSafe AI, a startup founded by Diogo Almeida, a former OpenAI researcher who contributed to the development of ChatGPT through his work on RLHF and InstructGPT, released Jev on September 15, 2026. Jev is not a chatbot or a text generator. It is a decision-making model built for software automation, designed to accept a state and a set of predefined options and return the most likely choice along with a confidence score. It operates in 70 to 500 milliseconds, far faster than typical large language model calls, and at a fraction of the cost.

This article breaks down what Jev does differently from standard LLMs, how its architecture produces structured decisions instead of generated text, and why developers building agentic workflows and classification systems should pay attention. It is written for software engineers, technical leads, and product teams evaluating AI tooling for programmatic tasks where reliability and speed outweigh conversational fluency.

A new kind of AI model from a ChatGPT: what actually changed

Jev replaces the text-generation loop that defines models like ChatGPT and Claude with a constrained decision interface. Where a conventional LLM receives a prompt and emits tokens one at a time until it hits a stop condition, Jev receives a structured state object and a finite set of options, then returns a single selection with an attached confidence score. There is no prompt engineering involved, no JSON parsing on the output side, and no risk of the model inventing a response format that breaks downstream code.

TypeSafe AI describes Jev as a "System One" model, borrowing the cognitive-science term for fast, automatic judgment. The practical consequence is that Jev completes calls in 70 to 500 milliseconds. That latency range puts it closer to a database lookup than to a typical LLM inference call. For tasks like intent routing, feature-flag evaluation, or multi-step agent orchestration, the speed difference compounds across every decision point in a pipeline.

A new kind of AI model from a ChatGPT: how it works

A developer integrating Jev defines a state, which is a structured representation of the current context, and a list of candidate actions or labels. Jev evaluates the state against those candidates and returns the top-ranked option along with a probability distribution. Because the output schema is fixed to the options the developer supplied, the response is always valid, typed, and machine-readable without additional validation code. This eliminates the common failure mode where an LLM returns malformed JSON or hallucinates a field name.

The model does not perform open-ended text generation at any point. It is trained specifically to map structured inputs to discrete choices, which means the compute cost per call is substantially lower than running a general-purpose language model for the same routing or classification task. Developers do not need to maintain prompt templates, retry logic for malformed outputs, or output parsers, all of which are standard overhead when using LLMs for programmatic decisions.

A new kind of AI model from a ChatGPT: why it matters now

Agentic AI systems, where software makes chains of autonomous decisions, have become a primary focus for development teams across the industry. The reliability problem with using LLMs in those chains is well documented: each step that depends on parsed text output introduces a failure surface. A single malformed response can cascade through a pipeline. Jev addresses this by guaranteeing that every response conforms to the caller's predefined schema, which removes an entire class of runtime errors from agent architectures.

Cost is the other pressure point. Running a large language model for every classification or routing decision inside an application adds up quickly at scale. Because Jev is purpose-built for decision tasks rather than general text generation, its per-call cost is lower and its throughput is higher. For teams already running LLM-based automation, swapping decision steps to Jev could reduce both latency and inference spending without changing the surrounding application logic.

A new kind of AI model from a ChatGPT: who is affected

The most immediate audience is backend and platform engineers building agentic workflows, classification pipelines, or multi-step automation systems. Anyone currently using an LLM to pick from a known set of options, route a request, or label an input stands to benefit from Jev's constrained-output approach. Teams that have invested engineering time in output validation, retry wrappers, and JSON-repair utilities for LLM responses are the ones most likely to see direct savings in both code complexity and runtime cost.

Startups and smaller teams that lack the infrastructure budget for high-volume LLM calls at scale are another natural fit. Jev's lower per-call cost and sub-second latency open up use cases where calling a general-purpose model was previously too expensive or too slow. Product teams that avoided embedding AI decisions in latency-sensitive paths now have a model that fits within the performance envelope of a typical API request.

A new kind of AI model from a ChatGPT: what to watch

Jev's value proposition depends on how broadly the decision-model paradigm applies. Routing, classification, and option selection are clear fits, but many real-world tasks require a blend of generation and decision-making. How TypeSafe AI positions Jev alongside conventional LLMs, whether as a complement or a replacement for specific pipeline stages, will determine adoption patterns. Developers should evaluate whether their workloads decompose cleanly into decision steps that Jev can handle and generation steps that still require a traditional model.

Ecosystem integration is the other variable. Jev's usefulness scales with how easily it plugs into existing frameworks, orchestration libraries, and deployment pipelines. Diogo Almeida's background at OpenAI lends credibility, but the model's long-term traction will depend on SDK quality, documentation depth, and whether TypeSafe AI publishes benchmarks that let teams compare Jev's accuracy and cost against LLM-based alternatives on standardized tasks. Early adopters should test Jev on their own routing and classification workloads and measure latency, accuracy, and cost per call against their current LLM setup.