The Co-Inventor of ChatGPT Built a New Intelligence That Doesn't Speak Words
Language models became superhuman at conversation years ago. Diogo Almeida, who helped build the instruction-following methods behind ChatGPT at OpenAI, spent four years asking why that skill never translated into widespread automation. His answer is a company called TypeSafe AI and a model called Jev, released September 15, 2026 after two years in stealth.
The premise is that chat models are the wrong tool for decisions buried inside software. A chatbot can afford to be creative, occasionally wrong, and slow. Code that routes a support ticket or scores a transaction cannot. Jev is built for that second job, and it works by giving up the thing that made LLMs famous: the ability to write.
What Jev Actually Does
Almeida describes Jev as a frontier-intelligence function call. Unstructured state goes in, typed probabilistic decisions come out. Instead of producing a stream of text that software then has to parse and validate, Jev returns values whose shape is defined in advance. The possible outputs are declared up front, and the model fills them in with answers plus confidence scores.
The company reports that Jev matches existing LLMs on what it calls System One tasks while running two orders of magnitude faster and cheaper. The name comes from Daniel Kahneman's distinction in Thinking, Fast and Slow between fast, intuitive System 1 thinking and slow, deliberate System 2 reasoning. The model is named after William Stanley Jevons, the economist whose observation that efficiency gains in coal use increased total demand. TypeSafe expects the same pattern for intelligence: each drop in cost unlocks more uses.
The training method behind it is called Reinforcement Learning for Calibrated Decisions. Where standard LLMs optimize for human preference or verifiable correctness, TypeSafe says it optimizes for calibration, meaning the probabilities the model reports are meant to track its actual accuracy. If a model can do a task correctly 95 percent of the time but cannot flag which cases fall in the other 5 percent, it cannot safely automate that task. Calibration is the attempt to fix that.
The Numbers TypeSafe Puts Forward
The company divides its claims into ones a user can verify immediately and ones that require more nuance.
On the verifiable side, Jev prices input tokens at $0.042 per million and does not meter output tokens at all. Frontier LLMs, by comparison, charge between $0.20 and $10 per million input tokens, with output tokens running roughly five times higher than input. TypeSafe reports end-to-end response times of 70 to 500 milliseconds against 3 to 329 seconds for frontier models, a range it translates to 40 to 200 times faster for comparable tasks.
The type-safety claim is the strongest stated. Because Jev's outputs are constrained to a schema declared in advance, the company says type errors are mathematically impossible rather than merely rare. A single counterexample would falsify it, and TypeSafe argues none can exist. On hallucination, it adds a flat zero to its charts on the same grounds, since schema matching is guaranteed.
How the Evaluations Work
TypeSafe built a new kind of benchmark to test how models perform inside code rather than in isolation. The setup assumes there is a correct compute graph, a workflow represented in code, and uses the predictions of the largest and most expensive external models as reference probabilities. Every model runs the same workflow, and each is scored against the average of two frontier models, identified as GPT-6 Astra and Fable 5.1.
On these workflow evals, TypeSafe reports that Jev owns the Pareto frontier by almost two orders of magnitude. The headline figures of 193.6 times faster and 444.6 times cheaper come from this test. The company is candid that the workflows were built by its own model capabilities team, which introduces possible bias, and that using OpenAI and Anthropic models as the reference answer likely biases scoring toward those models and understates Jev's relative performance.
A side-by-side demo shows the sampling difference directly. Jev produces all its probabilities in parallel in a single query, while an LLM generates token by token. TypeSafe notes that the demo query was simplified, with short and dense input state and readable output keys, and acknowledges this framing favors its model. In the recorded run against GPT-5.6 Terra, the only disagreement was on churn likelihood, which the company says looked genuinely ambiguous.
The Demos
Two demos show the model reacting in real time. In one, Jev plays Doom from a structured text representation of the game state, running about ten queries per second at a cost near $7 an hour. TypeSafe concedes a non-AI bot would play better, but the point was a bot reacting to different representations of game state while following instructions.
In the second, Jev plays Wikiracing, starting on one Wikipedia page and reaching a target using only links it encounters. Each step can mean choosing among hundreds or thousands of links, which the company uses to show the benefit of not hallucinating when the number of choices is high. Here the speedups are smaller, because the LLMs were tested in non-reasoning modes to keep the demo watchable. Jev tended to finish in fewer steps. The model supports a cardinality up to 255, and for larger choice sets it runs a two-stage process of scoring options independently and then choosing, which explains occasional slowdowns.
Where the Claims Need Scrutiny
TypeSafe presents its own caveats throughout, which is worth noting because the strongest claims rest on assumptions a reader should see clearly. The speed and cost figures come from workflows the company designed and ran, not from independent testing. The hallucination numbers for competing LLMs come from OpenRouter traffic, where more complex queries may be routed to stronger models, introducing bias the company flags but cannot control. The pricing may or may not be subsidized, and TypeSafe says only time will prove its sustainability.
The type-safety guarantee is the cleanest claim, since it follows from schema constraints rather than from measured performance. Whether that guarantee matters as much as raw intelligence depends on the use case. TypeSafe's argument is that a hallucinated tool call is a nuisance in a chatbot and a failure in a system with latency guarantees or deep dependency chains.
What It Is For
The intended uses are decisions inside ordinary software: classify, route, score, extract, or branch where hand-written rules are too brittle. TypeSafe frames these as smart if-statements, where the surrounding code constrains the model's freedom and makes it easier to compose into reliable systems. Other stated uses include running over large datasets to produce features, real-time applications where response time matters, and verifying or guardrailing the outputs of other LLMs.
Jev gives up string generation entirely, so it is not a replacement for chatbots, coding agents, or anything that needs to produce readable prose. It is a component meant to sit inside software and make fast, bounded decisions with a confidence number attached to each one.
The early access program is now open, and TypeSafe says it is pulling developers off its waitlist. The company's stated goal is to learn where Jev works and where it fails, which suggests the results outside its own benchmarks remain to be established.
