A Researcher Behind ChatGPT Built Jev, a Model That Gives Up Strings Entirely

TypeSafe AI released Jev in early access on September 15, 2026, a model that returns typed decisions with calibrated probabilities instead of text. Founder Diogo Almeida says he helped build the methods at OpenAI that made language models good at following instructions and talking with people, work that became the research behind ChatGPT, and the company spent two years in stealth before this launch. Jev prices input at $0.042 per million tokens with output tokens free, and the company puts end-to-end response time between 70 and 500 milliseconds. ASAP looks first at where this launch moves the bottleneck in automation rather than at the performance numbers.

Giving up strings is what buys the elimination of type errors

The defining property of the System One model class TypeSafe AI describes is that the output shape is fixed in advance. Possible outputs and their structure are defined ahead of time, so the model never makes a type error, and every answer arrives with calibrated probabilities and confidence scores. The company frames Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.

The company lays out the contrast with existing LLMs in a table. Existing models are optimized with RLHF for human preference or RLVR for verifiable rewards, and their output is a string. Strings can be anything, including chat responses, code, hallucinations, refusals, and even type-safe structured values, so software has to parse and validate them, and some risk of the model going off the rails always remains. Jev is trained with RLCD, Reinforcement Learning for Calibrated Decisions, and emits only predefined type-safe structured values.

The sampling method splits the two as well. Existing models generate one token at a time, each conditioned on the last. Jev generates all outputs in a single query in parallel. The company presents that parallel sampler as the main source of the speed and cost gap, and argues that giving up strings is precisely what grants the other capabilities.

The stated limits sit in the same place. Jev supports cardinality up to 255, and for larger choice sets it uses a two-stage system that scores options independently and then makes an explicit choice. The company attributes the occasional slowdown in its Wikiracing demo to that structure.

What blocked automation was not accuracy but the inability to say "I don't know"

The sharpest argument in this September 15, 2026 launch is about confidence, and it is the claim that separates Jev's design from every RLHF-trained model. The company notes that models are overconfident and inconsistent even when prompted for a confidence estimate, and writes that 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.

That sentence targets software structure rather than model intelligence. Putting a component with 95% accuracy into production does not require eliminating the remaining 5%; it requires a path that identifies those cases and routes them to a human. With calibrated probabilities, a developer sets thresholds in code for when the system acts autonomously and when it escalates for review. With decisions but no probabilities, the raw material for that branch does not exist.

The same reasoning binds hallucination and type safety into one problem. The company treats the latter as table stakes for automation, writing that a hallucinated tool call is inconvenient inside an agent but a deal-breaker in a system with latency guarantees or buried several layers deep in a dependency chain. A strange output is visible when a person reads the result, but in the middle of an automated pipeline a wrong value flows quietly downstream.

The company published the caveats on its own 444.6x headline

The headline figures on the TypeSafe AI home page are 193.6x faster and 444.6x cheaper. What makes this launch unusual is that the company documented the provenance and limits of those figures itself, and copying the headline without the caveats loses roughly half the information.

Four caveats are stated. First, the numbers come from workflow evaluations the company built, and it expects them to be at the higher end of real-world gains. Second, the workflow content was not deliberately chosen or constructed to flatter the model and is not in the training distribution, but it was made by people on the company's own model capabilities team, so some bias could exist. Third, the reference answer is the average of GPT-6 Astra and Fable 5.1, which biases results toward OpenAI and Anthropic models and, by the company's own reckoning, likely underestimates the relative performance of its own model and DeepSeek's. Fourth, the comparison LLMs run through the company's System One LLM wrapper, which it calls the most accurate way to get decisions from an LLM while noting that it is slower and more expensive.

The gap between the published numbers also deserves attention. In the side-by-side demo, the TypeSafe side shows $0.000081 and 0.114 seconds against $0.013880 and 8.566 seconds for the LLM. Dividing those pairs yields roughly 75x faster and roughly 171x cheaper, some distance from the headline 193.6x and 444.6x. The company explains that calls in the workflow evaluations are significantly more complex than the demo and more representative of real business automation loads. The two sets measure different work, and what a prospective adopter should expect depends on the complexity of their own workflows.

The 0% hallucination figure reads the same way. The company states plainly that this number is not empirical: schema matching is guaranteed, so it can confidently plot 0%. A guarantee that the shape is never wrong is a different claim from a guarantee that the content is right, and the company drew that line itself.

What the free-output price tag is actually aiming at

The most striking line item in the pricing is that output tokens are free. The company charges $0.042 per million input tokens and calls output too cheap to meter, while placing existing models at $0.20 to $10 per million input tokens with output roughly five times input. The home page states an input price 238x lower than Claude Fable 5.1.

Making output free follows from the architecture. When the output is a predefined structured value plus probabilities, its length is small relative to the input and bounded in advance. Output tokens are expensive in string-generating models because sequential generation is itself compute and latency, and that cost structure does not hold for a model that emits everything in parallel in one pass. The price sheet is a shadow of the architecture rather than a marketing decision.

The naming points the same direction. The model class draws on Daniel Kahneman's fast, intuitive System 1 thinking, and Jev is named after William Stanley Jevons. The company expects machine intelligence to follow coal after steam-engine efficiency raised demand, writing that every order-of-magnitude drop in the cost of intelligence unlocks orders of magnitude more use cases. Its Doom demo, running at roughly 10 queries a second for about $7 an hour, is offered as an example.

Sustainability remains unproven. The company says its pricing is transparent but that it cannot prove the prices are not subsidized and will need the long term to demonstrate otherwise, adding only that it expects prices to go down rather than up. That is reason enough not to treat an early-access price sheet as a fixed variable in an adoption decision.

The measurable question for a team today is the unit cost of classification and routing

The workloads this model targets are specific. The company lists classification, routing, scoring, extraction, and branching where hand-written logic is too brittle, map-reduce over large data to produce features and insights, real-time applications where 100-millisecond latency matters for UX, and scoring, judging, verifying, guardrailing, and jailbreak detection over LLM prompts, reasoning traces, and outputs.

The closest matches in most product organizations are support ticket triage, product catalog normalization, ad creative review, comment and review moderation, and payment anomaly detection, all of them high-volume work with short per-item judgments. Cost there is unit price times volume, so a two-order-of-magnitude drop in unit price changes the boundary of what is worth automating at all. Work that stayed manual for economic rather than technical reasons is exactly what moves.

The same definition marks what this model cannot replace. Drafting documents, handling conversations, and generating code all require strings as the deliverable and fall outside its scope. The company itself lists human-in-the-loop work, verifiable problems like math proofs and kernel optimization where correctness can be checked cheaply, and prototypes as the domain of existing LLMs. Splitting the two inside one system is the realistic adoption shape: LLMs where judgment is open-ended, Jev where the same decision repeats.

What remains unanswered

Four gaps in the TypeSafe AI launch are what bound how far its claims can be carried. First, parameter count and the concrete shape of the architecture are undisclosed, described only as a new architecture and a new sampler. Second, no public benchmark scores are presented, and the evaluation rests on four workflows the company built. Third, the origin of the training data appears only as an FAQ entry and is not explained in the body. Fourth, availability is early access, and the rate at which the waitlist is being cleared is not stated.

That narrows the useful validation method to one thing. A team should first check whether the decisions in its own work genuinely resolve to types, and whether attaching probabilities to those decisions makes it possible to set a threshold separating autonomous execution from human review. Where those two conditions fail, no multiple of speed or price matters. Where they hold, the comparison to run is not against another model's API price but against the cost of the people and rules engines making those calls today.

Source: Introducing System One Models & Jev (TypeSafe AI, Diogo Almeida, September 15, 2026)