TypeSafe AI, a startup founded by former OpenAI researcher Diogo Almeida, has released Jev — the first public model in a new architecture class the company calls 'System One Models.' The name references Daniel Kahneman's distinction between fast, intuitive System 1 thinking and slow, deliberate System 2 reasoning. Jev is explicitly not a large language model in the conventional sense: it does not generate strings. Instead, it takes structured program state as input and returns typed probabilistic decisions — classifications, scores, multiple-choice answers, or yes/no judgments — with calibrated confidence scores attached to every output.
### How Jev differs from LLMs
Conventional LLMs are trained with Reinforcement Learning from Human Feedback (RLHF) or Reinforcement Learning with Verifiable Rewards (RLVR) to produce text that human raters prefer. Their output is a sequence of tokens generated one at a time (autoregressive sampling), which makes them flexible — they can chat, write code, hallucinate, or refuse — but slow and expensive for software integration. A frontier LLM call takes 3–329 seconds end-to-end, with output tokens priced roughly 5x higher than input tokens.
Jev inverts this design. It uses a parallel sampler that produces all outputs in a single forward pass, yielding 70–500ms end-to-end latency (40–200x faster for comparable intelligence on structured tasks). Input tokens cost $0.042 per million ($42 per billion); output tokens are free, in the company's phrasing 'too cheap to meter.' The model is trained with a new algorithm, Reinforcement Learning for Calibrated Decisions (RLCD), which optimizes for epistemically honest probabilities on verifiable System One tasks rather than human-preferred text.
The API surface makes the difference concrete. A developer defines questions in advance — Choice (multiple choice), Score (numeric rating), or Noul (yes/no) — each with instructions. Jev returns answers that conform exactly to those types, accompanied by calibrated probability distributions. Because the output schema is fixed, type errors are mathematically impossible. The trade-off is that Jev cannot write code, hold a conversation, or produce any free-form text.
### What 'can't hallucinate' means — and doesn't
TypeSafe claims Jev 'can't hallucinate.' In the narrow sense this is true: the model cannot emit an invalid type or a field that wasn't defined in the question schema. A hallucinated tool call — a common failure mode in LLM agents — is structurally impossible. However, the model can still return a wrong answer with high confidence. Calibration means that when Jev assigns 95% probability to an answer, it is correct roughly 95% of the time across a distribution of tasks. But on any single call, a high-confidence error remains possible. The company acknowledges this: '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.' Jev solves the signaling problem, not the accuracy problem.
### Evaluation methodology and caveats
The article describes Jev's evaluation methodology and practical integration path. The adapter is the concrete tool that lets developers run side-by-side comparisons on structured-decision tasks, making the abstract API discussion tangible.
TypeSafe publishes 'workflow evaluations' that embed Jev and baseline LLMs into identical multi-step code workflows (for example, classifying support tickets, routing, scoring). Rather than comparing against ground-truth labels, they use the average predictions of GPT-6 Astra and Claude Fable 5.1 as reference probabilities. Jev sits on the Pareto frontier — the company cites 193.6x faster and 444.6x cheaper on their homepage — though they note these are likely the high end of real-world gains. Several caveats apply:
- The workflows were authored by TypeSafe's capabilities team, introducing potential bias. - LLMs are accessed through a TypeSafe 'System One LLM' wrapper that constrains them to structured output, which is slower and more expensive than unconstrained generation. - OpenRouter routing may favor stronger models for complex queries, skewing LLM baselines. - Jev's 0% hallucination rate is a structural guarantee (schema matching), not an empirical measurement.
### Intended use cases
Jev targets 'AI-powered workflows' — what the announcement calls 'smart if-statements' — where hand-written logic is too brittle: classifying, routing, scoring, extracting, or branching on fuzzy criteria. Examples include real-time application decisions (100ms latency enables UX-critical paths), map-reduce over large datasets, and guardrailing or verifying LLM outputs (jailbreak detection, reasoning-trace scoring). The company also demonstrates Jev playing Doom via structured game-state queries (~10 queries/second, ~$7/hour) and wikiracing (navigating Wikipedia link graphs), though the latter uses non-reasoning LLM baselines and shows smaller speedups.
### What remains uncertain
- Generalization: The published workflows are a narrow set; performance on unseen task distributions is unknown. - Sustainability: Pricing is not proven unsubsidized; TypeSafe expects costs to decrease but has no long-term track record. - Multimodality: Current demos operate on text-structured state, not images or raw bytes. - Cardinality limits: Jev supports up to 255 choices per question; higher cardinalities use a two-stage scoring-then-choice pipeline with occasional slowdowns.
Early access is open as of the announcement date (September 15, 2026). A GitHub adapter (`system-one-adapter-python`) lets developers compare Jev against OpenAI, Anthropic, and Gemini models on the same structured-decision API.