What Is TypeSafe Jev? System One Models Explained (2026)
The short answer
Jev is a new kind of AI model that does not generate text. TypeSafe AI, founded by former OpenAI researcher Diogo Almeida after two years in stealth, announced it on September 15, 2026 as the first public System One Model: you hand it unstructured state and a schema of typed questions, and it returns every answer at once as type-safe values with calibrated probabilities, in 70–500 ms, at $0.042 per million input tokens with output free. It is in early access with a waitlist.
The pitch is a “frontier-intelligence function call”: the intelligence of a large model, delivered as something software can depend on — no parsing, no validation, no hallucinated tool calls, and a confidence number your code can branch on.
Why a former ChatGPT researcher built it
Almeida’s framing: models have been superhuman at chat for years, “so where is all the automation?” His answer is that strings are the wrong interface for software. Text outputs must be parsed and validated, can go off the rails, arrive one token at a time, and cost 5x more on output than input. Chatbots and coding agents tolerate that because a human is in the loop; a decision buried several layers deep in a dependency chain with latency guarantees cannot.
TypeSafe built a new stack around that constraint: a new model architecture, a parallel sampler that emits all outputs in a single query, and a training method it calls Reinforcement Learning for Calibrated Decisions (RLCD) — optimising for epistemically honest probabilities rather than human preference (RLHF) or verifiable rewards (RLVR).
LLM vs System One Model, per TypeSafe
| Existing LLMs | System One + Jev | |
|---|---|---|
| Optimised with | RLHF / RLVR | RLCD (calibrated decisions) |
| Optimises for | Human-preferred writeups; verifiable outputs | Honest probabilities on decision tasks |
| Inputs | Text, emphasis on sequential messages | Text, emphasis on structured program state |
| Outputs | Strings — anything, incl. hallucinations; must be parsed/validated | Type-safe structured values defined in advance; no type errors |
| Sampling | Sequential, token by token | Parallel, all outputs in one query |
| Cost | $0.20–$10 / MTok in; output ~5x | $0.042 / MTok in; output free |
| Latency | 3–329 s end-to-end for frontier models | 70–500 ms |
| Confidence | Overconfident, inconsistent when asked | Calibrated probability on every output |
| Use cases | Chatbots, copilots, coding agents, verifiable problems, demos | AI-powered workflows, map-reduce over big data, real-time apps, guardrails |
Evidence TypeSafe published — and its own caveats
TypeSafe is unusually explicit that “extraordinary claims require extraordinary evidence” and lists what is and is not verifiable:
- Speed and cost per call are directly checkable; TypeSafe notes its published evals ran from laptops on the US West Coast, where its service is hosted, and that it cannot yet prove pricing is not subsidised.
- Zero type errors is guaranteed by construction (schema matching), so the 0% figure is not empirical.
- Workflow evals (evals.typesafe.ai) hold the compute graph fixed and score every model against the average of GPT-6 Astra and Claude Fable 5.1 as reference probabilities. Jev “owns the Pareto frontier for almost 2 orders of magnitude”; this is where the homepage’s 193.6x faster, 444.6x cheaper figures come from, which TypeSafe says are at the high end of real-world gains. Caveats it lists: workflows were written by its own capabilities team, the Astra/Fable reference biases toward OpenAI and Anthropic, and the LLM baselines run through TypeSafe’s open-source System One LLM adapter wrapper, which is slower and pricier than unconstrained prompting.
- Side-by-side demo against GPT-5.6 Terra (default reasoning), chosen as the closest LLM to Jev’s intelligence on average; the only disagreement in the recorded run was a genuinely ambiguous “churn likelihood” call.
- Hallucination comparison uses OpenRouter figures for LLMs, which TypeSafe flags as biased because harder queries may route to better models.
Two demos: a Doom bot driven by Jev at 10 queries per second (~$7/hour) over a structured text state, and Wikiracing, where Jev’s freedom from hallucination compounds across high-cardinality link choices; against non-reasoning LLM modes its speedups were smaller.
Where Jev fits an AI stack in 2026
Jev is not a competitor to GPT-6 Astra, Claude Fable 5.1 or Gemini 3.8 for generation. It competes with three things: hand-written rules and classifiers, small fine-tuned models used for routing and extraction, and LLM calls used as glorified if-statements — the “is this ticket urgent?”, “which queue?”, “does this output violate policy?” calls that agent frameworks scatter everywhere. Its calibrated confidence is the feature that matters for automation: a model that is right 95% of the time but cannot tell you when it is in the 5% cannot be trusted to act alone; a model that flags the 5% can escalate.
The obvious open questions: how well calibration holds off-distribution, whether $0.042/MTok survives scale, and whether structured-state inputs generalise beyond the demos (images are “not yet” supported). TypeSafe’s naming — after William Stanley Jevons, of the Jevons paradox — signals the bet: every order-of-magnitude drop in the cost of a decision unlocks orders of magnitude more decisions.