Blog: Introducing System One Models & Jev (2026-09-15)
TL;DR The launch post by Diogo Almeida announces System One models and Jev. Thesis: chat-shaped models are the wrong interface for automation, so TypeSafe built a new architecture, a parallel sampler, and RLCD training to return typed decisions with calibrated probabilities. It is unusually explicit about its own caveats — read the "Nuance" bullets before quoting any number.
Facts
| Field | Value |
|---|---|
| Title | "Introducing System One Models & Jev" |
| URL | https://typesafe.ai/blog/introducing-system-one-models-and-jev |
| Category | Company News |
| Date | 2026-09-15 (page shows "Sep 15, 2026") |
| Author | "Diogo Almeida, founder, TypeSafe" |
Thesis
Opening question: "Models have been superhuman at chat for years, so where is all the automation?" Almeida says this has been his "driving question for the last four years," that at OpenAI he "helped build the methods that made language models useful at following instructions and talking with people," and that "despite the hype it became obvious to me that there was something really big missing."
The announcement: "After two years in stealth… TypeSafe AI is releasing our first System One Model: a new class of frontier models built to make fast, structured decisions that software can use directly." The stack is described as new end to end — "a new model architecture, parallel sampler for maximum efficiency, and training method we call Reinforcement Learning for Calibrated Decisions (RLCD)."
The product claim: "Jev achieves similar levels of intelligence on System One tasks compared to existing LLMs, while being two orders of magnitude faster and more efficient. While Jev gives up string generation, it's optimized for structured outputs and can't hallucinate." And the mental model: "Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out."
Comparison table: "Frontiers, Old and New"
Reproduced from the post. Everything here is TypeSafe's characterization of both columns.
| Dimension | Existing LLMs | System One + Jev |
|---|---|---|
| Optimized with | RLHF / RLVR | RLCD |
| Optimizes for | Human preference; verifiable rewards | "Calibrated decisions: answers with epistemically honest probabilities on System One tasks" |
| Inputs | Unstructured data with an emphasis on sequential messages | Structured program state |
| Outputs | Strings / generated text; must be parsed + validated | Type-safe structured values, "defined in advance"; "The model never makes type errors"; all answers carry calibrated probabilities and confidence |
| Sampling | Sequential, one token at a time | Parallel: "Generates all outputs in a single query" |
| Cost | Input $0.20–$10 / MTok; output ~5x input | Input $0.042 / MTok ($42 per billion); output "FREE (too cheap to meter)" |
| Speed | End-to-end 3 to 329 seconds | End-to-end 70 ms – 500 ms; "40x-200x faster… for System One shaped queries" |
| Confidence | "models tend to be overconfident and inconsistent" | "Always communicates confidence and uncertainty… Calibrated: higher confidence means higher accuracy" |
| Use cases | Human-in-the-loop tasks; verifiable problems; demos | AI-powered workflows / "smart if-statements"; map-reducing over big data; real-time applications; "verify everything" (score, judge, guardrail, detect jailbreaks) |
Note the internal tension on speed: the table says 40x–200x for the demo shape, while the homepage advertises 193.6x faster / 444.6x cheaper from the workflow evals.
Evidence section
The post splits claims into "easily verifiable" and "bolder."
Easily verifiable, per the post:
- Speed per call — "We truly are that fast, though our published evals are generally run from our laptops on the West Coast (this is where our service is currently based)."
- Cost per call — "We make our pricing transparent. We can't prove it isn't subsidized; we'll need the long-term to prove the sustainability of our pricing (which we expect to go down, not up)."
- No type errors — "This would be an easy thing to falsify with just a single counter-example, but it is mathematically impossible."
Side-by-side demonstration. Jev "outputs all probabilities in parallel instead of autoregressively generating by token." The comparison model is "GPT-5.6 Terra with default reasoning… because we've found it to be the most comparable at intelligence to Jev on average." For early-access users the post links "the actual query" as a console playground share link.
Workflow evals. The methodology: "we assume there is a correct compute graph (a 'workflow' represented in code) and use the predictions of the largest, smartest, and most expensive external models as reference probabilities… every model gets the same workflow. We test how they compare to the average of the smartest models (in this case, Astra and Fable)." Result claimed: "Jev is off the charts – owning the Pareto frontier for almost 2 orders of magnitude." Four workflows are published, and this section is the stated origin of the homepage numbers: "This is where the claims of 193.6x faster, 444.6x cheaper on our home page comes from, and we expect that these are on the higher end of real world gains." Details: Workflow evals: how TypeSafe measures Jev.
Hallucination and type-safety. "Hallucination and type-safety are intrinsically related, and we think the latter is table stakes for automation… Existing models, no matter how smart, still hallucinate and have type errors."
Fun demos. Two: a Doom bot ("real-time intelligence and what can be doone with code + AI"; the engineer "was worried about making 10 queries a second (which ends up costing ~$7/hour)"), and Wikiracing ("Each step can mean choosing between hundreds to thousands of links!"). The Doom demo is the hook The Register used; see Press and third-party coverage.
Caveats the post states itself
These are the post's own "Nuance" bullets — quote them alongside any number taken from it.
| Claim | Caveat, as written |
|---|---|
| Side-by-side demo | "The query is highly simplified and questions were chosen to have descriptive, human-readable keys"; the state is "short, dense, and detailed… The relatively shorter input paints our model in an advantageous light" |
| Side-by-side demo | "the only disagreement with GPT-5.6 Terra is on 'Churn likelihood level'. The actual answer seems genuinely ambiguous to us" |
| Workflow evals | "These content of these workflows were not deliberately chosen nor constructed to make our model look good, and are not in our training distribution. However, they were made by individuals on our model capabilities team, so some bias could exist" |
| Workflow evals | "We use the average of GPT-6 Astra and Fable 5.1 as the reference answer, which biases answers towards OpenAI and Anthropic's models. We likely underestimate the relative performance of our model and DeepSeek's models" |
| Workflow evals | LLMs are run through "our System One LLM wrapper… this tends to be slower and more expensive than giving decisions without probabilities" |
| Homepage multipliers | "we expect that these are on the higher end of real world gains" |
| Pricing | "We can't prove it isn't subsidized" |
| Hallucination plot | "The numbers for LLMs are from OpenRouter i.e., there almost certainly is bias here"; "Our number is not empirical. Schema matching is guaranteed, thus we can confidently add 0% into the plots" |
| Speed measurements | "our published evals are generally run from our laptops on the West Coast" |
| Wikiracing | Run "against the non-reasoning modes of the models (except Astra which was set to the lowest reasoning setting)… The LLMs look much worse at this task than with reasoning enabled" |
| Cardinality | "Jev supports a cardinality up to 255. For the higher cardinality choices, we do a 2 stage-system of scoring independently then making an explicit choice, hence the occassional slowdown" |
Names and FAQ
On naming: "We were inspired by Daniel Kahneman, Thinking, Fast and Slow. The model class name draws on the distinction between fast, intuitive System 1 thinking and slow, deliberate System 2 reasoning," with the acknowledgement that "'System 1 thinking' has also implied error-prone." And: "We named Jev after William Stanley Jevons. We expect machine intelligence to follow a similar path to coal, after steam-engine efficiency led to an increase in demand. Every order of magnitude drop in the cost of intelligence unlocks orders of magnitude more use cases."
The post's FAQ headings (answers collapsed in the capture, except the naming one): "Why was a new training algorithm needed?", "What use cases is Jev good for?", "Is Jev just a smaller LLM?", "How does Jev perform against public benchmarks?", "Where does our training data come from?", "These are results are kinda crazy - how is it possible?"
Links the post references
- Console playground share link for the side-by-side demo's "actual query" (early access only) — console.typesafe.ai (console + playground)
- "our workflow evals site" → https://evals.typesafe.ai/ — Workflow evals: how TypeSafe measures Jev
- "System One LLM" wrapper → the adapter repo, https://github.com/typesafe-ai/system-one-adapter-python — system-one-adapter: LLM-backed drop-in for TypeSafeClient
- "early access" → the typesafe.ai waitlist
Related
- Jev (model) — the model as documented
- System One Models — the model class
- Workflow evals: how TypeSafe measures Jev — the eval methodology and numbers
- Manifesto: Composable AI — Build Prod, Not God — the argument this post productizes
- Blog: Lies, Damned Lies, and Benchmarks — why there is no benchmark table here
- Press and third-party coverage — how third parties read it
Sources
- raw/site/blog-introducing-system-one.txt (https://typesafe.ai/blog/introducing-system-one-models-and-jev)