TypeSafe AI's Jev offers an alternative to LLMs that claims to be 193x faster and 445x cheaper — System One type model is bespoke for probabilistic decision-making

TypeSafe AI's Jev Offers an Alternative to LLMs That Claims to Be 193x Faster and 445x Cheaper

TypeSafe AI emerged from stealth on September 15, 2026 with a $40 million seed round led by DCVC at a reported valuation of approximately $200 million, and launched Jev — the first model in a new category it calls System One.

Jev does not generate text. It takes unstructured input and returns typed, probabilistic decisions in a single parallel pass.

Who Built It

TypeSafe AI was founded by Diogo Almeida, an ex-OpenAI engineer and co-inventor of RLHF, whose work was foundational to InstructGPT and GPT-4. He co-wrote the core training techniques behind ChatGPT. The name Jev is a reference to the Jevons Paradox — the economic observation that efficiency gains often increase rather than decrease total resource consumption. It is, in its own way, a warning embedded in the product name.

What System One Means

The naming comes from Daniel Kahneman's Thinking, Fast and Slow. System 1 is fast, intuitive, automatic judgment — the kind your brain makes without deliberate reasoning. System 2 is slow, effortful analysis. Reasoning models like GPT-5.6 Sol or Claude Opus 5 are System 2. Jev is deliberately System 1.

The architectural difference from standard LLMs is at the foundation. A conventional LLM generates text token by token, autoregressively — each output depends on all the previous outputs, which is why it takes several seconds and why it can produce confidently wrong statements.

Jev skips all of that. It takes unstructured state data and a set of typed questions, then returns every answer in a single parallel forward pass. No text is generated. The output is a typed, probabilistic decision — a Choice, a Score, or a Noul (yes/no probability) — each attached to a calibrated confidence percentage.

The Three Primitives

Everything Jev can return maps onto one of three types.

Choice covers categorical decisions with up to 255 options — route this ticket to billing, support or escalate; classify this transaction as fraud or legitimate. Score covers ordered scales from 2 to 10 levels — urgency, sentiment, quality. Noul is a yes/no probability — is this a billing issue? (0.94 probability: yes).

Multiple questions can be batched into a single request and answered in parallel. The developer receives all answers simultaneously and decides what to do with the confidence values. Jev answers. The developer acts.

What It Cannot Do

The limitations are as important as the claims.

Jev does not chat. It cannot write, explain, summarise, generate code or produce any prose. It has no useful conversational interface. Providing it with extraneous context actively lowers its accuracy — the context window is capped at 32,000 tokens and the model performs best with minimal, directly relevant state.

It still misclassifies. TypeSafe reports approximately 67.8 percent accuracy on its four internal workflow benchmarks, which is close to mid-tier LLMs like GPT-5.6 Terra (68%) but meaningfully below Opus 5 (73%). That five to six percentage point gap is real and matters in high-stakes decisions.

Jev can still be adversarially attacked and can answer the literal wording of a question rather than its meaning. It does not hallucinate in the generative sense — it cannot invent statements — but it can confidently misclassify.

The Numbers

TypeSafe prices Jev at $0.042 per million input tokens. Output tokens are free, described by the company as too cheap to meter. At that price and with its parallel architecture, response times run 70 to 500 milliseconds compared to 3 to 329 seconds for frontier chat models on comparable classification tasks.

The headline comparisons come from TypeSafe's own benchmarks against GPT-6 Astra and Claude Fable 5.1: 193.6x faster and 444.6x cheaper. TypeSafe is explicit that this figure is not empirical — it represents peak performance on internal workflow evaluations, not a published head-to-head study.

The independent data is limited but directionally consistent. Every's head of evals ran 777 judgments in under 0.7 seconds for roughly a quarter of a cent. Every also found Jev approximately 25x faster and 580x cheaper than Claude Fable 5.1 on extraction tasks: 0.35 seconds versus 8.83 seconds per passage.

Arize AI's benchmarks found Jev at 68% accuracy for $0.0004 and 0.4 seconds per case, compared to GPT-5.6 Terra at 68% for $0.03 and 10 seconds, and Opus 5 at 73% for $0.18 and 38 seconds. The cost-benefit tradeoff at 440x cheaper for comparable accuracy is the part that changes architectural decisions.

What It Is Good For

The expected use cases are high-volume, repeated decisions over shared state where the possible answers are already known.

Content moderation at scale. Agent routing in multi-step pipelines. Real-time game AI decisions. Fraud classification. Sensor data interpretation. Any situation where an LLM is being used as an expensive function call to return a structured output that could be expressed as a choice, score or probability.

TypeSafe has demonstrated Jev controlling a Minecraft bot, navigating a simulated drone through obstacles, and playing a Subway Surfers-style game in real time. The Register demo had the model making continuous game-state decisions at sub-100 millisecond latency — where an LLM would time out before the first door opened. These were all simulator environments; real-world robotics performance is unproven.

The Model's Training

Jev is trained with TypeSafe's Reinforcement Learning for Calibrated Decisions, a proprietary variant of the RLHF approach Almeida helped develop at OpenAI. RLCD is designed to optimise for well-calibrated probability outputs rather than for fluent text generation, which is why type errors are structurally impossible rather than trained away — the schema is filled in by construction, so it cannot be violated.

The architecture paper has not been published yet, which means the specific performance numbers should be treated with appropriate scepticism. The directional claim — faster and cheaper at some accuracy cost, without hallucination in the generative sense — is consistent with how non-autoregressive architectures generally behave.

A community .NET SDK appeared four days after launch. Elixir and LangChain integrations exist. The Hacker News thread on launch drew 256 comments validating the concept while pressing hard on the marketing.

Sources: Tom's Hardware, Forkast, DataCamp, Arize AI, TypeSafe AI Blog