TypeSafe’s Jev Claims 444.6x Cheaper, 193.6x Faster

TypeSafe AI says its first model, Jev, is 193.6x faster and 444.6x cheaper than LLMs on what it calls System One tasks. Those are the vendor’s own claims. This publication read TypeSafe’s homepage, launch post, docs and evals site on 2 October 2026, ran nothing, and found no independent replication.

What TypeSafe Says Jev Is

TypeSafe describes System One Models as a new class of AI model built for decisions inside software. Jev is its first public one, in early access. It takes a state and typed questions and returns typed answers with probabilities, rather than generated text.

TypeSafe says it built a new architecture, a parallel sampler and a training method it calls Reinforcement Learning for Calibrated Decisions, or RLCD. Its docs say System One models do not write replies, produce code, or explain their reasoning. They answer Choice, Score and Noul questions. In the docs’ table, Noul is the question “Is this statement true?”, answered with a value from 0 to 1.

The Price List

TypeSafe lists Jev at $0.042 per million input tokens, which it also writes as $42 per billion, with output tokens free. For existing LLMs it gives a range of $0.20 to $10 per million input tokens, with output about five times dearer than input.

The homepage says that is a 238x lower input price than Claude Fable 5.1. This publication’s own arithmetic: 238 times $0.042 is about $10 per million tokens, which is the top of the range TypeSafe gives. TypeSafe adds in its blog that it cannot prove the pricing is not subsidized, and that proving it is sustainable will take the long term.

The Speed Claims

TypeSafe says its end-to-end response time is 70ms to 500ms. For frontier LLMs it cites 3 to 329 seconds, linking a third-party benchmarks site. From that it says Jev can be 40x to 200x faster for the same levels of intelligence on System One shaped queries.

It also says its published evals are generally run from its own laptops on the West Coast, where its service is currently based.

What the Headline Measures

TypeSafe’s blog says the 193.6x and 444.6x on its homepage come from its workflow evals, and that it expects them to be on the higher end of real-world gains. The evals use four example workflows: security incidents, agent trace observability, invoice processing and customer service.

The blog says the evals do not use a ground-truth classification. They use the predictions of the largest external models as reference probabilities, specifically the average of GPT-6 Astra and Claude Fable 5.1 at high thinking. So the intelligence measured is agreement with those two models.

The blog adds caveats. The workflows were made by people on TypeSafe’s own capabilities team, so some bias could exist. The reference biases results toward OpenAI’s and Anthropic’s models, and TypeSafe says it likely underestimates its own relative performance and DeepSeek’s.

The LLMs in the evals use TypeSafe’s own wrapper, which constrains them to structured decisions. TypeSafe says that is the most accurate way it has found to get decisions from LLMs, but slower and more expensive than asking without probabilities. Other models run at their providers’ default reasoning settings.

The Homepage’s Own Sample Run

Beside the headline, the homepage shows a single example, marked as proof. TypeSafe AI: cost $0.000081, completed in 0.114 seconds. LLMs: cost $0.013880, completed in 8.566 seconds. The footnote on the headline says it is based on workflows for System One tasks.

This publication’s own arithmetic on that example: $0.013880 divided by $0.000081 is about 171.4 times the cost, and 8.566 divided by 0.114 is about 75.1 times the time. Those differ from the headline multiples. The piece does not say either is wrong. TypeSafe’s blog gives the workflow evals, not this single run, as the source of the headline.

What Zero Hallucinations Covers

The homepage says Zero Hallucinations. The blog says Jev can’t hallucinate, and explains the claim for type errors: schema matching is guaranteed, so TypeSafe adds 0% to its plots. In the blog’s words, its number is not empirical. TypeSafe says a single counter-example would easily falsify the no-type-errors claim, but that a type error is mathematically impossible.

The docs draw the other line. They say calibration is measured across groups of predictions, and that it does not guarantee that an individual answer is correct. The docs also advise asking one narrow question at a time and decomposing anything that needs extended reasoning. The blog’s own demo records one disagreement with the comparison model, on a churn-likelihood level. For the LLM side of its hallucination chart, TypeSafe says the numbers come from OpenRouter, and that there is almost certainly bias there, because more complex queries might be routed to better models.

What Is Not Established

Not established, and therefore absent from this analysis: any independent replication of the speed, cost or accuracy figures; accuracy on any task outside TypeSafe’s four workflows; whether the pricing is sustainable, which TypeSafe says it cannot yet prove; how the full workflow results look beyond the summary chart; and any comment from the LLM providers. The sources read give no customer, revenue or usage figures.

None of the above is investment advice. It reports what a vendor says about its own model and what its pages show.