TypeSafe

TypeSafe is an AI startup that builds "System One Models", which it describes as "a new class of frontier models built to make fast, structured decisions that software can use directly." Its first public model, Jev, returns typed values with calibrated probabilities instead of generated text, and its launch on September 28, 2026 helped define the new category of decision models.

Founder and Origins

TypeSafe was founded by Diogo Almeida, who signs the launch post as "founder, TypeSafe." The company says Jev arrived "after two years in stealth, countless technical challenges, and research breakthroughs." TypeSafe has not disclosed its funding or investors.

Jev and System One Models

The name comes from psychology. TypeSafe says the model class is "inspired by Daniel Kahneman, Thinking, Fast and Slow," drawing on "the distinction between fast, intuitive System 1 thinking and slow, deliberate System 2 reasoning." Where reasoning models spend tokens deliberating, a System One Model answers a well-defined question quickly, choosing among a fixed set of allowed answers. TypeSafe pitches Jev for work that needs to "classify, route, score, extract, or branch," and to "score, judge, verify, guardrail, and detect jailbreaks," as well as for map-reducing over large datasets.

The stack

TypeSafe describes three pieces: "a new model architecture, parallel sampler for maximum efficiency, and training method we call Reinforcement Learning for Calibrated Decisions (RLCD)." Rather than generating text token by token, Jev returns one of the valid answers along with a probability that is trained to be calibrated. TypeSafe says the model "never makes type errors," and notes that its "0%" type-error figure is "not empirical. Schema matching is guaranteed." According to Cloudflare's comparison, Jev has a 32k-token context window.

Speed and pricing claims

TypeSafe says end-to-end response time is "70ms-500ms," which it says can be "40x-200x faster" than frontier LLMs, adding that it expects those numbers to be "on the higher end of real world gains." Pricing is $0.042 per million input tokens ($42 per billion), and output tokens are free, "too cheap to meter" in TypeSafe's words.

Access, Docs and Adapter

On September 28, 2026, TypeSafe opened early access and said it was "bringing developers off the waitlist as quickly as we can." Its documentation at docs.typesafe.ai lists three primitives: Choice, Score and Noul. A Python adapter, system-one-adapter-python, is published on GitHub.

Independent Assessments

Outside analysis so far has been cautious. DataCamp noted that "no large-scale independent reproduction has surfaced yet," that the benchmarks are vendor-reported, and that TypeSafe "can't prove the pricing isn't subsidized." innFactory argued that Jev is best understood as a classifier rather than an LLM and that the principle behind it "is not new"; the novelty is in offering it as a general-purpose, schema-driven API.

Competitors That Followed

Within days, two much larger companies shipped similar products. At DevDay on September 29, 2026, OpenAI reportedly previewed a Decisions API built on GPT-6 Luna, with OpenAI claiming about 150 ms per decision; it has not published a schema or pricing. On October 1, Cloudflare released Clef and Clef-flash as Apache 2.0 open-weight models that are "fully Jev-API compatible." Cloudflare reports median latency of 209.3 ms for Clef and 38.8 ms for Clef-flash against 524.1 ms for Jev in its own tests, and says Clef leads the community Jev Decision Index. Cloudflare's decision to copy Jev's API makes TypeSafe's request format a de facto interface for the category, even as it puts a free, open alternative next to TypeSafe's paid one.

Community Reaction

On Hacker News, much of the response has been developers building their own versions, with projects such as "von" and "Mini-Jev," and a community Decision Index that lists dozens of open reproductions. That reflects enthusiasm and how approachable the idea is. It is not evidence about Jev's quality.

Why It Matters

As agentic systems make thousands of small choices per task, the cost and latency of each choice add up. TypeSafe is betting that those choices should come from a specialized, calibrated model rather than a general-purpose one. The fast follow from OpenAI and Cloudflare suggests the bet is plausible. Whether TypeSafe can stay ahead of them is the open question.