Most AI companies right now are building bigger chat models. TypeSafe is building something that doesn’t chat at all. The company, founded by Diogo Almeida, an OpenAI veteran who helped build ChatGPT, has launched its first model, Jev, after two years in stealth. The pitch is simple: structured probabilistic decisions, delivered directly into your codebase, with no text generation involved.
That’s a sharp departure from how most developers currently use AI in production. The standard approach involves prompting a large language model, parsing its string output, wrapping it in guardrails, and hoping the schema holds. Jev skips all of that. It takes an unstructured state as input and returns typed, schema-constrained values in a single parallel query. There’s no autoregressive token generation happening. All outputs arrive simultaneously, which is why the latency numbers are striking: 70 to 500 milliseconds end-to-end, compared to 3 to 329 seconds across typical conversational frontier models.
The training methodology behind Jev is also different. Where models like GPT-4o or Claude train on Reinforcement Learning with Human Feedback, TypeSafe built what it calls Reinforcement Learning for Calibrated Decisions, or RLCD. The goal isn’t to satisfy a human reviewer’s preference for a good answer. It’s to produce confidence scores that actually correlate with output accuracy. For developers building automated branching logic, that distinction matters quite a bit.
On benchmarks run against GPT-6 Astra and Fable 5.1 across multi-step decision workflows, Jev recorded execution speeds up to 193.6 times faster. The internal sampler handles up to 255 discrete options using a two-stage scoring process, which is relevant for anyone managing high-cardinality selection problems at scale.
Pricing also breaks from convention. Input costs run at $0.042 per million tokens, well below the $0.20 to $10 range typical for conversational models. Output tokens aren’t metered at all, because parallel sampling doesn’t generate them in the traditional sense.
TypeSafe tested Jev across a few concrete scenarios worth paying attention to. Running a reactive bot across Doom game states at 10 queries per second cost roughly $7 per hour. A Wikiracing traversal test, where the model navigated dense encyclopedia link structures, completed in fewer steps than non-reasoning models without producing dead ends from hallucinated links.
The target use cases are real-time feature extraction, large-scale data workflows, output verification, and automated decision trees where hand-coded rules tend to break. This isn’t a tool for generating content or answering questions. It’s for teams replacing brittle rule engines with something that can handle probabilistic logic without the overhead of a full language model. Early developer access is open now, with engineering teams being onboarded from a deployment waitlist.
