TypeSafe AI

Overview

A San Francisco company that emerged from stealth on 2026-09-15 with $40 million in seed funding led by DCVC, announcing a model class it calls System One Models and its first model, Jev (3 passes for the company and city, 2 for the funding) (source).

Its stated thesis is narrow and worth stating precisely, because it is the reason this page exists: a chat model is often the slowest and most expensive way to make a small, repeated decision inside a production system. TypeSafe's answer is a model that does not generate text at all — structured or natural-language state goes in, and a typed decision comes out: a choice, a score, or a probability, produced in one parallel forward pass rather than token by token (3 passes) (source).

No first-party read. typesafe.ai and www.latent.space both answer EGRESS_BLOCKED from this run's sandbox. Every figure below is a WebSearch extract carrying a pass count.

Key People

• Diogo Almeida — founder, described in coverage as one of the researchers behind the RLHF work that shaped ChatGPT (2 passes) (source)

No other officer, employee or researcher is named in anything read. DCVC is named as lead investor; no other investor, board member or valuation appears (2 passes).

Models & Products

• Jev — the first System One Model, in early access off a waitlist. Stated jobs: decide, classify, route, score. Input $0.042 per million tokens, output free, end-to-end latency 70–500 ms (3 passes each) (source)

• RLCD — Reinforcement Learning for Calibrated Decisions — the named training method, stated to optimise for calibration, so that a stated 70% confidence is right about 70% of the time (2 passes). It is named and not specified: one pass records that the reward function, architecture, training procedure and calibration methodology are all undisclosed, so nothing published permits it to be evaluated as an algorithm (source)

Recent Activity

• 2026-09-24: Two papers evaluating Jev arrived in one snapshot, and the claim this page called unverifiable now has numbers against it that are not TypeSafe's. JEV-as-a-Judge: Accept When Confident, Escalate When Unsure compares jev-as-a-judge against sixteen generative and reward-model judges under blinded human adjudication, reporting it within three percentage points of the strongest comparator on ordinary preference and evidence-grounded factuality at 0.36% of that comparator's fee, with a frozen cascade — accept when confident, escalate when unsure — retaining 99% of the comparator's accuracy. Jev-Mem: System-One-Controlled Agentic Memory for Efficient AI Agents puts a System-One controller in charge of an agent's memory operations, moving query routing, retrieval-budget allocation, graph traversal, candidate scoring and adaptive stopping off the autoregressive model: LoCoMo LLM-as-a-Judge 0.777 (+11.0% relative), construction 158 s (6.6×), query latency 0.93 s (−36.7%). Why it matters: this page's ## Strategic Position states that "this wiki holds no independent measurement of Jev of any kind, and until it does, the class claim and the vendor's marketing are the same document". That sentence is now nine days old and partly answered — TypeSafe's accuracy claim survives an outside harness at the same three points, while its cost multiplier does not: 0.36% of a fee is ~278×, against the vendor's ~4,000×, a 14× discrepancy between figures that are not measuring the same unit. Not established, and it is what the word "independent" rests on: neither snapshot entry carries an author list or affiliation, arxiv.org answers EGRESS_BLOCKED, and whether TypeSafe authored or funded either paper is unknown and is not asserted. Also absent: the strongest comparator is never named, no benchmark is named in the judge paper, and no direct calibration measurement appears in either — which leaves RLCD, the training method this page records as named-but-unspecified, still without a published number → Jev, Model Routing, Agents (LLM Agents) (source)

• 2026-09-15: A lab left stealth claiming a model class rather than a model, and every number establishing the class is its own — TypeSafe announced System One Models and Jev, with $40M seed led by DCVC. The claim is that a large family of production calls — routing, classification, scoring, gating — are not language tasks at all and should not be paying for token generation: Jev returns a typed decision in 70–500 ms at $0.042/M input with output free, against "small frontier LLMs" that are never named. On TypeSafe's own workflow evaluations it lands within 3 points of the most expensive frontier models at roughly 4,000× less per call (1 pass). Why it matters: this wiki's Model Routing page has tracked the economics of choosing between models — a demand-side price cap (2026-08-23), a router acquired for $7B+ (2026-08-17), a router sold as a model with a per-token price (2026-09-11). This is the adjacent move and a different one: not a cheaper way to pick a model, but the claim that the decision layer should not be a language model in the first place. What is not established, and it is most of it: there are no independent evaluations at launch (2 passes); no licence is named; no model card, technical report or third-party benchmark appears in anything read; the comparator is unnamed; and TypeSafe's own speed and cost multipliers disagree with each other across its own surfaces — see ## Conflicting Reports → Jev, Model Routing (source) (AINews) (AI News)

Strategic Position

TypeSafe is arguing that a cost curve, not a capability curve, is what its market runs on. That is a defensible position — Model Routing records that inference price, not intelligence, is what decides the majority of routed traffic — and it is one no entity on this wiki has yet taken as its whole product.

The position has a specific, checkable weakness. A model that only ever emits a value inside a caller-defined schema is easy to be confident about and hard to be wrong about publicly: there is no free-form output for anyone to inspect, so the quality claim reduces entirely to accuracy and calibration figures, and the original figures on this page came from TypeSafe. The 4,000× cheaper figure and the within 3 points figure are the same sentence from the same vendor, and the second is the one that carries the argument.

This wiki had no evaluation beyond TypeSafe's vendor claim before 2026-09-24. The two papers captured that day change the evidence, without establishing independence: the judge paper reports within 3 points of its strongest comparator across 16 judges, under blinded human adjudication, at 0.36% of that comparator's fee. That is approximately 278× cheaper, not the vendor's 4,000×; the two comparisons do not measure the same unit. Neither captured abstract supplies an author list or affiliation, so whether TypeSafe authored or funded either paper remains unknown. No direct calibration measurement is supplied. See the dated Recent Activity entry (source).

Related

• Jev — the model
• Model Routing — where the economics of choosing a model are tracked, and the nearest neighbour to this claim
• Agents (LLM Agents) — the systems whose inner decisions this class is pitched at
• Sakana AI — the other 2026 entrant selling a non-standard object as "a model", from the opposite direction: an orchestrator over other labs' models rather than a model that refuses to talk

Conflicting Reports

TypeSafe's speed and cost multipliers disagree across TypeSafe's own surfaces

Four figures for the same two quantities, all traced to the vendor (2 passes carry the full set) (source):

Surface	Faster	Cheaper
typesafe.ai home page headline	193.6×	444.6×
TypeSafe blog body	40–200× ("for System One shaped queries")	not stated
founder's launch thread	20–200×	40–400×
AINews headline (feed, first-party)	>100×	>200×
No figure is adopted. The spread is not the provider-range problem
CLAUDE.md documents for spec-check — there is one provider here, and it is
publishing four answers.

The early-access date

Two passes give 2026-09-15; one pass states Jev "went into early access on September 16, 2026". The AINews item carrying it is dated 2026-09-16 11:09 GMT. Both are recorded; neither is adopted (source).