If your software needs fast, repeated machine-to-machine decisions rather than a chatbot, TypeSafe's Jev is worth a look: it takes up to 64k tokens of text per request and bills input only. Its Vanta-run trust portal publishes SOC 2 Type II status, though API access is still waitlisted.
TypeSafe AI is a San Francisco startup, founded in 2024, whose CEO and co-founder is Diogo Almeida, a former OpenAI researcher. It sells one product, the Jev decision model, which returns typed answers with calibrated confidence in roughly 0.4 seconds per case on its own benchmark, for software automation rather than chat.
Founded: 2024 · HQ: San Francisco, CA, USA · Team: 10-25 · CEO: Diogo Almeida · Funding: $40M seed (announced Sep 2026, lead: DCVC)
About TypeSafe AI
TypeSafe AI Inc. is a San Francisco AI lab founded in 2024 by Diogo Almeida (CEO), Erik Gafni (CTO), and Sasha Sheng (COO). Almeida spent four years at OpenAI working on post-training for ChatGPT, is credited as a co-inventor of RLHF, and is a listed author on the InstructGPT paper and the GPT-4 technical report; he previously worked at Google Brain, part of Google. Gafni is a repeat founder who built Ravel, a machine learning company for DNA sequencing, and was an early engineer at Invitae and Freenome. Sheng was a research engineer at Meta's FAIR lab, working on News Feed and AI Experiences, with papers at NeurIPS and ECCV. The company worked in stealth for roughly two years before its public launch, operating from an office near San Francisco's Embarcadero station with a stated five-day in-person policy and staff drawn from OpenAI, Google Brain, Meta FAIR, Stripe, Airbnb, Plaid, and Docker.
TypeSafe's first release is Jev, what the company calls a System One Model: a non-autoregressive model that returns typed decisions and calibrated confidence scores in a single pass instead of generating text token by token. It is trained with a technique TypeSafe calls Reinforcement Learning for Calibrated Decisions (RLCD), which the company says optimizes for calibration, so a decision Jev states as 70 percent confident should be correct about 70 percent of the time, rather than for pleasing a human rater the way RLHF does. On TypeSafe's own four-workflow benchmark, Jev scored 67.8 percent accuracy, close to GPT-5.6 Terra's 67.9 percent and a few points behind GPT-5.6 Sol (74.1 percent) and Claude Opus 5 (73.1 percent), while running in roughly 0.4 seconds per case against 10 to 38 seconds for those models, and at a claimed cost near $0.0004 per case against $0.0304 to $0.1761 for them. Jev is priced at $0.042 per million input tokens with no charge for output tokens, and is reachable only through a waitlisted early access program at console.typesafe.ai.
TypeSafe AI raised a $40 million seed round, announced alongside its stealth exit, led by deep-tech investor DCVC. No other investors were named publicly, and the company has not disclosed a valuation.
TypeSafe positions Jev against general-purpose frontier chat models such as OpenAI's GPT-5.6 family and Anthropic's Claude Opus 5, for a narrower job: fast, repeated, machine-to-machine decisions embedded inside production software such as games, robotics, simulations, and automation pipelines, where a fixed, typed output matters more than open-ended reasoning or conversation. Public demos shown at launch include a Minecraft-playing bot, a Doom-playing agent, a self-driving-style simulation, and a drone navigating an obstacle course. Almeida has argued publicly since leaving OpenAI that RLHF-trained chat models are tuned to satisfy the human rater in the training loop, which he says makes them a poor fit for unsupervised, high-volume automation, the gap TypeSafe is trying to fill.
On security, TypeSafe runs a Vanta-hosted trust center at trust.typesafe.ai. When HokAI read it on 8 October 2026, its compliance panel showed SOC 2, its resources included a document titled "SOC 2 Type II - 2026", and its control list covered infrastructure, organizational, product and data security, including restricted encryption key access and deletion of customer data when an account leaves. The named subprocessors are Amazon Web Services for databases, caches and compute, Modal for processing prompts without storing them, and Slack plus Google Workspace for customer support; a notice dated 3 October 2026 added Nebius and CoreWeave. The four original subprocessors are all listed in the USA, so a buyer with EU data-residency rules should ask TypeSafe before sending production data.
Who should not pick TypeSafe yet: Jev is reachable only through a waitlist, accepts text input only, and caps each request at 64k tokens, per its model documentation. Its accuracy and speed figures come from TypeSafe's own four-workflow benchmark rather than an independent test. Teams that need open-ended reasoning, image or audio input, or a generally available API with published uptime terms are better served today by Claude Opus 5 or OpenAI's GPT-5.6 models. TypeSafe has also not published a formal model card, so buyers who must document model risk will be working from blog posts and docs.
Mission
Machine-native intelligence infrastructure for software decisions, not conversation ('Build Prod, Not God').
Products
- Jev (System One decision model / API): https://docs.typesafe.ai/models
TypeSafe AI Models on HokAI
Compliance
SOC 2 Type II
Links
Frequently Asked Questions
Who founded TypeSafe AI, and who is the CEO?
Diogo Almeida, a former OpenAI and Google Brain researcher credited as a co-inventor of RLHF, is TypeSafe AI's chief executive. He started the company with Erik Gafni, who serves as CTO, and Sasha Sheng, who serves as COO. The three ran it in stealth until 15 September 2026, when they announced a seed round led by DCVC.
What products does TypeSafe AI make?
TypeSafe AI's only shipped product is Jev, a System One decision model that returns typed outputs with calibrated confidence scores instead of generated text. Input tokens cost $0.042 per million, output is free of charge, and access currently runs through an early-access waitlist at console.typesafe.ai.
How does TypeSafe AI handle data privacy and compliance?
TypeSafe AI's trust center (trust.typesafe.ai, checked 8 October 2026) shows SOC 2 under compliance and offers a SOC 2 Type II report dated 2026 among its resources. On data handling, the company says Jev serves every account from the same base model weights rather than fine-tuning on one customer's data, and its subprocessor list names AWS for storage and compute and Modal for processing prompts without keeping them.
Which companies does TypeSafe AI compete with?
Jev's indirect rivals are large conversational systems, namely OpenAI's newest GPT line and Anthropic's Opus 5, both stronger on open-ended reasoning and backed by a bigger existing developer base. TypeSafe argues its own model instead wins on latency and cost for narrow, repeated, machine-to-machine decisions inside production software, a use case those larger systems were not built around.
How do you start using TypeSafe AI's products?
Developers can read the technical documentation at docs.typesafe.ai and request access through the waitlist at console.typesafe.ai. TypeSafe has said it is issuing API keys to approved applicants in batches rather than opening Jev to everyone at once.