TypeSafe AI
TypeSafe AI is the San Francisco lab that emerged from stealth on 15 September 2026 with Jev, the first of what it calls System One models: AI that returns typed decisions with calibrated probabilities for software to use, instead of generating text. It raised a $40 million seed round led by DCVC, and its CEO, Diogo Almeida, previously worked on RLHF, InstructGPT, ChatGPT and GPT-4 at OpenAI.
TypeSafe AI is a San Francisco lab that came out of stealth on 15 September 2026 with an unusual premise: most AI is built to talk to people, and TypeSafe builds AI for software to consume. Its first model, Jev, does not generate text at all. It reads a program’s state and returns typed decisions, such as a choice from a fixed list, a score or a yes/no probability, each with a calibrated confidence estimate that code can act on directly (TypeSafe).
The company calls this new category System One models, a name that echoes the fast, intuitive System 1 of dual-process psychology. The pitch is that many of the decisions software now hands to a large language model, like routing a ticket, classifying a document or checking a guardrail, do not need a model that writes paragraphs. They need a fast, cheap and consistent answer of the right type.
Quick facts
| Emerged from stealth | 15 September 2026 |
| Headquarters | San Francisco, California |
| CEO | Diogo Almeida, formerly of OpenAI |
| COO | Sasha Sheng, formerly of Meta and FAIR |
| CTO | Erik Gafni, repeat founder |
| Funding | $40 million seed round, led by DCVC |
| Reported valuation | $200 million |
| First model | Jev, in early access |
| Access | API and Python and JavaScript SDKs, via a waitlist |
What TypeSafe is building
TypeSafe’s manifesto argues that today’s models are already capable enough to create large economic value, and that the bottleneck is deployment. Its answer is to treat AI as a composable software primitive, closer to a database query than to an assistant: a component that returns inspectable, predictable outputs that developers can stack into larger systems. The manifesto sets a target of global total factor productivity growth reaching 3 per cent within five years.
The first product of that thesis is Jev. TypeSafe trains it with a method it calls Reinforcement Learning for Calibrated Decisions (RLCD), which optimises for honest probabilities on decision tasks rather than for human preference, the goal of the RLHF training behind chat assistants (TypeSafe).
Models and products
| Model | Released | What it does | Price |
|---|---|---|---|
| Jev 1.13 | 15 September 2026, early access | Returns choices, scores and yes/no probabilities with confidence, from text input | $0.042 per million input tokens; output free |
Jev is not a general-purpose chat or coding model, and TypeSafe says so plainly: it does not write text, code or explanations. It is not on our ranking of AI models, which scores general models on independent intelligence benchmarks that do not apply to a decision-only model. The Jev page covers what it does, what TypeSafe claims and where those claims come from.
The team
TypeSafe’s founders come from the labs and companies that built the current generation of AI:
- Diogo Almeida, CEO, previously worked on reinforcement learning from human feedback, InstructGPT, ChatGPT and GPT-4 at OpenAI (SiliconANGLE). TypeSafe’s team page credits him as a co-inventor of RLHF and InstructGPT.
- Sasha Sheng, COO, was a research engineer at Meta and FAIR who worked on News Feed and AI research.
- Erik Gafni, CTO, founded Ravel, a multimodal AI company for DNA sequencing, and was an early employee at Invitae and Freenome.
Almeida framed the company’s shift away from chat assistants in one line to The Register: “people can’t be the only consumers of intelligence” (The Register).
Funding
TypeSafe raised a $40 million seed round led by DCVC, announced as it emerged from stealth. Forbes reported a $200 million valuation, citing an unnamed source (SiliconANGLE).
Frequently asked questions
What is TypeSafe AI?
TypeSafe AI is a San Francisco AI lab that emerged from stealth on 15 September 2026. It builds System One models, AI that returns typed decisions with calibrated probabilities for software to use, rather than generating text. Its first model is Jev.
Who founded TypeSafe AI?
TypeSafe is led by CEO Diogo Almeida, who worked on RLHF, InstructGPT, ChatGPT and GPT-4 at OpenAI, alongside COO Sasha Sheng, formerly of Meta and FAIR, and CTO Erik Gafni, a repeat founder.
How much has TypeSafe AI raised?
TypeSafe raised a $40 million seed round led by DCVC, announced on 15 September 2026. Forbes reported a $200 million valuation, citing an unnamed source.
What is a System One model?
A System One model is TypeSafe’s name for an AI model that evaluates a state and returns typed answers with probabilities, such as a choice, a score or a yes/no probability, in a single fast pass. The name echoes the fast, intuitive System 1 of dual-process psychology. It is built for high-volume decisions inside software, not for conversation or writing.
Is Jev a large language model?
Not in the usual sense. Jev reads text but does not generate it: every answer is a typed value with a probability attached. That makes it a different tool from ChatGPT, Claude or Gemini, and it is why Jev does not appear on our ranking of general AI models.
Models
| Model | SWE | Context | In | Out | Status |
|---|---|---|---|---|---|
| Jev | — | 64K | $0.042 | $0 | Preview |