ABAB Pedia

TypeSafe AI

TypeSafe AI is an artificial intelligence company based in San Francisco, founded in 2024, that develops “System One” models which software calls directly and which output structured judgments with probabilities rather than text; its first public model is Jev. The company came out of stealth on September 15, 2026, announcing a $40 million seed round led by DCVCDCVCDCVC (Data Collective) is an American deep tech venture capital firm founded in 2010 by Matt Ocko and Zachary Bogue, with offices in Palo Alto and San Francisco, California. It invests in companies that use computing and AI to solve real-world problems in fields such as energy, climate, biomedicine, defense, space, and industry; as of October 2026 it had more than $4 billion in assets under management across 13 funds.Open the full entry, and on October 9 of the same year announced an $870 million Series A led by Andreessen HorowitzAndreessen HorowitzAndreessen Horowitz (a16z) is a venture capital firm founded in 2009 by Marc Andreessen and Ben Horowitz. It invests from the seed stage through the growth stage and provides operational support to its portfolio companies.Open the full entry at a $7.5 billion valuation.

Contents10 sections
Key facts

Founding

TypeSafe AI was founded in 2024 by Diogo Almeida, who serves as chief executive; its co-founders are chief technology officer Erik Gafni and chief operating officer Sasha ShengSasha ShengSasha Sheng is a software engineer and entrepreneur, co-founder and chief operating officer of the artificial intelligence company TypeSafe AI. She worked on mobile apps and machine learning systems at Yahoo and Facebook (Meta), won a number of AI hackathons after leaving Meta, and in 2024 co-founded TypeSafe AI with Diogo Almeida and Erik Gafni.Open the full entry. Almeida previously worked at OpenAI on reinforcement learning from human feedback (RLHF), InstructGPT, ChatGPT, and GPT-4; the leadership team’s backgrounds span OpenAI, Google Brain, and Meta FAIR.⁠[1][2][2][3]

The company then spent about two years developing in stealth. Almeida wrote that at OpenAI he helped build the methods that made language models good at following instructions and talking with people, but he came to believe that chat models were missing something important, which led him to ask: models have been superhuman at chat for years, so where is all the automation?⁠[4][4][4][4]

Research Essays Published

Before leaving stealth, the company published “The Bitterest Lesson” on its blog, arguing, on the basis of Rich Sutton’s “bitter lesson,” that in machine learning choosing the right task matters more than data, compute, and algorithms; the next day it published a further essay criticizing “benchmaxxing,” the practice of optimizing for public benchmark scores.⁠[5][5][6][6]

Out of Stealth with Jev and a Seed Round

On September 15, 2026, TypeSafe AI came out of stealth and released its first System One model, Jev, opening it through an early-access waitlist; at the same time it announced a $40 million seed round led by the deep tech venture capital firm DCVC. A person familiar with the transaction put the round’s valuation at about $200 million.⁠[4][4][4][7][1][3]

The company defines System One models as a new class of models built for decisions inside software: they take unstructured program state as input and return typed values in a structure defined in advance (such as yes/no, a choice from a list, or a score on a scale), with calibrated probabilities and confidence attached to every decision, so that developers can set when the software should act automatically and when it should hand a case to a person for review. The model uses a new architecture, a parallel sampler, and a training method called Reinforcement Learning for Calibrated Decisions (RLCD), and it does not generate strings.⁠[8][1][8][4][4]

At launch the company’s published price was $0.042 per million input tokens ($42 per billion tokens), with output tokens free of charge; end-to-end response times were 70 to 500 milliseconds. The company said that in its workflow evaluations Jev was 193.6 times faster and 444.6 times cheaper than the language models it was compared with, and noted that these figures came from workflows written by its internal team and were likely at the high end of real-world gains; the figures have not been independently verified.⁠[4][4][4][4][1]

The name “System One” comes from the fast, intuitive “System 1” thinking described in Daniel Kahneman’s Thinking, Fast and Slow; Jev is named after the economist William Stanley Jevons, reflecting the company’s expectation that machine intelligence will follow the path of coal after improvements in steam-engine efficiency, with falling costs leading to rising demand.⁠[4][4][4]

James Hardiman, a general partner at lead investor DCVC, said TypeSafe was addressing “one of the biggest remaining challenges in AI”: making models reliable enough to embed in products at scale.⁠[1]

Guest on the a16z Podcast

Jack & Jill Case Study

On October 7, 2026, the company published a customer case study: within 10 days of its first test, the talent marketplace Jack & Jill switched all calls for a key stage of its candidate-matching pipeline from Gemini 3.1 Flash Lite to Jev; according to the company, accuracy stayed the same, cost fell by 88%, completion time was about 50% faster, and projected savings were $500,000 a year.⁠[10][10][10][10]

$870 Million Series A Led by a16z

On October 9, 2026, TypeSafe AI announced an $870 million Series A at a $7.5 billion valuation, led by Andreessen Horowitz with participation from Sequoia Capital, existing investor DCVC, and a group of angel investors; Martin Casado joined the company’s board. The company said that a third of the Fortune 500 were using Jev and that it had already saved customers millions of dollars in production.⁠[11][11][11][11][11]

In the announcement a16z argued that the real economic unlock should be for all software to carry the intelligence of the latest models natively, at performance and cost consistent with traditional software economics; Jev costs roughly 1/100 to 1/500 as much as frontier models and is about 100 times faster on classification tasks at comparable accuracy. The announcement coined the “Jev-ons paradox”: make intelligence cheap enough to call anywhere, and it gets called everywhere. It also said that AI should make software more powerful rather than just automate the recreation of existing software, and that “the true bitter lesson” is not about building a new being but about returning the primitives of intelligence to builders. The announcement closed with the line “Software continues to eat the world.”⁠[12][12][12][12][12][12]

jev-1.14.0 Released

On October 14, 2026, the company released jev-1.14.0 and, behind “danger zone” request headers, opened experimental features including image input, a box type that returns a single bounding box, and longer context, noting that these features were not yet recommended for production; the company said the new version’s largest gain was in multilingual support, planned to point jev-latest to it a week later, and said it would give 60 days’ notice before deprecating the older version.⁠[13][13][13][13][13][13][13]

Enterprise Services and Data Policy

As of October 2026, the company’s enterprise page stated that it had a SOC 2 Type II report, offered zero data retention to enterprise customers, and did not use customers’ prompts or other inputs to train or fine-tune models.⁠[14][14][14]

Related Organizations, Websites, and Public Accounts

TypeSafe AI: Official website: https://typesafe.ai/ (opens in a new window).⁠[8]

Sources

  1. TypeSafe AI exits stealth with $40M to build AI for use by software (opens in a new window)
  2. ChatGPT Co-Creator Launches TypeSafe AI With Jev (opens in a new window)
  3. a16z leads TypeSafe AI’s Series A at a $7.5 billion valuation (opens in a new window)
  4. Introducing System One Models & Jev - TypeSafe AI Blog (opens in a new window)
  5. The Bitterest Lesson - TypeSafe AI Blog (opens in a new window)
  6. Lies, Damned Lies, and Benchmarks - TypeSafe AI Blog (opens in a new window)
  7. TypeSafe emerges from stealth with a new way of doing AI (opens in a new window)
  8. Home - TypeSafe AI (opens in a new window)
  9. AI Can Write Code. Why Isn’t Software Better? (opens in a new window)
  10. How Jack & Jill Cut Candidate Screening Costs By 88% With Jev - TypeSafe AI Blog (opens in a new window)
  11. TypeSafe A raises Series AI (opens in a new window)
  12. Investing in TypeSafe AI (opens in a new window)
  13. Jev: For the builders AND dreamers - TypeSafe AI Blog (opens in a new window)
  14. Enterprise - TypeSafe AI (opens in a new window)