AI
TypeSafe Shipped an AI Model That Decides Instead of Talking, and a 19th Century Economist Explains the Bet
The smartest bet in AI this month might be named after a 19th century economist. TypeSafe AI, a San Francisco startup founded by former OpenAI researcher Diogo Almeida alongside Erik Gafni and Sasha Sheng, emerged from stealth on September 15 with a $40 million seed round led by DCVC and a first model called Jev. The name honors William Stanley Jevons, the economist behind the Jevons paradox, when a resource gets cheaper, total consumption of it tends to rise. TypeSafe is betting the same logic applies to AI decisions. Make each decision radically cheaper, and software will ask AI to make vastly more of them.
Jev is a different species from the chat models that dominate the headlines. It produces zero text and answers in structured data instead. You hand it a chunk of unstructured state, like a support ticket, a row of transaction data, or a game's current position, plus a set of typed questions, and it returns structured answers with calibrated confidence scores in a single parallel pass. Ben Tossell, who writes the Ben's Bites newsletter, put it memorably. A new kind of AI arrived that skips the conversation and moves straight to the verdict. Text goes in, a number comes out.
The question types come in three shapes. A noul call answers a binary question with a probability. A choice call picks one option from a list of up to 255 and returns the full probability distribution. A score call places the state on a 2 to 10 scale. Because the space of possible outputs is fixed before the call runs, the model returns only values the code already knows how to handle. TypeSafe calls it a frontier intelligence function call, and DigitalOcean has already added it to its Model Catalog through serverless inference, with day one integrations from LangChain and PydanticAI.
The numbers explain the excitement. Jev answers in 70 to 500 milliseconds at $0.042 per million input tokens, with output free since every answer arrives as structured values. That undercuts even the cheapest frontier tiers by orders of magnitude. A DataCamp benchmark put Jev at about $0.0004 per decision against $0.03 to $0.18 per case for frontier models, while landing within a point of a leading frontier model on accuracy. LangChain's own guardrail test had Jev matching a human reviewer across 500 repeated calls at $0.00035 per call, against $28.17 total for a leading rival model on the same workload. The launch thread crossed 30 million views in days, and the model pulled more than 1,500 points on Hacker News.
The honest caveats matter too. Jev returns a number and a confidence score instead of an explanation, so teams still need to test it before trusting it on anything that counts. Its benchmarks so far come from the vendor's own workflows, and independent reproduction is still catching up. But the direction is clear. Classification, routing, scoring, and moderation now have a model built for them instead of borrowed from chat. For readers shipping software, the practical move is the two layer pattern the early adopters are already using. Let Jev make the fast calls, and escalate only the genuinely ambiguous cases to the big model. That is the Jevons paradox in action, and TypeSafe just made it cheap enough to run at scale.
Quick answers
What is this story about?
The smartest bet in AI this month might be named after a 19th century economist. TypeSafe AI, a San Francisco startup founded by former OpenAI researcher Diogo Almeida alongside Erik Gafni and Sasha Sheng, emerged from stealth on September 15 with a $40 million seed round led by DCVC and a first model called Jev. The name honors William Stanley Jevons, the economist behind the Jevons paradox, when a resource gets cheaper, total consumption of it tends to rise. TypeSafe is betting the same logic applies to AI decisions. Make each decision radically cheaper, and software will ask AI to make vastly more of them.
Why does this story matter?
The honest caveats matter too. Jev returns a number and a confidence score instead of an explanation, so teams still need to test it before trusting it on anything that counts. Its benchmarks so far come from the vendor's own workflows, and independent reproduction is still catching up. But the direction is clear. Classification, routing, scoring, and moderation now have a model built for them instead of borrowed from chat. For readers shipping software, the practical move is the two layer pattern the early adopters are already using. Let Jev make the fast calls, and escalate only the genuinely ambiguous cases to the big model. That is the Jevons paradox in action, and TypeSafe just made it cheap enough to run at scale.
Sources
New to crypto? Read the crypto glossary, browse frequent questions, read our story, or explore the story archive.