Graded by the Models It Replaces

A San Francisco lab called TypeSafe AI left two years of stealth on September 15 with $40 million in seed funding and a model that returns decisions in place of prose¹². The model is Jev. Give it data and a list of typed questions, and it returns a verdict with a probability, one option picked from a list, or a score on a scale, each carrying a confidence number³. Developers noticed fast. Jev held the top of Hacker News for most of launch day, and by September 18 Vercel called it the fastest adopted model in the history of its AI Gateway⁴⁵. TypeSafe also published the number that did the rest of the work: in peak in-house testing, Jev ran 193.6 times faster and 444.6 times cheaper than frontier language models⁶.

What Jev returns

Diogo Almeida spent his research years on the other side of this problem. He worked at OpenAI on reinforcement learning from human feedback, the method behind InstructGPT and ChatGPT². His explanation for leaving that work reads as a correction to it. "I spent years working on models designed to make AI better at interacting with people. But if AI is going to fundamentally change how work gets done, people can't be the only consumers of intelligence. Most intelligence should eventually live inside software, running quietly in the background," Almeida said when the company left stealth¹. Read that last sentence twice. He puts intelligence inside software and seats the human outside the default path, which is the entire design of a model that hands back a typed answer a program can branch on.

The shape of the output decides the use. A model that returns a typed verdict with a probability slots into the places software already branches: routing a ticket, scoring a lead, picking an offer, flagging a document for review, choosing which creative to serve. TypeSafe pitches Jev at that layer, calling System One Models a class built for decisions inside software³. Marketing operations sit squarely in range, because a media buy and a personalization call are both high-volume decisions carrying a cost per call and a latency budget. Almeida founded the company in 2024 with Erik Gafni and Sasha Sheng².

Selling speed

Speed is the product. TypeSafe says Jev uses parallel processing to answer in under 100 milliseconds, and it puts the peak gap at 193.6 times faster and 444.6 times cheaper than frontier language models³⁶. Attention arrived ahead of verification. The founder's announcement passed four million views, and Forbes valued the company at $200 million on the DCVC round⁴⁸. Reporters at The Register watched the lab teach Jev to play Doom⁷. Adoption followed the numbers, and Vercel, which routes model traffic for a wide slice of the web development market, reported its fastest adoption on record within three days⁵.

Who wrote the answer key

Here the record turns. TypeSafe measured Jev's accuracy by agreement with two other frontier models, GPT-6 Astra and Claude Fable 5.1, and the reference answer for each question is the average of what those two models said⁹. Ground truth sits outside that arrangement. So what does 193.6 measure? It measures speed against models whose answers double as the grading key, which leaves correctness an open question for all three. One independent test exists. Every ran Jev on an extraction task and measured roughly 25 times faster and 580 times cheaper than Claude Fable 5.1⁹. Compare that to the headline. The cost saving came in higher than TypeSafe advertised, the speed came in near an eighth of it, and the whole result rests on a single task shape⁹.

The pitch beneath the pitch

Calibration is the deeper claim, and it explains why developers cared. TypeSafe argues that training on human feedback taught models to sound right, and that a confidence number trained a different way lets software automate above a chosen threshold and escalate below it⁸. Almeida gave Forbes the diagnosis in one line: "We've been optimizing for humans and we're super human at pleasing humans"⁸. Sit with the verb. He says pleasing, which describes a model tuned for approval, and approval is precisely what a benchmark scored on agreement with two other models captures. James Hardiman, a general partner at DCVC, framed the investment around the same gap: "TypeSafe is approaching one of the biggest remaining challenges in AI: turning increasingly capable models into technology that developers can reliably build into products at scale"¹. Hardiman's word is reliably. A confidence number that holds under load would earn it, and the published evidence for that number runs through two rival models and one outside test.

By the numbers

• $40 million in seed funding, led by DCVC, carried TypeSafe out of stealth on September 15¹.
• 193.6 times faster and 444.6 times cheaper: the peak in-house comparison against frontier models⁶.
• 100 milliseconds is the ceiling TypeSafe quotes for a Jev response³.
• 25 times faster and 580 times cheaper: Every's independent result on one extraction task⁹.
• Forbes valued the company at $200 million on the seed round⁸.
• Three days after launch, Vercel called Jev its fastest adopted model⁵.
• Four million views reached the founder's announcement post in the first week⁴.

What to watch

Independent benchmarks carry the story from here. A second and third outside test, run across varied workloads against a real answer key, would either confirm the speed claim or shrink it the way the extraction test did⁹. Watch the calibration numbers next, because a confidence score that stays honest under pressure is what lets a company automate an offer decision or a media buy at volume. Enterprise buyers will ask the plain question: what happens at the threshold, and who pays when a 0.91 turns out wrong? Watch the category too. Should Jev hold up, System One Models becomes a product line other labs ship, and the cost of a routine decision inside software falls toward a rounding error.

Sources

1. Business Wire, "TypeSafe AI Emerges From Stealth With $40M in Funding With New Model for Composable AI," Business Wire, Sept. 15, 2026, https://www.businesswire.com/news/home/20260915525333/en/TypeSafe-AI-Emerges-From-Stealth-With-$40M-in-Funding-With-New-Model-for-Composable-AI
2. SiliconANGLE, "TypeSafe AI exits stealth with $40M to build AI for use by software," SiliconANGLE, Sept. 16, 2026, https://siliconangle.com/2026/09/16/typesafe-ai-exits-stealth-with-40m-to-build-ai-for-use-by-software/
3. TypeSafe AI, "Introducing System One Models & Jev," TypeSafe AI, Sept. 15, 2026, https://typesafe.ai/blog/introducing-system-one-models-and-jev
4. TechCrunch, "A new kind of AI model from a ChatGPT inventor is thrilling developers," TechCrunch, Sept. 18, 2026, https://techcrunch.com/2026/09/18/a-new-kind-of-ai-model-from-a-chatgpt-inventor-is-thrilling-developers/
5. Startup Fortune, "TypeSafe AI's Decision Model Jev Becomes Vercel's Fastest Adopted Launch," Startup Fortune, Sept. 18, 2026, https://startupfortune.com/typesafe-ais-decision-model-jev-becomes-vercels-fastest-adopted-launch/
6. Tom's Hardware, "TypeSafe AI's Jev offers an alternative to LLMs that claims to be 193x faster and 445x cheaper," Tom's Hardware, Sept. 17, 2026, https://www.tomshardware.com/tech-industry/artificial-intelligence/typesafe-ais-jev-offers-an-alternative-to-llms-that-claims-to-be-193x-faster-and-445x-cheaper-system-one-type-model-is-bespoke-for-probabilistic-decision-making
7. The Register, "TypeSafe AI debuts model for machines that plays Doom," The Register, Sept. 16, 2026, https://www.theregister.com/ai-and-ml/2026/09/16/typesafe-ai-debuts-model-for-machines-that-plays-doom/5296711
8. Forbes, "This $200 Million Startup Wants To Fix AI's Overconfidence Problem," Forbes, Sept. 15, 2026, https://www.forbes.com/sites/the-prompt/2026/09/15/this-200-million-startup-wants-to-fix-ais-overconfidence-problem/
9. KDnuggets, "What Everyone Is Getting Wrong About TypeSafe AI's Jev," KDnuggets, Sept. 21, 2026, https://www.kdnuggets.com/what-everyone-is-getting-wrong-about-typesafe-ais-jev