TypeSafe must meter 23.8 trillion input tokens to bill a million dollars for Jev

At $0.042 per million input tokens, a million dollars of billings means 23.8 trillion tokens through the meter ^[3][5]. A million routing calls carrying 2,000 tokens of state each bills $84 ^[3]. Raise the state to 20,000 tokens for a rubric evaluation and the same million calls bills $840 ^[4]. The $5 entry price is consistent with input-only metering: $5 divided by $0.042 per million is 119 million tokens, near the roughly 120 million TypeSafe quotes ^[6].

The record on demand is thinner than the record on distribution. Surging demand last week briefly cost TypeSafe the ability to serve users from its API, according to the Indian Express ^[16]. Neither that account nor Vercel's model page puts a number on revenue, paying customers or funding ^[7]. TypeSafe has kept Jev's architecture under wraps, and the Indian Express reports that some observers speculate it is built on top of an open-weight LLM ^[14].

Diogo Almeida invented reinforcement learning from human feedback and was behind the instruction-following that underpins ChatGPT ^[8]. He left OpenAI nearly two years ago and founded TypeSafe AI with Erik Gafni and Sashsa Sheng ^[9]. "The problem is we are optimising for human language. We have been super good at human language for four years, but it's not useful for automation because computers speak a different language," he said ^[10]. Jev was trained only on synthetically generated data, using a method the company calls Reinforcement Learning for Calibrated Decisions ^[13], and TypeSafe says the model cannot hallucinate because the user supplies the context and defines the outputs in advance ^[15].

The company named the model after William Stanley Jevons, whose paradox holds that cheaper coal leads to more coal being consumed ^[12]. "We think that there's just going to be smart software all over the place in a way that's emergent and distributed ... much more like the early internet than you know like the mega apps that people are trying to build right now," Almeida was quoted as saying by TechCrunch ^[11].

In my view the pricing is the main risk. The buyers the Indian Express identifies are agent vendors switching in order to spend less ^[19], output is free ^[3], and the volume that makes input-only metering work has not been shown. The case on the other side is the company's own name: if each agent task fires dozens of cheap decisions, 23.8 trillion tokens is a year of traffic, not a decade ^[5]. And if the accuracy gap one developer found against Gemini matters more on a safety or routing call than the price gap does ^[18], the frontier model stays in the loop and Jev stays a small line item. TypeSafe says it will build more versions of the model in new modalities ^[20].