Most of what an AI agent asks a frontier model to do is not writing. It is choosing. Which tool to call next, whether to retry, whether a command is safe to run. Jev, released September 15 by TypeSafe AI, answers those questions in under half a second for $0.042 per million input tokens and nothing on output. Within three days Vercel, Cloudflare, LangChain and Langfuse had added it to their stacks.

DCVC led a $40 million seed round to back founder Diogo Almeida, who helped build the instruction-following research behind ChatGPT at OpenAI before spending two years in stealth with co-founders Erik Gafni and Sasha Sheng. James Hardiman, general partner at DCVC, described the bet as turning capable models into components developers can build into products at scale. Agent vendors are the buyers. Decagon, which raised $250 million at a $4.5 billion valuation in January, runs support agents whose every routing, scoring and escalation step is currently a language model call. Move those steps to a model with free output and the margin expands for every agent company at once, while the lab owning the decision layer sits under all of them. TypeSafe named the model after William Stanley Jevons, whose paradox observed that cheaper coal produced more coal consumption, and expects intelligence to follow the same curve.

The mechanics are simple enough to explain in one sentence. Code sends a block of state and typed questions, and Jev returns a Choice among up to 255 options, a Score on an ordered scale, or a Noul, a yes-no probability, with calibrated confidence attached to each. Every question in a request is evaluated in parallel. TypeSafe trained it with a method it calls Reinforcement Learning for Calibrated Decisions, so confidence tracks accuracy and code can act automatically above a threshold and route to a human below it. Schema matching is guaranteed, which means an agent deciding between four tools cannot invent a fifth.

Every’s Mike Taylor ran 37 documents against 21 writing checks and got 777 judgments back in under 0.7 seconds for about a quarter of a cent. Dan Shipper’s head-to-head on 12 passages with planted defects clocked Jev at a median 0.35 seconds against 8.83 seconds for Claude Fable 5.1 at high effort, roughly 25 times faster at about one 580th of the cost, catching six of seven defects to Fable’s seven. One developer pointed Jev at 9,081 product-matching pairs his own engine had parked for human review and cleared the whole queue for 32 cents in 13 minutes, a stage he had shelved in June because frontier pricing made it not worth building. That is the Jevons effect in miniature: work that did not exist at $15 per million output tokens exists at zero.

TypeSafe’s own workflow evals across security alerts, agent review, invoice processing and customer service put Jev at 67.8 percent agreement with the averaged answers of GPT-6 Astra and Claude Fable 5.1, level with GPT-5.6 Terra and Claude Sonnet 5, at $0.0004 per case and 0.4 seconds. Sonnet 5 reaches the same score in 78.1 seconds at 294 times the cost per case, which is where the 193.6x speed figure comes from; the 444.6x cost figure is against Claude Opus 5. Gabriel Anhaia, a Berlin software engineer who built a cost model from the table, found that 10,000 decisions a day runs $120 a month on Jev against $35,220 on Sonnet 5 at identical accuracy, and that a 30-step agent run waits about 12 seconds on Jev decisions versus five minutes on Terra’s. The same evals showed decomposing tasks into typed questions lifted every LLM, which means the design pattern pays off before the model swap and compounds after it.

Vercel added Jev to its AI Gateway on September 16 and listed tool selection, next-action choice, risk checks and output validation as target use cases. Cloudflare listed it in its AI docs as typesafe/jev. LangChain shipped a TypeSafe classifier and Langfuse published a guide to eval scoring with it.

The caveats are the ordinary ones for a model three days old; the evals are vendor-run, the reference labels are other models’ answers rather than ground truth, and TypeSafe says the peak multiples sit at the higher end of real-world gains. The company published those qualifications itself under every claim.

For builders, ignoring Jev is a luxury most can’t afford. Pull last month’s model calls and flag every one whose output became an enum, a boolean or a number. Decompose those into Choice, Score and Noul questions, batch them into one request per agent step, and read fresh state after each action. Gate every autonomous action on the confidence field, keep the language model for anything that needs words, and let Jev run the safety reviewer in shadow mode until it agrees with your labelled commands.

For investors, the seed prices a new layer in the stack. If decisions and text separate into different model classes, every agent framework gains a fast, cheap primitive underneath its language model, and the volume of automated decisions grows the way Jevons predicted. The researcher who helped build RLHF at OpenAI just shipped the first model of that class with pricing designed to bend the demand curve. The next round will be priced on how far it bends.