AI Developers are Suddenly Obsessed With Jev—And Swear By Its Low Latency

Developers are turning their attention to Jev, a new AI model that takes a different approach from ChatGPT, Claude, and Gemini. It makes structured decisions that software can use directly, rather than generating text.

The model is the brainchild of TypeSafe AI, a San Francisco-based startup that emerged from stealth with $40 million in seed funding led by DCVC. The company was founded by Diogo Almeida, who co-invented reinforcement learning from human feedback—the machine learning technique behind ChatGPT; Erik Gafni, who created multi-modal AI for DNA sequencing; and Sasha Sheng, an ex-research engineer at Meta.

Since its launch on September 15, the model has become one of the more talked-about developer stories on X, with programmers using Jev for everything—from routing AI agents and filtering content to playing games and managing context in coding agents.

Jev Evolves From ChatGPT

Almeida spent years at OpenAI working on methods that made language models better at following instructions and interacting with people. He was involved in research that became part of ChatGPT and co-authored work around InstructGPT and reinforcement learning from human feedback.

Almeida, in a post on X, said that he spent the past two years in stealth working on a new training method called Reinforcement Learning for Calibrated Decisions (RLCD) and a new type of frontier AI model. He said Jev is built for decision-making and can be 20–200 times faster and 40–400 times cheaper than existing models, with output tokens free.

“Computers speak a different language,” Almeida told TechCrunch, arguing that models optimised around human-readable responses are not necessarily the best tools for software that needs to make millions of small decisions.

That thinking led to Jev.

TypeSafe calls Jev its first System One model, a reference to the fast, intuitive form of decision-making described in Daniel Kahneman’s Thinking, Fast and Slow. The company says its goal is to build models that make fast decisions for software rather than models primarily designed to communicate with people.

“Jev achieves similar levels of intelligence on System One tasks compared to existing LLMs, while being two orders of magnitude faster and more efficient,” the company wrote in its blog post.

“While Jev gives up string generation, it’s optimised for structured outputs and can’t hallucinate. Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out,” it added. 

Jev Does Not Generate Text

The simplest way to understand Jev is to think of it as an AI decision layer sitting between an application and a larger language model.

A conventional LLM receives a prompt and generates a sequence of tokens. If an application wants to know whether an email is spam, which AI agent should handle a request or whether a tool call is safe, the LLM still has to generate an answer that the surrounding software then interprets.

Jev takes a different approach.

An application sends it a contextual raw data payload and a set of predefined questions. Jev then returns structured answers with probabilities and confidence scores.

For example, on whether a customer message is related to billing, technical support, or sales, Jev returns those choices as probabilities instead of writing a paragraph explaining its decision.

The company currently describes three core question types. A Choice selects one option from a predefined list, a Score evaluates something against a defined scale, and a Noul handles yes-or-no decisions.

The answers are generated in parallel rather than one token at a time. That is central to TypeSafe’s speed argument.

TypeSafe also claims Jev cannot produce type errors because its output structure is mathematically constrained. For hallucination comparisons, however, the company acknowledges that the LLM figures in its evaluation come from OpenRouter and that its own figure is based on guaranteed schema matching rather than an empirical measurement.

The startup says Jev can respond in between 70–500 milliseconds for the workloads it targets.

The company lists the model at $0.042 per million input tokens, while output tokens are free. That puts its pricing far below most frontier language models for the kinds of classification, routing, and scoring tasks it targets.

Developers are Already Finding Unusual Uses

The model has also drawn attention from developers testing it in real-world workflows.

Razorpay Co-founder and CEO Harshil Mathur, who has been using Jev, described the model as a different approach to AI infrastructure. “High-throughput, low-latency intelligence means AI can run inside every loop, reacting in real-time,” Mathur wrote on X. “This is a bigger deal than a better model. It’s a different kind of intelligence.”

Software engineer and creator of DiceDB Arpit Bhayani, who spent around 2.5 hours testing Jev, said the model could lead developers to rethink how they build AI workflows. “It is soooo very different and, more importantly, unlocks so many different use cases,” Bhayani wrote on X. “I can already see so many LLM-based workflows being rewritten around this.”

Developers are using Jev in AI workflows to route requests between coding agents, select files before sending them to an LLM, review pull requests, classify emails, detect fraud, rank content, and choose which tools an AI agent should use.

An AI agent might use a large language model to reason through a complicated task but use Jev for hundreds of smaller decisions around that model. Jev could decide which agent should receive a request, whether a tool should be called, whether an answer needs another verification pass, whether a workflow should be retried, or whether a human should take over.

That distinction matters because an agent can make dozens or hundreds of these decisions during a single workflow. Using a frontier model for every small decision can quickly become expensive and slow.

Hassan El Mghari, Director of Developer Experience at cloud platform Together AI, used Jev to classify 1,018 research papers for about $0.08 with a median latency of 256 milliseconds.

Romàn Czerny, Co-founder of B2B AI platform Gojiberry AI, used Jev to evaluate 700 high-intent leads and personalised outreach messages. “In 40 seconds, it predicted how each message would perform, assigned a confidence score, and detected lead-message mismatches. All for just $0.09,” Roman wrote on X.

He added that Jev could also score leads, analyse buying signals, match prospects with messages, and identify campaigns likely to perform based on available data.

Vercel’s early data indicates how quickly developers have latched onto the idea. Within 24 hours of Jev appearing on Vercel AI Gateway, nearly 13% of paid teams on the platform were using it. Vercel said that was more than twice the adoption of any previous model launch on its gateway during the comparison period.

Vercel engineer Pranit Sharma also reported that replacing a GPT-5.6 Luna classifier with Jev made the system 5–18x faster while improving accuracy in the company’s test. 

Jev Can Even Play Doom

One of TypeSafe’s most eye-catching demonstrations shows Jev controlling the first-person shooter video game Doom.

The company says the demonstration was built to show what low-latency AI decisions could enable in real-time applications. The model receives structured game-state information and decides what to do next.

There is an important caveat. Jev is not looking at the Doom screen like a human would. The demonstration provides structured information about the game environment, including positions and other state information.

The point was to show how quickly a model can make decisions inside a software environment.

Architecture Remains a Black Box

Despite the attention, plenty about Jev is still not publicly known.

TypeSafe describes it as a transformer-based model and says it uses a new architecture, parallel sampling and RLCD, or Reinforcement Learning for Calibrated Decisions.

The company says Jev was trained exclusively on synthetic data and that TypeSafe has built a research programme around statistically understood synthetic data. Almeida has not disclosed the full architecture or model size. 

Developers have already started developing their own projects. Open-source projects inspired by Jev have appeared for browser automation, agent routing, and other workloads, although these projects should not be treated as replicas of TypeSafe’s model.

Cheaper Intelligence

The name Jev comes from 19th-century economist William Stanley Jevons and the Jevons paradox, which describes how efficiency improvements can increase consumption rather than reduce it.

TypeSafe is applying the same idea to AI.

If intelligence becomes dramatically cheaper and faster, the company argues, developers will use it in places where AI is currently too expensive or slow.

That could mean an AI decision inside every software workflow rather than a handful of large AI interactions.

Almeida sees a future where intelligence is embedded throughout software, with small models making decisions continuously while larger models handle tasks that require deeper reasoning or generation.