Tech • AI • Robotics • Game
JEV Just Broke the Internet: This Is a New Kind of AI
9.4/10TL;DR
Typesafe AI has launched Jev, a non-LLM, transformer-based model designed to return fast, typed probabilities for software automation, betting that machine-native decision systems can outperform chat models inside code.
KEY POINTS
From chat to automation
Diego Almeida, a former OpenAI researcher who helped develop RLHF, argued that modern AI has been optimized for human language rather than for software systems. His view is that chat models became superhuman at conversation years ago, yet failed to deliver broad automation because computers need structured outputs, not free-form text.
A new model class
After leaving OpenAI two years ago, Almeida co-founded Typesafe AI, which has raised $40 million and has now unveiled Jev. The company describes it as a system one model for fast structured decisions: unstructured state goes in, typed probabilistic decisions come out, with no natural-language response required.
How Jev works
Inputs can be JSON or plain text such as a customer complaint. Developers query the model with primitives including choice, score and null, which return fixed outputs such as ranked categories, probabilities or yes-no judgments. In a support-routing example, Jev assigned a ticket to technical with 0.85 probability and 0.82 confidence, making the result easy for code to branch on directly.
Type safety over text generation
Unlike an LLM, which emits a string that may contain a useful answer, a refusal, invalid structure or a hallucination, Jev is designed so outputs cannot fall outside a predefined schema. Typesafe AI says that means no parsing or validation layer is needed and that schema mismatches are mathematically impossible within the system’s design.
Speed and cost claims
The company says Jev responds in 70 to 500 milliseconds, compared with 3 to 329 seconds for frontier language models in comparable workflow tests. In one demo, Jev answered in 0.114 seconds while a reference model took 8.566 seconds. Pricing is also sharply lower: $0.042 per million input units, or $42 per billion, with output free, versus mainstream LLM pricing that can run from cents to double-digit dollars per million tokens.
Training and architecture
Jev is transformer-based but not an LLM and does not generate text token by token. Typesafe AI says it uses a parallel sampling architecture and a training method called RLCD, short for reinforcement learning for calibrated decisions. The goal is to produce probabilities that are not only accurate but well calibrated, so high confidence actually correlates with correctness.
Confidence as a product feature
A central pitch is that automation depends less on raw accuracy than on knowing when the model is uncertain. Typesafe AI argues that if a model gets a task right 95% of the time but cannot identify the risky 5%, it is unsafe to automate. By returning explicit confidence estimates, Jev is intended to allow software to act on strong signals and escalate ambiguous cases to humans.
Evidence and caveats
The company published workflow-based evaluations rather than public benchmark scores, saying production tasks are made of many small probability-driven decisions. It reported gains of 193.6 times faster and 444.6 times cheaper in selected workflows, though the tests were built by its own team and critics note they may favor the product. The company also concedes some demos were simplified and that speed tests depend partly on geography and deployment conditions.
Early developer traction
Demand surged immediately after launch. Vercel said Jev became the fastest-adopted model in its AI Gateway history, reaching nearly 13% of paid teams within 24 hours, more than twice the level of the GPT 5.6 family and over six times Fable 5.1 in the same window. The rush briefly strained Typesafe AI’s API capacity.
Use cases and limits
Developers are testing the model for command safety, business-email classification, model routing, risk scoring and agent control. Vercel said swapping Jev into a command safety classifier yielded results 5 to 18 times faster and with better accuracy, while another user found Gemini slightly more accurate on email tasks but 10 to 20 times more expensive. LLMs are still expected to dominate chatbots, copilots, coding agents and open-ended reasoning tasks, while Jev is aimed at what Typesafe AI calls smart if-statements.
CONCLUSION
Jev reflects a widening split in AI between models built for human conversation and models built for software execution. Its long-term significance will depend on whether fast, calibrated, schema-bound decisions prove more valuable in production than ever-better text generation.
Ask a question