TypeSafe's Jev: A New AI Model for Ultra-Fast, Zero-Hallucination Decisions

ChatGPT Co-Developer Unveils Jev: A New Class of AI for Software Automation

Diogo Almeida, a former OpenAI researcher who helped build the methods behind ChatGPT, has emerged from stealth with a bold new venture: TypeSafe AI. On September 15, 2026, the company introduced Jev, the first of its System One Models—a new class of frontier AI designed not for conversation, but for making fast, structured decisions that software can use directly.

The launch addresses a persistent paradox in AI: models have been superhuman at chat for years, yet automation remains elusive. Jev aims to bridge that gap by abandoning string generation entirely, instead outputting type-safe, probabilistic decisions in parallel—at speeds and costs that make AI viable for real-time, production-grade applications.

What Makes Jev Different from Traditional LLMs?

Jev's architecture represents a fundamental departure from autoregressive LLMs. Instead of predicting tokens sequentially, Jev processes queries in parallel, generating all outputs in a single pass. This enables end-to-end response times of 70-500ms, compared to the 3-329 seconds typical of frontier LLMs—a 40-200x speedup for System One tasks.

The model is trained using a proprietary method called Reinforcement Learning for Calibrated Decisions (RLCD), which optimizes for epistemically honest probabilities rather than human preference or verifiable rewards. Every output includes calibrated confidence scores, meaning higher confidence correlates with higher accuracy—a critical feature for automation.

Jev's outputs are type-safe structured values, defined in advance. This makes hallucinations mathematically impossible for type errors, a claim TypeSafe asserts with confidence. The model never generates free-form text; it only chooses from predefined options, functioning as a "frontier-intelligence function call."

Pricing and Performance: A Game-Changer for Developers

The economics of Jev are striking. Input costs are $0.042 per million tokens ($42 per billion), while output is free—"too cheap to meter." This compares to GPT-5.6 Terra's $2/MTok input and $12/MTok output, making Jev approximately 238x cheaper than Claude Fable 5.1 on the same tasks.

In a side-by-side demo, Jev returned a response in 0.114 seconds versus 8.566 seconds for GPT-5.6 Terra. For an app making one million calls with an average of 1,000 input tokens each, the cost would be just $42 before infrastructure—enabling developers to check every AI decision rather than sampling.

TypeSafe acknowledges that some claims require long-term validation. "We can't prove it isn't subsidized," Almeida noted, "but we expect pricing to go down, not up." The company also published workflow evals showing Jev owning the Pareto frontier for almost two orders of magnitude compared to leading LLMs.

Use Cases: From Customer Service to Doom

Jev's design targets AI-powered workflows—smart if-statements that classify, route, score, or branch where hand-written logic is too brittle. TypeSafe's example use cases include customer service request classification, invoice handling, security alerts, and reviewing AI agent outputs for jailbreaks or errors.

The demos are equally compelling. One shows Jev playing Doom in real-time, making about 10 queries per second at a cost of roughly $7 per hour. Another demonstrates wikiracing, where Jev navigates Wikipedia links with high-cardinality choices (up to 255 options) without hallucinating.

"Most intelligence should eventually live inside software, running quietly in the background," Almeida told Forbes. This vision positions Jev as a standard component for developers building reliable, AI-augmented systems.

Why This Matters: The Jevons Paradox in AI

The model's name references 19th-century economist William Stanley Jevons, who observed that efficiency gains in coal led to increased consumption. TypeSafe bets that dropping the cost of intelligence by orders of magnitude will unlock exponentially more use cases, even as token prices decline.

This wager assumes the market for AI decisions will be as broad as the market for energy—a claim that remains unproven given adoption barriers. However, the combination of speed, cost, and reliability could make Jev a critical enabler for embedding AI into software infrastructure, potentially reshaping how applications are built.

TypeSafe has opened early access and is onboarding developers from its waitlist. The company plans to publish more technical details, including documentation and workflow evals, while continuing to refine Jev's capabilities.

The Bottom Line

Jev represents a significant shift in AI architecture, prioritizing decision-making over conversation. While it won't replace LLMs for chat, coding, or creative tasks, its speed, cost, and reliability make it uniquely suited for automation. As Almeida puts it, "AI needs an interface software could depend on"—and Jev is TypeSafe's answer to that challenge.

For developers and enterprises, the implications are clear: AI can now handle routine decisions at scale, with confidence scores that build trust. Whether Jev lives up to its extraordinary claims will depend on real-world deployments, but the direction it points to—faster, cheaper, and more reliable AI for software—is undeniably compelling.