ChatGPT's Inventor Just Killed the Chatbot

Jev AI recently grabbed headlines by playing the video game Doom in real time, processing 10 queries per second. This isn't just a parlor trick; it showcases Jev's core value: blistering speed. The model boasts response times between 70 and 500 milliseconds, making it 40x to 200x faster than conventional large language models for similar tasks, all for about $7 per hour during the Doom demonstration.

This groundbreaking AI comes from TypeSafe AI, co-founded by Diogo Almeida, a co-inventor of ChatGPT. Almeida deliberately pivoted from conversational AI, aiming to solve a more fundamental problem in software automation. He envisioned an AI that acts, not just talks.

Jev defines a new category as the first System One Model, functioning as a "frontier-intelligence function call" that makes typed, probabilistic decisions directly within code. It entirely abandons text generation, taking unstructured input and delivering type-safe structured values in a single parallel query, which inherently prevents hallucinations. This model is also incredibly cheap, charging $0.042 per million input tokens with output tokens free—238x cheaper than models like Claude Fable 5.1.

Jev doesn't operate like your typical chatbot, which generates text token by token. That sequential process inherently limits speed and makes LLMs unsuitable for real-time decisions. Instead, Jev employs a novel hardware-aware parallel sampler, evaluating and delivering all structured values simultaneously. This fundamental architectural shift fundamentally redefines how AI processes information, moving beyond conversation to direct action.

This isn't just a hardware trick; the training is different too. TypeSafe AI developed Reinforcement Learning for Calibrated Decisions (RLCD), a new methodology ensuring the model's confidence scores directly reflect its accuracy. This makes Jev exceptionally reliable for critical production systems, a stark contrast to the often-unpredictable outputs of general-purpose LLMs, where accuracy is often a guess.

The performance gains are staggering, positioning Jev for serious enterprise applications. It delivers response times between 70 and 500 milliseconds, making it 40x to 200x faster than conventional LLMs for structured tasks. Crucially, its outputs are schema-bound by design, making it immune to the hallucinations plaguing open-ended models. This isn't a chatbot; it's a decision engine built for speed and reliability, ready for real work.

Jev's pricing model redefines AI affordability. It charges approximately $0.04 per million input tokens, making the cost of feeding data into the AI almost negligible. Crucially, output tokens are "too cheap to meter," meaning you pay nothing for the AI's actual structured decisions.

This isn't just cheap; it's disruptive. For tasks like classification, Jev is an estimated 238x cheaper than a model such as Claude Sonnet 5. Such dramatic cost savings finally make real-time, high-volume AI applications economically feasible across entire software stacks, where prior models were simply too expensive to scale.

TypeSafe AI, Jev's developer, chose its name carefully, referencing the Jevons Paradox. This economic principle suggests that making a resource radically more efficient and cheaper doesn't reduce its consumption; it expands its use into countless new areas. For further technical insights, read Introducing System One Models & Jev - TypeSafe AI Blog.

Expect an explosion in how software uses AI decision-making. Jev's affordability will embed intelligent logic into every corner of the digital experience, from backend systems to user interfaces, making always-on AI a practical reality for developers and businesses alike.

Jev isn't here to replace ChatGPT in your browser. Its true target is the messy, expensive backend infrastructure developers build to get structured data from LLMs. Think of it as a specialized tool for when you need precise, machine-readable output, not creative prose. Traditional LLMs offer "JSON mode," but it’s often slow, unreliable, and still requires extensive validation and parsing.

This is where Jev shines. It bypasses the token-by-token generation entirely, delivering type-safe structured values in a single, parallel query. This architectural shift eliminates the slow, brittle layers developers currently bolt onto LLMs, saving significant time and compute cycles. For production systems, this is a game-changer for reliability and cost.

This capability makes Jev ideal for high-throughput, mission-critical tasks where speed and accuracy are paramount. imagine using it for:

• Classification
• Routing
• Scoring
• Data extraction
• Implementing fuzzy decision-branching directly in code where handwritten logic is too rigid.

These are the repetitive, high-volume tasks that grind down traditional LLMs and developer resources.

Ultimately, Jev isn't just another AI model; it signals a crucial market shift towards specialized, production-grade AI tools. It offers a new primitive for developers, moving beyond chat to embed intelligence as a core, efficient component of software logic. This isn't about talking to AI; it's about making software smarter, faster, and cheaper to build.

What is Jev AI?

Jev is a new type of AI model called a 'System One Model' from TypeSafe AI. It's designed for extremely fast, structured, programmatic decisions within software, not for generating conversational text.

How is Jev different from models like ChatGPT or Claude?

Unlike traditional LLMs that generate text token-by-token, Jev takes an unstructured state as input and outputs a type-safe, structured value in a single parallel query. This makes it orders of magnitude faster and cheaper for specific automation tasks.

Why can't Jev AI hallucinate?

Jev cannot hallucinate because its outputs are constrained to predefined schemas and choices. It doesn't generate open-ended text; it selects from or populates a rigid structure, eliminating the possibility of inventing facts.

What is a 'System One Model'?

Coined by TypeSafe AI, a 'System One Model' refers to an AI built for fast, intuitive, and automatic decision-making, akin to 'System 1' thinking in humans. It contrasts with slower, deliberative 'System 2' models like large language models.