Introduction of Jev by TypeSafe AI
TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, has launched a novel AI model named Jev that fundamentally redefines AI decision-making. Unlike typical large language models (LLMs) that generate text or tokens sequentially, Jev operates by answering multiple structured questions in parallel with typed outputs and accompanying probability confidence scores. This approach delivers dramatic improvements in speed-with latencies between 70 to 500 milliseconds-and cost effectiveness-at $0.042 per million input tokens with output tokens free indefinitely. It is reported to be 20 to 200 times faster and 40 to 400 times cheaper than comparable LLMs, such as GPT-5.6 Terra or Claude.
Core Technology and Philosophy
Jev was developed over two years with the motivation that many real-world applications require fast, discrete decision-making rather than verbose natural language generation. It embodies a “System 1” model philosophy, inspired by cognitive science distinctions between fast, intuitive thinking and slower, deliberative reasoning. Jev focuses on making calibrated decisions-assigning confidence to each prediction to enable software to determine when human oversight is required-thus minimizing costly hallucinations common in traditional LLM usage. Its architecture circumvents the token-by-token generation paradigm, instead leveraging parallel probability computations to yield rapid, typed results. This model is optimized for embedding inside software workflows as a “smart if-statement,” enabling fine-grained decision automation within diverse applications.
Use Cases and Practical Deployments
Developers have integrated Jev into a wide variety of real-time and large-scale applications:
– Workflow and Agent Orchestration: Jev serves as a decision engine for routing tasks between AI agents and determining subsequent action steps, replacing slower LLM-controlled dispatch logic and reducing processing latency by up to 90%, while lowering costs by over 70%.
– Classification and Triage: It supports real-time message or email classification with confidence scores, enabling automated triage (e.g., fraud detection where uncertain cases are escalated to heavier LLMs), content moderation, and social media intelligence at fractions of traditional costs.
– Code Review and Security Checks: Jev enables extremely cost-effective automated pull request evaluations, scoring diffs for security issues and generating probabilistic verdicts that reduce reliance on expensive LLM calls and human review by escalating only uncertain cases.
– Real-time Game Playing and UI Automation: The model has demonstrated superhuman speed and decision accuracy in games such as Super Mario Bros, Subway Surfers (running 50 games simultaneously), Tetris, and Smash Bros, consistently making rapid sequential decisions better than or on par with existing LLMs, but vastly faster and cheaper. It is also used for fast, intelligent browser automation by selecting and executing UI elements with ~90ms decision latency without requiring pixel processing or LLM-level reasoning.
– Customer Support and CRM: Jev aids in rapidly classifying customer inquiries for routing, prioritization, and determining whether human intervention is needed, improving automation reliability while minimizing extraneous LLM usage.
– Financial Applications and Trading: The model contributes to fraud detection and early trading decision systems, classifying large volumes of emails and research papers quickly and at minimal cost, often acting as a lightweight front-line classifier with higher-accuracy LLMs as fallback.
– Data Analysis and Compaction: Integrated with tools like DuckDB and ClaudeCode, Jev helps manage context by scoring and pruning tool call histories, removing redundant or irrelevant data to reduce context size instantly without expensive summarization.
Strategic and Economic Impact
Jev introduces a new architectural paradigm by separating generation-heavy reasoning tasks from lightweight, fast, probabilistic decision-making. It empowers software to embed intelligent decisions directly into code flows, thereby reducing reliance on heavier LLM generations and enabling AI-enabled systems to scale with minimal incremental cost. Its “output tokens free” pricing further enhances unit economics by eliminating token generation costs, which traditionally dominate LLM expenses.
The model’s design emphasizes reliability through calibrated confidence scores, enabling safer automation by escalating ambiguous cases for human review or more complex AI processing. This innovation not only enhances AI performance but aligns well with regulatory expectations on explainability and risk management.
Experts consider Jev as a “missing building block” that complements rather than replaces existing LLMs. It constitutes an enabling layer for massively scalable, real-time AI decision workflows essential for driving next-generation software automation.
Developer Ecosystem and Access
Since its launch, Jev has been rapidly adopted by early users and developers worldwide. A public API and SDKs enable integration into diverse platforms. The model has been embedded in projects ranging from Discord moderation bots to browser automation frameworks and agent orchestration pipelines. The TypeSafe AI team supports open discussions through Discord communities and provides an official waitlist for API access at https://typesafe.ai.
Several open-source projects have emerged around Jev, including alternative implementations and utilities to better leverage its decision capabilities. The low overhead and high throughput have made Jev a favorite for developers seeking to infuse intelligent decision primitives into production systems without incurring prohibitive latency or cost.
Conclusion
Jev by TypeSafe AI represents a significant breakthrough in AI infrastructure, delivering fast, accurate, and cost-effective decision making as a core service. By foregoing natural language generation in favor of parallel structured outputs with confidence scoring, it unlocks new dimensions of efficiency and safety in AI-powered software automation. This model is setting the stage for a new generation of AI applications characterized by massive scalability, real-time responsiveness, and judicious human oversight. Its advent signals a paradigm shift in how intelligence is embedded in software, heralding a future where AI decisions are pervasive yet transparent and economically viable.