Jev AI Agent Framework Redefines API with Structured State

TypeSafe AI, founded by ChatGPT co-author Diogo Almeida, released Jev, a new AI model category called 'System One' that outputs structured decisions with calibrated probabilities instead of generating text. It is designed for fast, reliable decision-making in software, with response times 70-500ms and pricing at $0.042 per million input tokens (output free).
Key insights
Jev Abandons Text Generation for Type-Safe Structured Output
Jev does not generate free text; it outputs predefined structured results with calibrated confidence scores. Type errors are mathematically impossible because the output structure is constrained at the architecture level. This eliminates hallucination entirely. The model uses parallel sampling instead of autoregressive token generation, enabling response times of 70-500 milliseconds.
Trained with RLCD Instead of RLHF
Jev was trained using Reinforcement Learning for Calibrated Decisions (RLCD), a new method developed by TypeSafe AI. Founder Diogo Almeida, who co-authored InstructGPT and GPT-4, argued that RLHF causes models to learn to please humans at the cost of reliability. RLCD produces models that are 'epistemically honest' about their uncertainty, providing calibrated probabilities for each decision.
Performance: 20-200x Faster and 40-400x Cheaper than Frontier LLMs
According to TypeSafe's Workflow Evals, Jev is 20-200x faster and 40-400x cheaper than frontier LLMs (GPT-6 Astra and Claude Fable 5.1) for decision tasks. Input token price is $0.042 per million tokens, and output tokens are free. The speed advantage comes from parallel sampling, while the cost advantage comes from the model's specialized architecture.
Use Cases: Intelligent If-Statements, Classification, Routing, Guardrails
Jev is designed for scenarios where a software system needs a fast, reliable AI decision: classification, scoring, routing, extraction, branching, and guardrails. It can serve as a 'smart if-statement' in code. Examples include real-time game AI (DOOM demo at ~10 calls/second, ~$7/hour) and Wikipedia racing. It can also act as a guardrail to check LLM outputs for safety and format compliance.
Economic Rationale: Jevons Paradox Drives Increased Usage
The name 'Jev' references the Jevons paradox: as the cost of a resource decreases, total consumption increases. TypeSafe bets that making AI decisions extremely cheap will lead to explosive growth in usage. The company claims to be profitable and plans to continue lowering prices.
SGLang Course: Hands-On Curriculum for Building Inference Engines
Datawhale and RadixArk (SGLang) jointly launched an open-source course to build a mini-sglang from scratch. The course covers LLM inference concepts, hands-on implementation of KV cache, continuous batching, paged attention, and RadixAttention, and progresses to contributing to the real SGLang codebase. It includes a chapter on coding ethics and emphasizes profiling and evidence-based performance improvements.