A new class of AI models is emerging that refuses to generate text at all, instead returning structured decisions with confidence scores in milliseconds at a fraction of the cost of traditional large language models. TypeSafe AI launched Jev on September 15, 2026, marking the first "System One" model designed specifically for automation workflows that need answers, not prose. The model costs $0.042 per million input tokens, with no charge for output, and processes requests in 70 to 500 milliseconds.

What Makes a "System One" Model Different from Traditional AI?

Traditional large language models like Claude Opus, Claude Sonnet, and Claude Haiku generate text one word at a time, a process called autoregressive generation. This approach works well for creative writing, explanation, and open-ended conversation. But for tasks where software needs to make a decision and branch on the result, text generation adds unnecessary cost and latency. Jev inverts this logic entirely.

Instead of generating text, Jev accepts a state (unstructured text or program data) and a set of typed questions, then returns structured answers with calibrated probabilities. The model cannot emit an output outside the schema you define. This architectural guarantee means developers no longer need to parse text responses or worry about the model returning unexpected formats.

"Chat-shaped models are the wrong interface for automation: software does not want strings, it wants values it can branch on," explained Diogo Almeida, founder of TypeSafe AI, who previously worked on instruction-following methods behind ChatGPT at OpenAI.

Diogo Almeida, Founder at TypeSafe AI

The three core decision primitives Jev supports are Choice (pick one option from a defined set), Score (rate a state against ordered descriptive levels), and Noul (a boolean yes-or-no question). Each question is evaluated in parallel against the same state, so adding more questions barely increases latency or cost.

How Does Jev Achieve Such Dramatic Cost and Speed Improvements?

The cost advantage stems from a fundamentally different architecture. Instead of generating tokens sequentially, Jev uses a parallel sampler that enumerates possible outputs in advance. This guarantees schema conformance without sampling, eliminating the computational overhead of token-by-token generation. The training method, called Reinforcement Learning for Calibrated Decisions (RLCD), optimizes for "epistemically honest" probabilities rather than human preference, meaning higher confidence correlates with higher accuracy.

On TypeSafe's own workflow evaluations, Jev matches mid-tier frontier models while delivering dramatic efficiency gains. The benchmarks show Jev achieving comparable accuracy to GPT-5.6 Luna and DeepSeek V4 Flash while costing 25 to 145 times less latency and hundreds to thousands of times less per call. TypeSafe's own Doom demo bot fires 10 decisions per second for approximately $7 per hour, making decision-making effectively free at scale.

How to Integrate Jev Into Your Automation Workflow

• REST API Access: Send a POST request to https://api.typesafe.ai/v1/systemone with your state and typed questions, receiving all answers in a single call with probabilities and confidence scores.
• Official SDKs: Install the Python SDK via pip install typesafe-sdk or use the JavaScript SDK for Node.js environments, with agent skill integration available for Claude Code and other agent platforms.
• Decompose Decisions: Break complex judgments into atomic questions, then combine them with weights and thresholds you control in your own code, keeping decision logic transparent and auditable.
• Threshold on Confidence: Use the confidence channel to route low-confidence cases to human review, treating the model as a classifier rather than a final authority.
• Parallel Question Batching: Submit multiple questions about the same state simultaneously; TypeSafe reports a 13-question briefing costs 12.2 times less and runs 10 times faster than sequential per-question calls.

What Tasks Is Jev Actually Built For?

Jev is deliberately incapable of open-ended work. It cannot write code diffs, answer arbitrary questions, or generate creative content. The fit is work that ends in a JSON object anyway: routing and escalation decisions (which team should handle this, does a human need to review first), scoring and ranking (sentiment analysis, document relevance, review priority), guardrails (screening prompts and outputs for jailbreaks), and high-volume extraction and classification.

The architectural pattern TypeSafe advocates is "keep code in control." Rather than delegating judgment to a model and parsing its text output, developers decompose a judgment into atomic questions, then combine answers with weights and thresholds they own. This mirrors the model routing orchestration patterns already used in production AI systems, but adds a confidence channel that most router comparisons ignore.

The honest boundary is clear: Jev can no more write a diff or answer an open question than a traditional router can. The interesting question is which share of an agent's calls were ever really generative. If agents mostly classify, score, and branch, the decision lane just got a price point that changes the arithmetic. The division of labor between string models and decision models becomes a real architectural choice rather than a vendor's slide.

How Does Jev's Pricing Compare to Claude and Other Frontier Models?

TypeSafe claims Jev is 238 times cheaper on input tokens than Claude Fable 5.1, at $0.042 per million input tokens. Output tokens are still counted in usage but not billed. The company is candid that the price may be subsidized and expects it to go down over time, not up.

On TypeSafe's published workflow evaluations, which use reference answers from the average of GPT-6 Astra and Fable 5.1, the cost-per-call comparison is stark. Claude Opus 5 costs $0.2152 per call on these decision tasks, Claude Sonnet 5 costs $0.3616, and Claude Haiku 4.5 costs $0.0047. Jev costs $0.0001 per call on the same workflows. The company notes these workflows were designed by TypeSafe's own capabilities team and bias toward OpenAI and Anthropic models, and calls its homepage multipliers "the higher end of real world gains".

Jev is available in early access as of September 15, 2026, with access gated behind a waitlist. There is no OpenCode provider yet, so integration is limited to the REST API and official SDKs. The model id is jev-latest, and you reach it at POST https://api.typesafe.ai/v1/systemone.