TypeSafe AI Launches Decision-Only Intelligence Model to Eliminate Text Hallucinations
San Francisco, Tuesday, 22 September 2026.
TypeSafe AI introduced Jev, a specialized decision model operating up to 200 times faster than standard LLMs by returning typed choices instead of text, eliminating structural hallucination risks for enterprise automation.
Launch of Decision-Only Architecture
TypeSafe AI emerged from stealth in mid-September 2026 with the release of Jev, a specialized artificial intelligence model designed to output structured choices rather than traditional text responses [1][5]. Positioned as a System One architecture, the model addresses enterprise concerns regarding text hallucinations by forcing direct, typed decision-making within predefined schemas [1][8]. The company, founded by InstructGPT co-author Diogo Almeida, announced the technology alongside a $40 million seed funding round led by DCVC [1][8]. Unlike standard large language models that generate tokens sequentially, Jev processes parallel questions to deliver calibrated probabilities for choices, scores, or binary outcomes [1][5].
Performance Metrics and Cost Efficiency
Technical benchmarks indicate significant latency and cost advantages for decision-focused tasks compared to frontier reasoning models. Jev processes inputs with latencies ranging from 70 to 500 milliseconds, whereas comparable reasoning tasks on standard LLMs can take between 3 to 329 seconds [1][3]. In terms of pricing, the model is listed at $0.042 per million input tokens with no charge for output tokens [1][7]. Comparative testing highlighted in industry reports suggests a 14-question call costs approximately $0.000043 on Jev, while a similar task on Claude-Haiku-4-5 was benchmarked at $0.0018 [3]. This price differential represents a cost reduction factor of 41.86 when utilizing the decision-only architecture for applicable workflows [3][7].
Enterprise Integrations and Ecosystem
Adoption accelerated rapidly following the initial release, with Vercel integrating Jev into its AI Gateway on 2026-09-16, making it accessible via the AI SDK 7 evaluate method without a waitlist [1][8]. Subsequently, the Spring AI Community announced Spring AI TypeSafe on 2026-09-20, a project integrating the hosted Jev API for enterprise decision automation within Java environments [3]. Developer tooling also expanded through LiteLLM v1.103.0-rc, which enabled proxy access with logging and cost tracking for the model [4]. These integrations allow developers to implement guardrails, such as evaluating coding agent commands for irreversibility or off-task status, with recorded accuracy rates for holds reaching 88% across 17,000 calls in early tests [1][5].
Market Implications and Limitations
While Jev guarantees schema compliance, meaning it cannot invent labels or return broken JSON, it is noted that the model can still return incorrect valid values from a provided list [1][5]. The technology is recommended for use alongside rather than instead of traditional LLMs, specifically for tasks like reranking, citation checks, and routing [1][3]. TypeSafe team members have indicated that future releases will focus on coding capabilities, specifically addressing context management and semantic linting against configuration files [1]. As of 22 September 2026, the model remains a hosted API service with undisclosed weights and no self-hosting option, requiring users to verify compliance terms for sensitive data processing [5][8].