TypeSafe AI's Jev Emerges as a Faster, Cheaper Judge for Agent Evals

TypeSafe AI's Jev, a 'System One' model that returns typed answers instead of generating text, launched last week and spread faster through the Vercel AI Gateway than any prior model — then LangChain's own tests found it beat LLM-as-judge evaluators like Claude Sonnet 4.6 on both consistency and cost. Together the adoption data and the LangSmith benchmark point to a real, if early, third option for agent evals beyond code-based checks and LLM judges.

• Vercel reported Jev reached ~13% of AI Gateway teams in its first day — 2x the adoption of GPT-5.6 and 6x Fable 5.1 (Latent Space, Sep 19).
• LangSmith found Jev's quality-score variance was 92–913x lower than GPT-5.6 Luna, Terra, and Claude Sonnet 4.6 on the same eval set (LangChain, Sep 20).
• Jev averaged 0.44s and $0.00035 per call, totaling $0.34 for the test run versus $28.17 for Claude Sonnet 4.6.
• Caveat: LangChain frames this as a narrow, early test of one model on one eval task, not a general verdict on Jev as an evaluator.
• Six clones of Jev appeared within two days of launch, alongside a 36M-view launch video, per Latent Space's Sep 19 roundup.

Bottom line If agent evals are part of your stack, pilot Jev against your current LLM-judge setup now, but validate on your own workload before trusting LangSmith's early numbers.