TypeSafe Jev tops GPT-5 in intent test: 96.2% at 255.2 ms

The comparison is built around a practical complaint: using a generative LLM such as GPT-5 for intent classification means waiting seconds for a response and then handling format breaks and prompt-injection risks. AI Nexus's answer is to split the work—route simple choice-and-triage tasks to judgment-specialized System One models and reserve generative LLMs for cases that need free-form reasoning.

The 80-question set is divided into four categories of 20: typical clear cases, overlapping intents, sarcasm and indirect phrasing, and adversarial attacks. That breakdown explains why the overall numbers differ from the category scores: GPT-5 and GPT-5-mini both hit 100% on typical clear cases, while Jev's edge comes from consistent performance across the harder categories. On adversarial inputs, Decision-1 reached 100% and Jev 95.0%, whereas GPT-5 and GPT-5-mini each scored 80.0%—and Azure's jailbreak filter caused HTTP 400 crashes on those same tests, a failure mode the specialized models avoid because they do not generate free text.

The practical takeaway from the benchmark is a routing pattern rather than a single winner. Because System One models return a probability distribution, teams can watch the top-two margin and hold low-confidence cases for a human or a downstream model. A threshold of margin < 0.25 for fallback is a way to cut misclassification risk while keeping fast, cheap triage on the front line.