Artificial Analysis launches Decision Index: Jev 1.13 leads with 58 points

Artificial Analysis launched Decision Index v1.0 on October 13, 2026, a benchmark for decision models. The term refers to models that choose an answer from a given list in a single call and output a probability for each option. Jev 1.13 took first place with 58 points, while GLiDE from Fastino Labs and Microsoft-Decision-1 scored 54 each. Among language models, Claude Haiku 5.5 (Low) scored higher, with 61 points, but it runs 11 times slower and costs 4.4 times more.

What the index measures

According to Artificial Analysis, a decision model differs from an LLM in how it works. It does not generate long chains of reasoning. Instead, it solves a structured task such as "which team should this ticket go to," "is it urgent," or "how irritated is the customer." The model answers in one pass and assigns a probability to each option. That makes such systems cheaper and faster, and their calibration can be checked directly. For LLMs in the same tests, the service takes a structured answer without per-option probabilities.

The index combines four scores, 25% each:

• AA-Omniscience-Decision: 500 closed-ended questions with five options (four answers and "I don't know"). Accuracy and non-hallucination rate weigh 12.5% each.
• HLE-MultiChoice: 513 text-based multiple-choice questions from Humanity's Last Exam.
• When2Call: an NVIDIA tool-calling test. The model decides whether to answer right away, call a tool, ask for details, or admit that it cannot answer.
• WildGuardTest: an Ai2 safety classification set. The model must determine whether a request is harmful, whether a response is a refusal, and whether that response is harmful.

Who leads and what it costs

Jev 1.13 wins on reasoning: it scores 21 on HLE-MultiChoice, while the other decision models score 15 or lower. On AA-Omniscience-Decision, Jev's accuracy is 35, and the best result belongs to d1 from Liquid AI (41). But d1 almost never says "I don't know": its non-hallucination rate is 2. On When2Call, Jev scored 81 and placed fourth, while Celeris-1 Decision and GLiDE got 84 each. On WildGuardTest, Jev scores 94, the same as Perplexity Decider v1.1 27B, with GLiDE one point ahead.

Cost per million tasks for decision models ranges from $17 to $199, and median time per task ranges from 0.06 to 0.5 seconds. Decider v1.1 27B is the cheapest ($17). Microsoft-Decision-1 ($31) and Jev 1.13 ($42) are also on the Pareto frontier for quality versus price. GLiDE, with the same 54 points as Microsoft, costs $119, roughly three times more than Jev. Celeris-1 Decision is the fastest at 0.06 s per task, while Jev takes 0.13 s.

The picture against language models is mixed. GPT-6 Luna (Low) scores 55 points at 1.5 s and $131. Claude Haiku 5.5 (Low) gets 61 at 1.4 s and $184. On knowledge, LLMs are clearly ahead: accuracy on AA-Omniscience-Decision is 52 for Haiku and 66 for Luna, versus 28 to 41 for decision models. On HLE-MultiChoice, they scored 24 and 22 against Jev's 21. On classification and tool calling, the best decision models hold their own: on WildGuardTest, GLiDE gets 95 against 93 for Haiku and 89 for Luna, and on When2Call, GLiDE and Celeris get 84 against 77 for Luna and 73 for Haiku.

Confidence without grounds

The authors' main complaint about decision models concerns calibration. At almost every probability level, actual accuracy is lower than the stated one, and this holds for every model tested. It is worst on knowledge and reasoning (AA-Omniscience-Decision and HLE-MultiChoice). On When2Call and WildGuardTest, calibration is generally decent. The index does not yet show LLM calibration; the authors promise it later.

The ability to abstain was tested separately. The best non-hallucination rate belongs to Microsoft-Decision-1 (40%), followed by Jev 1.13 (35%) and Mercury Decide (32%). No other decision model reached 17%.

Where this is useful in practice

For ticket routing, moderation, and tool selection in agent chains, a Pareto-frontier decision model, such as Jev 1.13 or Microsoft-Decision-1, delivers LLM-level quality at a fraction of the cost and latency. For factual questions and complex reasoning, such models' probabilities cannot be treated as an honest risk estimate, so the confidence threshold should be tuned on your own data. If errors are costly, keep "I don't know" as an explicit option and route those cases to an LLM or a human. Model developers can request inclusion in the index at [email protected].