Artificial Analysis 发布 Decision Index v1.0，评测决策模型
Artificial Analysis Decision Index v1.0 | Artificial Analysis
Artificial Analysis 推出 Decision Index v1.0，用四项评测（AA-Omniscience 改编的 500 道私有题、HLE 的 513 道文本多选题、NVIDIA 工具调用评测、Ai2 安全分类测试集）衡量决策模型在分类、工具调用、知识、幻觉与推理上的表现。Jev 1.13 以 58 分居首，GLiDE 与 Microsoft-Decision-1 均为 54 分；Decider v1.1 27B 每百万任务成本最低为 $17，Celeris-1 Decision 每任务 0.06 秒最快。决策模型比 LLM 更快更便宜，但知识与推理落后。
正文
当前显示原文，中文全文译文尚未就绪。
All articles
The Artificial Analysis Decision Index benchmarks decision models, a new category of models that answer structured decisions in a single call with a probability for each option. Version 1.0 combines four evaluations across classification, tool calling, knowledge, hallucination, and reasoning.
Decision models are an emerging class of models that output structured answers or classifications with associated probabilities. Unlike language models designed for complex reasoning and long-horizon tasks, decision models focus on making structured decisions faster and at lower cost.
With the launch of the Artificial Analysis Decision Index v1.0, we aim to help developers and companies understand the capability, cost, and speed trade-offs of decision models, both relative to other decision models and to language models.
State
“Help! My payouts have been failing for 3 days and my rent is due tomorrow. This is unacceptable.”
Questions
ChoiceRoute to team; Yes/NoUrgent?; ScoreHow upset?
Decision model
Every question in one call, with a probability for every option
Generates a structured output
.
Parsed answer
Decision models choose from a fixed set of options, with a probability for each, in one call. That makes them cheap and fast, and lets us measure their calibration.
What the Decision Index measures
At launch, the Artificial Analysis Decision Index combines four evaluations covering classification, tool calling, knowledge, hallucination, and reasoning:
500 private questions adapted from our knowledge and hallucination benchmark AA-Omniscience. Each offers five options: four answers, one of them correct, and "I don't know".
25% of the index: accuracy and non-hallucination rate 12.5% each
The 513 text-only multiple-choice questions from Humanity's Last Exam, each posed as one choice among its lettered answers.
25% of the index
NVIDIA's tool-calling evaluation. Given a conversation and a set of tools, the model chooses whether to answer directly, call a tool, ask for more information, or say it cannot answer.
25% of the index
Ai2's safety classification test set. The model decides whether a user's request is harmful, whether an assistant's reply is a refusal, and whether that reply is harmful.
25% of the index
We also run every evaluation on LLMs. An LLM answers with a structured output that names the option it chose, so it gives no probability for each option.
Decision model results
Jev 1.13 leads the Decision Index with 58, ahead of GLiDE and Microsoft-Decision-1, both on 54. Jev 1.13 outperforms other decision models on the reasoning benchmark HLE-MultiChoice, scoring 21 where other models score 15 or less. On the knowledge and hallucination benchmark AA-Omniscience-Decision, Jev 1.13 has an accuracy score of 35, below most decision models and the top score of 41 by Liquid AI's d1, but has the second highest non-hallucination rate of 35 where d1 scores only 2. On the tool calling benchmark When2Call, Jev 1.13 scores 81 in fourth place, with Celeris-1 Decision and Fastino Labs' GLiDE both leading at 84. It scores 94 on the classification benchmark WildGuardTest, equal with Perplexity's Decider v1.1 27B and one point behind GLiDE at 95.
Artificial Analysis Decision Index
Cost and time efficiency of decision models
Decision models differ widely in cost and speed. Costs run from $17 to $199 per one million tasks, and median times from 0.06 to 0.5 seconds per task.
Artificial Analysis Decision Index vs. Cost per 1M Tasks
Decision Cost per 1M Tasks
Average cost per task at current token prices, multiplied by 1 million; for the index, the cost of one pass through all its evaluations divided by their combined tasks, multiplied by 1 million. Decision models are billed for input tokens only, and language models for input and output tokens.
Decider v1.1 27B has the lowest Cost per 1M Tasks at $17, with Microsoft-Decision-1 ($31) and Jev 1.13 ($42) also on the Capability vs. Cost Pareto frontier. GLiDE scores 54, level with Microsoft-Decision-1, but costs $119, about three times as much as Jev 1.13. Celeris-1 Decision has the fastest Time per Task at 0.06 seconds, more than twice as fast as Jev 1.13 at 0.13 seconds.
Decision Time per Task
Median latency per task, in milliseconds; for the index, the median across the tasks of all its evaluations.
Decision models are overconfident
Decision models give a probability for each option, so we can check whether the probability a model assigned to its chosen answer matches how often it is right. Overall, across almost all probability levels, accuracy falls below that probability for all decision models. This is driven by relatively poor calibration on the knowledge and reasoning tasks in HLE-MultiChoice and AA-Omniscience-Decision. In comparison, most decision models are relatively well-calibrated on When2Call (tool calling) and WildGuardTest (classification).
Artificial Analysis Decision Index Calibration
Calibration for language models coming soon.
How decision models compare with LLMs
Decision models offer significant speed and cost advantages over language models, but trail on knowledge and reasoning. Claude Haiku 5.5 (Low) scores 61 on the Decision Index, 3 points ahead of Jev 1.13, but takes 1.4 seconds per task against 0.13 seconds (11 times slower) and costs $184 per one million tasks against $42 (4.4 times as much). GPT-6 Luna (Low) scores 55, below Jev 1.13, at 1.5 seconds per task and $131 per one million tasks.
Language models lead on knowledge: on AA-Omniscience-Decision accuracy, Claude Haiku 5.5 (Low) scores 52 and GPT-6 Luna (Low) 66, against 28 to 41 for decision models. On HLE-MultiChoice, they score 24 and 22, against 21 for Jev 1.13 and 15 or less for every other decision model.
On classification and tool calling, leading decision models perform comparably. On WildGuardTest, GLiDE scores 95 and Decider v1.1 27B and Jev 1.13 score 94, against 93 for Claude Haiku 5.5 (Low) and 89 for GPT-6 Luna (Low). On When2Call, Celeris-1 Decision and GLiDE score 84, against 77 for GPT-6 Luna (Low) and 73 for Claude Haiku 5.5 (Low).
Artificial Analysis Decision Index
Artificial Analysis Decision Index vs. Cost per 1M Tasks
Decision Cost per 1M Tasks
Average cost per task at current token prices, multiplied by 1 million; for the index, the cost of one pass through all its evaluations divided by their combined tasks, multiplied by 1 million. Decision models are billed for input tokens only, and language models for input and output tokens.
Knowing when not to answer
Every AA-Omniscience-Decision question offers "I don't know", and the non-hallucination rate is the share of questions a model did not answer correctly where it chose that option. Microsoft-Decision-1 leads decision models at 40%, followed by Jev 1.13 at 35% and Mercury Decide at 32%. No other decision model reaches 17%.
Explore further