Market-Bench - Fill Rate: leaderboard

Market-Bench (2026) places LLM-driven retailer agents together in one simulated supply-chain market where they bid for stock in budget-constrained auctions, set prices and write slogans for persona-driven buyers; values are means over 10 runs of a 6-step simulation at temperature 0. This board is the fraction of the buyer demand directed to the agent that it fulfilled (0 to 1); higher is better.

Metric: Fill rate (fraction of directed buyer demand in units that the agent fulfilled, 0 to 1); all 20 models compete as retailer agents in one simulated supply-chain market (sealed-bid procurement auctions, retail pricing and slogans; 6 steps, temperature 0), mean over 10 runs; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 19 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1O30.418#237
2Gemini 2.5 Pro0.221#276
3Gemini 2.5 Flash0.194#463
4Claude Sonnet 4.50.097#261
5Phi-40.091#1931
6GPT-4o0.079#751
7Qwen 2.5 VL 72B Instruct0.05#841
8ERNIE 4.5 300B A47B0.05#1122
9Market-Bench Gemma 3 27B (checkpoint unspecified)0.043
10Market-Bench Llama 3.1 70B (checkpoint unspecified)0.035
11Market-Bench InternLM2.5 20B (checkpoint unspecified)0.035
12InternLM3-8B-Instruct0.031#2703
13Market-Bench Qwen2.5 32B (checkpoint unspecified)0.031
14Qwen 2.5 VL 32B Instruct0.026#1062
15Market-Bench Qwen2.5 72B (checkpoint unspecified)0.022

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=market-bench-fill-rate · How It Works · Data refreshed daily, snapshot 2026-11-21.