Atlas

Benchmarks

← All benchmarks

OpenAI-MRCR 2-needle - 128K

General QA · 2025-04-14

OpenAI-MRCR (multi-round co-reference resolution) inserts two identical needle requests into a long synthetic conversation and asks the model to reproduce the response to a specified needle. This row is the 128K bin, averaging the string match ratio over prompts with 64K-128K input tokens. Higher is better.

Top models (higher is better)

ModelScore
GPT-595.2
GPT-5 Mini84.3
GPT-4.157.2
O4 Mini56.4
o355.0
GPT-4.1 Mini47.2
GPT-5 Nano43.2
GPT-4.538.5
GPT-4.1 Nano36.6
GPT-4o (2024-11-20)31.9
GPT-4o Mini24.5
o1 (2024-12-17)22.1
o3-mini18.7
Loading Atlas data…