What matters in AI.

Subscribe

Research

Top AI model gets 59.7 percent on a new test of task steps

The test shows a model 2 paths in a long task. The model must find the right path.

Claimed, not confirmed

Researchers from Microsoft and City University of Hong Kong made Taste-Bench. It tests the steps that AI agents do in long tasks. They gave the test to 14 models. The top model, GPT-5.6 Sol, got 59.7 percent, and GPT-5.5 got 59.5 percent. A random selection gets 25 percent. The test has 502 questions, and each question has 2 paths.

Taste-Bench score: best model 59.7 percent, random 25 percent Three horizontal bars on a 0 to 100 percent scale: GPT-5.6 Sol 59.7 percent, GPT-5.5 59.5 percent, random selection 25 percent. Taste-Bench gave 502 two-path questions to 14 models. Taste-Bench score 502 questions, 2 paths each GPT-5.6 Sol 59.7 percent GPT-5.5 59.5 percent Random selection 25 percent RANDOM Other models: well below 14 models tested
Taste-Bench gave 502 two-path questions to 14 AI models. The best scored 59.7 percent; random guessing scores 25 percent.

Sources

  1. AI agents have bad taste in long-term decisions and a new benchmark proves itthe-decoder.com
  2. AI agents have bad taste in long-term decisions and a new benchmark proves itThe Decoder
AI MATTER · NEWS · AI MATTER · NEWS ·18 OCT2026

Posted

Tags

More in Research

All Research news