Skip to content

Context-Bench: Benchmarking LLMs on Agentic Context Engineering Recent

Published: October 30, 2025 | Original Post


An open-source benchmark evaluating LLM performance on file operations and multi-step information retrieval.

Top Performers

ModelScoreCost
Claude Sonnet 4.574.0%$24.58
GPT-572.67%~$49
GLM-4.656.83%--
Kimi K255.13%--

Even top models achieve only 74% accuracy, highlighting the complexity of multi-step information retrieval.

Benchmark Design

Three core properties:

  1. Contamination-Proof -- Questions derive from SQL databases with fictional entities
  2. Multi-Hop Tool Calling -- Requires agents to chain multiple file operations
  3. Controllable Difficulty -- SQL-generated questions enable progressive difficulty scaling

Evaluation

Agents access two tools -- open_files and grep_files -- and are assessed on search query construction, operation chaining, tool selection, and hierarchical data navigation.

Resources