AgentBench vs Inspect AI

AgentBench has no write-up on Specifiction yet; its figures are above.

Inspect AI is a free, open-source, self-hosted framework for evaluating large language models. It covers coding, agentic tasks, reasoning, knowledge, behavior, and multimodal understanding. Evaluations combine datasets, solvers, and scorers, with more than 200 pre-built evaluations available. Inspect View provides web-based monitoring and visualization, while a VS Code extension supports log browsing, task execution, debugging, and option configuration. The framework has built-in support for more than 20 model providers and local inference through Hugging Face, vLLM, and SGLang. Documented provider groups include OpenAI, Anthropic, Google, AWS Bedrock, Azure AI, Ollama, and others. Standard tools cover web search, code execution, text editing, computer use, and web browsing, with MCP support. Agents from OpenAI Agents SDK, Pydantic AI, LangChain, Claude Code, Codex CLI, and Gemini CLI can be integrated. Untrusted model code can run in Docker, Kubernetes, Modal, Proxmox, Vagrant, and other systems through an extension API.