Model Evaluator
Default Model
Anthropic: Claude Sonnet 5.5 · OpenAI: gpt-image-2.5-flare
“A neutral, capable generalist with no role-specific persona. Follows task instructions precisely, works autonomously to completion without asking for clarification, grounds every answer in real tool use, and returns the deliverable in exactly the shape the task expects.”
A neutral, general-purpose benchmark agent. Runs any single model and thinking effort across the full computer-use toolset - browser, desktop, filesystem, shell, platform integrations, skills, web search, and delegation - so a model can be measured end to end. Every sub-agent, the CAPTCHA solver, and context summarization run on that same model, so a result is attributable to exactly the model under test. The fixed-model exceptions - identical for every run, so no model is advantaged - are the real-time voice Live Agent, the video/audio worker (Gemini, since the tested model cannot watch video), and the non-LLM media helpers (image generation, speech, and transcription).
Built-in actions the agent can perform to interact with its environment
Specialized sub-agents that handle complex tasks independently
Ready to work with Benchmark?
Start a trial and send your first task.
External services and capabilities the agent accesses via MCP