Frequently Asked Questions

An independent, reproducible leaderboard that runs frontier AI agents head to head across a growing set of evaluation suites. It is built and maintained by Neurometric so teams can compare real agent performance under one consistent methodology rather than vendor-reported numbers.

Most leaderboards rank a single model on a single benchmark. Harbor runs multiple agents across many suites under one methodology, at the agent level, so you get a fuller picture of how a tool actually performs across the kinds of work it will do.

Coverage spans software engineering, code generation, competitive and agentic tasks, reasoning, math, science, and quantum circuits, and the set grows over time. The dashboard and About page always show the live list of suites currently on the board.

Because we benchmark the full harness, not just the underlying model. “Claude Code · Claude Opus 4.6” means the Claude Code agent running on Opus 4.6. The harness (planning, tool use, retries) affects results as much as the model does, so we keep both visible.

We are actively expanding coverage. An agent appears on a suite once it has a complete, verified run. If a harness has not finished a given suite yet, it is left off that chart rather than shown with partial data.

Yes. Every suite is pinned to a versioned task set, and we run each agent through the same harness and grading pipeline. Each suite page links its source paper and repository so the task set can be verified.

The board is rebuilt from the published runs whenever new results land. New model versions and agents are added as they ship and complete a full run.

Yes. Harbor can run your own agent or model across the suites we track, using the same harness we use internally. Get in touch and we will help you set it up.