Start with a job, not a leaderboard
“Best AI” hides the decision you need to make. Drafting a campaign, editing a product image, fixing a repository bug, and answering a research question need different inputs, controls, and outputs. Write down the deliverable and the conditions for an acceptable result before you choose a tool.
Compare the same unit
A chat application, an API model, a coding agent, and an orchestration framework are different things. A product may call multiple models or connected tools. Record the product surface, plan, selected model, input files, and enabled integrations before comparing results.
Inspect three kinds of evidence
A vendor page can establish a documented claim. A reproducible task run can establish what happened in that configuration. A comparative evaluation can support a narrower statement about relative performance. One kind of evidence does not automatically establish the next.
Make the result usable
Check how many revisions were needed, whether the result can be exported in the right format, and whether a person can inspect and correct it. For agents, also check approvals, permissions, failure recovery, and execution records.
Use this catalog as a shortlist
Our profiles expose documented capabilities, sources, access caveats, and research dates. They are a starting point for a task-specific trial. Numerical performance rankings will require a published evaluation protocol and observed results.