DocsConcepts
Capabilities and levels
What an offering can do, and how sure we are.
Every answer carries capabilities for that offering, because the same model differs by provider (PDF input, prompt caching and tool use all vary). The names are input.text, input.image, input.pdf, input.audio, input.video, output.text, output.image, output.audio, output.video, output.embedding, tools, tool_choice, parallel_tools, structured_output, reasoning, prompt_caching, web_search, code_execution, computer_use, citations, batch, image.edit, transparency, fine_tuning, zdr, no_training.
Levels
Each capability has a level, in increasing trust: claimed, parsed, native, probed.
claimed- A catalog or document says so.
parsed- Accepted only after conversion, e.g. a PDF turned into text by the router.
native- The provider's own docs or API say the model handles it directly.
probed- A real call on the date shown succeeded.