Claude Sonnet 5.5 or Opus 5.5 for an agent that calls a video tool?

Either Claude model can drive Sume's video tools, because the tools and their spend controls are the same for both. What differs is on Anthropic's pages: Sonnet 5.5 is listed at $2 input and $10 output per million tokens, Opus 5.5 at $4 and $20. Model choice changes the token bill of the agent; it does not change what a Sume clip costs.

This page compares only facts printed on Anthropic's Sonnet 5.5 and Opus 5.5 pages, read 2026-09-29. It makes no ranking claim, and neither vendor page publishes a video-tool test.

What do Anthropic's pages list for each model?

Both are current, both run on the three big clouds, and both are in Claude Code.

What stays the same when the tool is Sume?

The tool list, the argument shapes and the safety controls belong to the MCP server. generate_video and jobs_wait behave identically, paid calls need an idempotency_key, dry_run=true previews cost, and max_spend_usd caps a call only when passed.

Video generation itself is priced by the video model and its length, not by which LLM asked for it. A five-second clip costs the same whether Sonnet or Opus wrote the request.

Where does the model choice show up?

In tokens. A video agent reads tool schemas, writes prompts, reads job results and reports back; so the per-token rate matters mostly when the agent runs many clips or a long plan around them.

Anthropic's Sonnet page says its testers found the model batched tool calls together more than Sonnet 5, and Sume's jobs_wait accepts up to 20 job ids per call. Whether that lowers your total depends on your run; measure a real task on both before you decide.

How do I test the choice cheaply?

Run the same three-clip task under each model with dry_run=true so nothing is submitted, compare the plans and the token counts your client reports, then turn dry run off for the winner. Keep max_spend_usd on for both.