Claude Opus 5.5 programmatic tool calling: when video jobs fit it
Opus 5.5 supports programmatic tool calling. Fan-out video jobs fit it, sequential renders do not. Which Sume calls batch well and which stay one by one.

Claude Opus 5.5 is on Anthropic's list of models that support programmatic tool calling, along with Opus 5, Sonnet 5.5 and Fable 5.1, according to the docs read on 2026-10-04. Haiku 4.5 accepts the tool version but does not support the feature. For a video workflow, the useful question is which calls batch well, because the same page says the feature fits fan-out across many items and fits poorly for strictly sequential steps or a few small calls.
Which video calls are fan-out?
Sume's script_run tool description names the shape it is for: three or more independent calls of the same shape in one turn, such as a tts_create per sentence, a generate_image per scene, video_frames_create at many timestamps, or jobs_wait then jobs_result over a wave of jobs. Those items do not depend on each other, so a loop is the natural form.
Two or fewer calls should be made directly. The same description says to call them directly rather than script them.
Which stay sequential?
A talking-head shot is a chain. Sume's avatar tool notes describe it as a still, then tts_create, then an avatar image-to-video call on that still. Each step needs the previous output, so a loop adds nothing. Treat each chain as one unit and fan out across shots, not within a shot.
Rendering is the other limit. A script run is bounded by timeout_seconds (5 to 55, default 45), so it should submit creates and return job ids, then jobs_wait outside it. The Jobs and results page covers polling and results.
A decision table
| Task | Shape | Route |
|---|---|---|
| 10 scene stills | independent, same tool | one script_run with 10 generate_image creates |
| One voice line | single call | call tts_create directly |
| Talking-head shot | chain of 3 steps | direct calls in order |
| Check 20 finished jobs | independent reads | jobs_wait with up to 20 ids |
| Frames at 8 timestamps | independent, same tool | script_run with video_frames_create |
What about cost and tokens?
Anthropic suggests measuring billed input tokens with and without allowed_callers before assuming a saving, and Sume's guidance is similar in spirit: the bound is per run. A script run defaults to 32 calls and 16 paid calls, with ceilings of 64 and 32, and every paid create needs its own idempotency_key.
Set max_spend_usd on paid calls if you want a hard ceiling, because Sume enforces it only when you send it. A dry_run previews admission and cost without submitting. See MCP tools and gates for the gate list.
Sources
Related posts
More in Agents
- Opus 5.5 hands cyber tasks to Opus 4.8: log the model id
Anthropic says Opus 5.5 re-routes most cybersecurity tasks to Opus 4.8. Keep the model id that served each Sume Format run in your logs to explain odd results.
- Codex Cloud background tasks: three Sume guardrails to set first
OpenAI's Codex Cloud runs tasks in the background while you are away. Before one can call Sume, set read-only scope, idempotency keys and a hard spend cap.
- Cursor Projects coordinator fan-out: size waves from generation_limits
A Cursor coordinator that delegates to subagents can overrun a Sume workspace. Budget new in-flight jobs from generation_limits, not from wave_size_hint.
- Edit a photo from an AI agent: generate_image with input_references
The hosted MCP generate_image tool takes the same body as POST /v1/images, including input_references. Call tools_schema first, pass an idempotency_key.
Written by Sume