Opus 5.5 or Fable 5.1 for a video agent: when to step up

Start a video agent on Claude Opus 5.5 and step up to Claude Fable 5.1 only if your own evals at higher effort still fall short. That is Anthropic's own advice on its models overview: Opus 5.5 for most workloads, Fable 5.1 for demanding reasoning and long-horizon agentic work. Fable 5.1 costs 2.5 times as much on both input and output.

Everything about Anthropic's models below is from its models overview, pricing page and effort page, read on 2026-10-02.

How do Opus 5.5 and Fable 5.1 differ on paper?

Both think adaptively and always, and both support five effort levels, from low up to max. Opus 5.5 is the cheaper one to leave on its default, because its default is medium while Fable's is high, and effort changes how many tokens the model spends on thinking, tool calls and text.

What does one turn cost on each?

Say a planning turn reads 20,000 uncached input tokens and writes 4,000 output tokens, a shot list for a short ad. On Opus 5.5 that is 20,000 times $4 per million, $0.08, plus 4,000 times $20 per million, $0.08, so $0.16. On Fable 5.1 it is $0.20 plus $0.20, so $0.40.

That gap only matters if Fable finishes in fewer turns or with fewer retries. A model that costs 2.5 times as much per turn has to save more than 60% of the turns to break even, and Anthropic's own page says only that Fable is for work where Opus 5.5 at higher effort falls short. Anthropic publishes no figure for how often that happens on video work, and neither does this post.

When does stepping up make sense for video work?

If none of those is true, stay on Opus 5.5 and raise effort before changing models. Anthropic recommends running an effort sweep on your own evals rather than carrying a setting over from an older model.

• The agent has to hold a long plan across many scenes, such as a 30-scene series, and loses the thread at medium.
• Your eval shows Opus 5.5 at high or xhigh still gets the same class of mistake, for example reusing the wrong reference image.
• The cost of a bad render is much larger than the cost of the planning turn. Video generation is usually the larger line, so a better plan that avoids one retry can pay for itself.
• You have already tried tightening the brief or giving the agent a schema to fill, which is cheaper than a bigger model.

How do you try both on Sume?

Send the same Format input twice and change only the model field. The Formats API documents model as the Agents catalog id of the LLM that orchestrates the run, and it selects the orchestrator only: image, video and audio models are chosen by the Format's tools. An id outside the catalog is 400 invalid_request, and the receipt echoes the id that ran.

In the repo, Opus 5.5 is on every catalog, so openrouter/anthropic-claude-opus-5-5 is admitted everywhere. Fable 5.1, openrouter/anthropic-claude-fable-5-1, sits behind the OpenRouter catalog gate, so a workspace without it gets the 400. Retired Opus 5 and Fable 5 requests move onto their successors rather than failing. Compare usage.debited_usd_micros on the two receipts, which includes the turn's own LLM cost.

What Sume does not do is pick between them for you inside a run. Auto routes by its own rules, and a pinned model stays pinned.