Claude 5.5: Opus, Sonnet or Haiku? Which size to use at each stage of a brand film

Between September 22 and October 7, 2026, Anthropic renewed the entire Claude family. First came Opus 5.5, the largest. Six days later, Sonnet 5.5, the mid-size model. On October 7, Haiku 5.5, the smallest.

Three sizes in just over two weeks change a practical question. It is no longer “which AI should I use?”. It is “which size should I use at each stage?”.

For anyone who makes or commissions brand films, the answer affects deadlines, cost and quality. This post lays out what Anthropic itself says about each model and proposes a stage-by-stage map.

Three sizes, three jobs

All three launched on every platform the company offers, including Amazon Web Services, Google Cloud and Microsoft Azure. The numbers below are the manufacturer’s, published in the official announcements. Prices are in dollars per million tokens (the unit of text models read and write).

• Opus 5.5 (9/22). According to Anthropic, it leads in agentic coding and knowledge work. The company says it costs 40% less to run than Opus 5 on typical workloads: the per-token price fell 20% (input at $4, output at $20) and it uses fewer tokens per task.
• Sonnet 5.5 (9/28). According to the company, it generates output more than 30% faster than Sonnet 5 and costs up to 30% less per task. The per-token price stayed the same ($2 and $10). On October 7, cache reads became 50% cheaper, which, according to Anthropic, cuts the cost of most agentic work by around 20%.
• Haiku 5.5 (10/7). Anthropic calls it the cheapest, fastest and most capable small model it has ever released. It says the model is built for high-volume work: summarization, subagents and browser use. On average, it would cost 75% less than Haiku 4.5. For prompts up to 100,000 tokens, input costs $0.10 and output $0.50.

The company sums up the split itself. Opus 5.5 is for “complex work requiring careful judgment”. Sonnet 5.5 is strongest at well-scoped everyday tasks and at polished documents, slides and spreadsheets. Haiku 5.5 pairs with the other two as a subagent.

A caveat before the map

All these numbers come from Anthropic. TechCrunch covered the Opus and Sonnet launches, but repeated the company’s claims, with no independent testing. None of the cost and speed gains cited here has been checked by an independent test on creative work.

So the map below is a starting point, not a rule. Every studio should test with its own material.

The map: which size at each stage

A brand film goes through stages with very different demands. Some call for judgment. Others call for volume and speed. Using the largest model for everything is expensive; using the smallest for everything is risky.

Research and briefing: Opus

The brief is where a film is won or lost. You have to cross what the brand says about itself, what the audience expects and what competitors have already done. That is judgment work, and it is what Anthropic positions Opus for.

The announcement also mentions a style change that helps here: Opus 5.5 puts the most important information up front and uses less jargon. For a brief that will circulate between client, agency and production, that matters.

Script drafts: Opus for structure, Sonnet for variations

The backbone of the script (concept, arc, tone) is a directing decision. It makes sense to build that base with the model that has the most judgment.

Then come the variations: 15, 30 and 60-second cuts, alternative lines, adaptations for each platform. That is well-scoped, repeated work, the profile Anthropic describes for Sonnet 5.5, faster and cheaper per task.

Either way, the final script goes through people. AI speeds up the draft; choosing what stays belongs to whoever directs.

Quick summaries: Haiku

A film project produces a lot of support text: meeting transcripts, client feedback emails, comments scattered across several files. Summarizing that is volume work, not creative work.

That is the use Anthropic points to for Haiku 5.5: summaries, classification and quick, repetitive tasks. The promised gain is in cost per request and speed.

Production documents: Sonnet

Schedule, call sheet, budget spreadsheet, client presentation. These need to come out correct and well formatted, but they do not call for invention. Anthropic highlights Sonnet 5.5 precisely for documents, slides and spreadsheets.

Agents: Opus coordinates, Haiku executes

Agents are programs that carry out several tasks in a row without asking for a command at each step. This is where the combination of sizes shows most clearly.

According to the announcement, Opus 5.5 delegates to subagents far more effectively than its predecessor. And Haiku 5.5 was designed to be that subagent. In practice: the larger model plans and reviews; the smaller ones search, summarize and organize in parallel.

It is the same design as Claude Duet, Anthropic’s proactive agent, which runs on Opus 5.5 and asks permission before acting.

A behind-the-scenes note

This blog is made with Claude. Research, writing and publishing of the posts are done by Opus 5.5 inside Claude Code. Topics and approvals are Enzo’s decision.

There is no cost measurement or comparison between the three sizes here. The note only makes clear how the text was produced.

This blog’s point of view

The news is not that there is a better model. It is that the family became clearer: one size to think, one to produce, one to handle volume.

For anyone making AI brand films, this looks like a production crew. Nobody puts the director on meeting transcripts, or the assistant in charge of the concept. Choosing the model by stage follows the same logic: each task with the level of judgment it calls for, and the final decision with whoever answers for the film.