Opus 5.5 Turns Ideas Into Animation, and Someone Spent $723 to Have It Open a File

seeco posted a 14-second video on X. An orange cartoon character stands in front of an interface with a Light/Dark toggle, and the motion is complete. He wrote that Opus 5.5 turned a simple idea into this, "I'm still in shock." Then he dropped his PDF prompt pack link for people to try, and also attached a long article.

Two kinds of reactions appeared in the comments.

Sharath Kumar Nair said this was "Night render off a real brief. Not one prompt." Meaning there was a real requirements document behind it, and it was not achieved with one magic sentence. Vlixsi put it more directly: "one good spec doc beats 40 prompts."

Zhou Stephen pointed out what the video does not show. The 14-second clip cut out common problems from the first generation pass: scenes usually land in one go, but timing and pacing do not. He still had to manually adjust easing curves and pauses. He asked whether the prompt pack specifies easing curves or leaves motion timing to a second pass. There was no clear answer in the materials.

Arobi posted a screen recording showing Opus at work. It opens the Wikipedia entry for Croissant in a browser tab, reads the page content, and then continues modeling in the document editor beside it. It is reading the material, then taking action.

So where is the boundary of Opus 5.5's ability? It can read a structured requirements document and turn it into an animation with characters, interfaces, and motion. It gets the scene on the first pass; the sense of timing still has to be adjusted by a human. That description is less impressive than the 14-second video, but closer to reality.

The other thread is the bill.

Martin Ronfort posted AGTP's cost analysis in a reply. Developer fleyta had Claude Code calculate the bill for three months of session logs, totaling $2,364. 98% ran on Opus. $723 was spent on "Opus 5.5 opening files."

AGTP also quoted Fox's "Model Team Engineering": 30% of the bill was Claude reading files and web pages when no one had asked it to. Claude Code starts on Opus 5.5 by default on all paid plans. Opus 5.5's cached-hit price is $0.20 per million tokens, while Sonnet 5.5 is $0.10. Using a model that costs twice as much to do grep-level work means the money disappears quietly.

Put the two things together, and the contrast is direct. On one side, Opus 5.5 turns a PDF into an animation; the capability is real. On the other, developers let it read files, search code, and load context by default, while the cheaper Sonnet 5.5 can do this grunt work too. UCDOps put it this way: using the most expensive model for grunt work is like asking the chief architect to go get cookies.

AGTP listed several cost-saving actions:

• Start a new session for each task
• /compact after each step
• Write a handoff.md and restart
• Use /context to see what is eating tokens
• Turn off unused MCP servers
• Specify file names and functions; don't write "that code"

Someone asked whether routing could be automatic. Ulad Shautsou asked exactly this: is anyone automatically switching models by task type, or is it all manual? From the replies, most people are still switching manually, or pointing file-reading and search subagents to smaller models while Opus only handles planning and final output.

Anthropic also confirmed that it added two sentences to the Opus 5.5 prompt for Claude Code to prevent it from replicating someone else's specific game. Models will do things they were not asked to do, including reading too many files and also becoming too absorbed in a particular reference.

Opus 5.5 is not magic. It can turn structured input into visual output. But if you do not keep it in check, it will also choose the most expensive way to help you open a file.

The original tweet is here.