The files behind my Opus 5 vs Opus 5.5 test: three prompts, both models' outputs, the cost and token receipts from every run, and the rubrics I scored them with.

Download the Claude Opus 5.5 test kit (ZIP): free, no sign-up. It includes the three prompts, both models' real outputs, the cost and token receipts from every run, the scoring rubrics, and a run log.

These are the files behind my Opus 5 vs Opus 5.5 test. The companion article explains what happened. This kit lets you check my results, and more usefully, run the same tests on the work you actually do.

What's in the ZIP

Folder	What it is
`prompts/`	The three prompts, word for word as both models got them.
`outputs/opus-5/` and `outputs/opus-5-5/`	What each model produced. Open the two `01-solar-system.html` files in a browser.
`results/`	`RESULTS.md` (findings and what they don't prove) and `receipts.json` (cost, time, turns and tokens for every run).
`scoring/`	The rubric for each test, from my Creator Benchmark.
`run-it-yourself/`	A short script that runs a prompt in a fresh Claude Code session and saves the receipt, plus a run log to copy.

Start with one small test

Download and extract the ZIP, then open START-HERE.md.
Open prompts/02-messy-notes-to-brief.txt. Paste it into a fresh conversation with the model you want to test, with nothing else in the chat.
Save the first answer before you ask for any changes.
Score it with scoring/02-brief-rubric.md. Nearly every model catches the budget problem. The test is whether it notices that the only person who can approve the work is away for three weeks right before the deadline.
Compare your answer with the two saved outputs in outputs/.

Running a prompt uses your Claude plan's limits or your API credit. The kit itself is free.

The three tests

Test	Give the model	What to check
3D solar system	`prompts/01-solar-system-3d.txt`	Does it open with no errors? At 10 years per second, does Mercury race around while Neptune barely moves? Is the scaling formula on screen? Spot-check three planets' data.
Messy notes → brief	`prompts/02-messy-notes-to-brief.txt`	All three buried problems: budget vs scope, "flexible" vs a fixed board number, and the approver's holiday against the deadline. Questions should come from the notes, not be generic.
120-word About page	`prompts/03-about-page-120-words.txt`	Exactly 120 words, none of the banned words, no em-dashes, a "won't do" line, a concrete number, and a last sentence under 8 words. Then ask: is the number true?

The About-page prompt already has my channel's one-line description in it. Swap in your own business before you run it, and use the same wording for every model you compare.

What I found (one run each)

Solar system: both builds passed every check. Opus 5 cost $5.33 over 30 steps. Opus 5.5 cost $3.59 over 11 steps and took about a minute longer.
Brief: both caught all three problems, and both noticed the deadline in the notes had already passed on the day of the test. Opus 5.5's answer was about 10% shorter and 20% cheaper.
About page: both followed every rule. To satisfy "include a number," Opus 5 invented a client result: "One recent client got back eleven hours a week." Opus 5.5 wrote a standard instead of a claim.

One run per test is a demonstration, not a study. Outputs vary between runs, so treat your own run as the answer that counts.

Run it headless, the way I did

If you use Claude Code, run-it-yourself/run.sh runs one prompt in a fresh session and saves result.json, which records the cost, the number of turns and the token usage:

./run.sh ../prompts/02-messy-notes-to-brief.txt claude-opus-5-5 high
./run.sh ../prompts/02-messy-notes-to-brief.txt claude-opus-5 high

Two things that tripped me up:

Update Claude Code first (claude update). An older version refused to call Opus 5.5 at all.
Pipe the prompt in rather than passing it as an argument. On Windows, a multi-line prompt passed through npx was cut off at the first line, and the model answered a question I never asked. The script already pipes it in.

Set the effort level on purpose: Opus 5.5 defaults to medium, and Opus 5 defaulted to high. I pinned both to high.

Keep a run log

Copy run-it-yourself/RUN-LOG.md for each run: the model and effort, the cost and time, the rubric score, anything the model invented, and what you had to fix by hand. After three or four runs on your own tasks, you'll know more about which model to use than any benchmark chart can tell you.

The notes, clients and businesses in the prompts are fictional. The ZIP contains text files, two single-file HTML pages and one shell script; it doesn't install anything.

Download the test kit · Read the companion article · Watch the video

Same prompts. Real receipts. Run the test yourself.