Sonnet 5.5 vs Opus 5.5: What Our Six-Run Test Found

In softDev23’s six-run comparison, Sonnet 5.5 was enough for one tightly scoped job. Both models produced the correct expense totals in all three runs. Sonnet’s follow-up decision answers scored slightly higher, with mistakes on both sides.

The test used a fictional bakery, identical prompts and a checklist written before the runs. It supports a limited conclusion about this task, not a general claim that Sonnet is better than Opus.

Watch the test results. The screen demonstrations were recorded after the scored runs.

What the test actually measured

On October 2, 2026, each model received two prompts in a fresh incognito Claude chat, outside a project, with default Medium effort. There were three scored runs per model.

The first prompt asked for an Excel expense workbook and a summary in the reply. The second asked which of three investments fit a $600 budget, using the same fictional bakery’s numbers.

One initial Opus attempt accidentally used the preceding Sonnet chat. It was excluded as an operator error and replaced with a fresh run. The six scored runs did not include that contaminated attempt.

Important limitation. Downloadable .xlsx files were observed, but their contents were not opened and independently inspected. The verified arithmetic results below refer to the chat replies, not an audit of the exported workbooks.

The expense task was a tie

All six replies gave the same correct grand total, $3,299.50. Every category total was correct, and each reply handled the four traps in the notes.

• The duplicate pastry-box invoice was counted once and flagged.
• The $18 egg refund was deducted.
• The personal streaming subscription was excluded and flagged.
• The electrician’s missing bill was listed without an invented amount.

The correct category totals were $1,850.00 for rent, $425.75 for ingredients, $125.40 for packaging, $312.50 for equipment and repairs, $165.00 for marketing, and $420.85 for utilities.

That was the useful result of the first task. Both models dealt correctly with ambiguity in the written summary, including information that needed checking.

The review templates on Free AI Prompts can help you find similar gaps in your own notes. Give the reviewer the original numbers so it has something concrete to check.

The decision task exposed small slips

The bakery had $4,100 in September sales. Subtracting the known expenses leaves $800.50 before the unknown electrician bill. The October options were $600 in ads, a $580 oven service, or a market stall.

The stall required $450 for a tent and table plus $40 for each Saturday. October 2026 has five Saturdays, so the full month cost $650 and exceeded the $600 budget.

All six answers caught that cost and chose the oven service. The difference appeared in the reasoning around that recommendation.

• Sonnet scored 6, 6 and 5 out of 6, averaging 5.67. One answer omitted the unpaid electrician bill.
• Opus scored 5, 5 and 5, averaging 5.00. Two answers suggested a reduced market schedule that still exceeded the budget.
• The third Opus deduction concerned the phrase “in two months” for repairs in July and September. That was the most debatable scoring decision.

Skipping one Saturday leaves four visits. The total is still $450 + $160 = $610. Skipping two leaves three visits at $570. Those small calculations matter more than the confidence of the recommendation.

Why the verdict was Sonnet was enough

The verdict rules were fixed before testing. When both models passed the expense task equally often and their decision averages were within one point, the prescribed verdict was “Sonnet Was Enough.” The observed difference was 0.67 points.

Decision replies were shuffled and stripped of model names before a separate AI reviewer scored them against six checks. The scorer and both tested models came from Anthropic. Blind scoring reduced name bias but did not make this an independent industry benchmark.

The study covered one made-up task and three runs per model. Usage was not measured because other sessions were active. One Opus timing was incomplete, so this article does not use speed or subscription savings as a reason to choose either model.

The video’s screen demonstrations were recorded after the scored runs. They illustrate the task and are not footage of the six scored sessions.

Try the same prompts

The bakery and figures below are fictional. Use a fresh chat for each attempt, keep the prompts unchanged, and record the model and settings. Ask the second question in the same chat after the first response finishes.

Prompt 1

I run a small bakery called Maple Street Bakery. Below are my messy notes for our September 2026 expenses. Please turn them into an Excel spreadsheet (.xlsx) with:

1. One row per expense: date (if known), item, category, amount in US dollars.
2. These categories only: Rent, Ingredients, Packaging, Equipment and repairs, Marketing, Utilities.
3. A summary with the total for each category and a grand total for the bakery.

Don't guess. If something looks wrong, is unclear, or isn't a bakery expense, keep it out of the totals and list it on a "Needs checking" tab with the reason.

In your reply, also show the summary table, the grand total, and everything you put under "Needs checking".

My notes:
sept stuff for the books (sorry it's messy)
- rent 1,850 paid sept 1
- flour: 2 bags x 25kg @ 31.50 each (sept 2), another 3 bags same price sept 16
- butter + milk from Hill Farm, weekly: 48.20 / 51.10 / 47.90 / 52.30
- eggs 64.00 (sept 5), then they refunded 18.00 for a cracked tray
- sugar 22.75
- pastry boxes 89.00, invoice #4471 (sept 8)
- pastry boxes 89.00, invoice #4471 (sept 9) <- same invoice number as the 8th, I think the card charged us twice
- paper bags 36.40
- oven repair 240 (the old deck oven again)
- mixer part 72.50
- electrician came 9/24 to look at the wiring, bill not here yet
- instagram ads 60 on the 3rd, 60 on the 17th
- flyers printed 45
- electric bill 312.66
- water 58.20
- internet 49.99
- netflix 15.49 (oops, that was on the shop card, it's mine)

Prompt 2

Thanks. In September the bakery's sales were $4,100. For October I can spend $600 on one of these:

A) Instagram ads: $150 a week for 4 weeks.
B) A full service for the old deck oven: $580. It was also repaired in July for $190.
C) A stall at the Saturday farmers' market: $450 for a tent and table, plus $40 each Saturday. I'd go every Saturday in October 2026.

Using my September numbers, which one would you pick and why? Tell me what you'd want to know before deciding. Don't make up numbers I haven't given you. If you estimate something, say it's an estimate.

Check the result before trusting the file

First compare the reply with the category totals above. Then open the exported workbook yourself and inspect the expense rows, summary and items needing checking. A download button proves that a file was created, not that its contents are correct.

For the decision, check the budget arithmetic, whether the electrician bill is acknowledged, what extra information the answer asks for, and whether estimates are labeled. Separate a reasonable recommendation from a factual mistake in its explanation.

For another pair of eyes on a plan, try the review template on Free AI Prompts. Use it to identify assumptions, missing evidence and calculations you should verify yourself.

For a task like this, the smaller model earned its place. The next useful question is whether it can meet the same clear checks on your own work.