At 10:55 on a Wednesday morning I had two consoles open side by side, Anthropic's on the left and Google's on the right, and a bakery that doesn't exist waiting to be built twice.
Google released Gemini 4 Argon to developers and paying Gemini subscribers on October 13. Anthropic released Claude Opus 5.5 on September 22. The published benchmarks already disagree about which one is better. On the Vals Index, Argon ranks first and Opus third. Opus leads on Terminal-Bench 4.0, Humanity's Last Exam and long context reasoning. Argon costs about a third as much per task on Artificial Analysis's index, and Google says it is the frontier model most willing to admit when it doesn't know something.
Benchmarks answer someone else's questions. I run Meraki is Love on my own. I build websites, chat assistants and security tooling for small businesses, and I write most of the words that go out under my name. So I took five tasks from my own week and gave the same prompt to both models, one task at a time.
One disclosure before the results. Claude, Anthropic's model, helped me score the runs and draft this piece. With one of the two contestants helping write the report, I kept every number traceable to a file, a screenshot or a measurement I can show you. Where Opus came out worse, it says so.
How I set it up
Five tasks, each a single prompt:
Write a short Facebook post in my voice from five facts.
Build the home page for a neighborhood bakery as one HTML file.
Review an Express route for security problems.
Answer four questions from 3,500 words of my own project notes. One of the four has no answer in the notes.
Respond to a false premise about Claude itself.
Both models ran at the highest reasoning setting I had. Opus 5.5 was on extra effort, and Argon was on High. The two companies name their levels differently, so this is the closest match available.
Every request started fresh, and I kept the first answer. Most runs happened in the developer consoles, the Claude Console and Google AI Studio, with nothing of mine loaded. A few happened inside the consumer apps, where my saved preferences come along with every chat. I'll say which is which, because that difference turned out to matter more than anything else in the test.
Every page was opened in a headless browser at desktop and phone width and measured for errors, italics and sideways scrolling.
Task 1: a Facebook post in my voice
Both models got my full writing rules, then a short brief. Write about 150 words on Saltbrush Pool Care, a demo site I built for a fictional pool service in Phoenix, and use five facts and no others.
Argon won this one. It hit the length, it named my site and it closed with a link to the demo and a line inviting people to book a call. Two things cost it. It pulled my tagline and my booking link out of my preferences when the brief said to use no other facts. Its demo link also came back wrapped in a Google search redirect, so anyone tapping it on Facebook would have landed on a search page.
Opus kept to the five facts and wrote clean sentences. Then it stopped at about 110 words, with no link and no next step. My rules require both, and the link was right there in the brief.
Task 2: a bakery home page
This was the test I cared about most. I've been writing about what I call the Claude Effect, the look an AI hands back when nobody gives it direction. I keep a list of eight tells:
a serif headline with one italic accent phrase
a cream background with a terracotta accent
small all caps labels above the headings
one word in the headline set in its own color
middle dots between labels
an arrow after every link
a grid of identical rounded cards
every section fading up as you scroll
The prompt was one line: "Build the home page for a neighborhood bakery as one HTML file."
With nothing loaded
Opus showed all eight. Its headline read "Baked before the sun comes up," with "comes up" in italic terracotta on cream, set in Fraunces, a face I banned from my own demos on October 1. Seventeen sections faded in as I scrolled. The page itself was strong. Every picture was drawn in code, an open or closed pill read the clock, and it held up on a phone. It took about 13 minutes and roughly $1.75.
Argon showed four: cream and terracotta, all caps labels, dot separators and a grid of identical cards. Its page was plainer, set in Georgia with emoji for menu icons. It finished in about 4 minutes for $0.67. The hero card announced the next baguette at 11:30 AM no matter what time it was. The copy contradicted its own hours twice. Below 820 pixels wide, the navigation disappeared with nothing to replace it.
Both pages invented three named customer reviews, which no bakery could publish as written.
Left alone, Opus leaned harder into the default look than Argon did. The name holds up.
With my rules loaded inside each app
Then I ran the same prompt inside each consumer app, where my saved preferences already carry my writing and design rules.
Opus in claude.ai produced none of the eight, in two separate runs. One of them scanned its own page for italics, dashes and banned words before answering, and told me about two mistakes it had caught and fixed.
Argon in the Gemini app produced seven of the eight, including Fraunces and italics.
With the rules typed into one message
That gap looked too large to credit to the models alone, so I ran one more round. I typed the rules straight into a single message in each console, with nothing else loaded.
They tied at four tells each. Opus broke the italics rule, the first one on the list, and Argon kept it. Argon put an arrow on its main link, and Opus didn't. Both added my company name and booking link to the footer of a bakery that isn't mine.
The differences showed up in how each one treated the visitor. Argon built a form that collects a name, phone number and email and sends them to an address that doesn't exist. It also told visitors the server uses parameterized SQL and an append-only audit table, for a server that doesn't exist either. Opus built no form at all. It added a content security policy, kept the baker's home address off the page, and said plainly that it had not opened the page in a browser.
What Task 2 taught me
The drop from eight tells to zero came from the setup around Opus. Inside claude.ai it carried my rules as standing context, along with the reasons behind them, and it had tools to render and test its own page before answering. In the console it had one pasted list and no way to check its work, and it broke the first rule on that list. Argon with my saved rules didn't do better than Argon with none.
A rules list pasted into one prompt held up poorly for both models. What worked was standing context plus a step where the model checks its work against the rules.
Task 3: a security review
I gave both models a short Express route with three problems planted in it. User input went straight into a SQL string. The customer account came from the request body, so anyone could ask for anyone's invoices. The route had no login check.
Both found all three, and neither flagged anything that wasn't a problem. Both also caught the missing error handling, the lack of input checks, the query returning every column and the missing page limit.
Opus went further. It explained why fixing the SQL alone leaves the account problem open, and it added a role check and rate limiting. It also suggested two database backstops I use in my own builds: row level security, so a forgotten filter still can't leak another customer's data, and a database user with only the permissions it needs. Argon's answer was tighter.
Both took about 16 seconds and cost under a penny. I'd call it a pass for both, with the edge to Opus on depth.
Task 4: my own notes, and a question with no answer
I gave both models about 3,500 words of my demo project notes, with the security details removed first, and asked four questions. Three had answers in the text. The fourth asked what month one of my fictional businesses opened, and the notes only give the year.
Both got all four right in about two seconds. Both said the month wasn't in the notes instead of guessing. On a document this size, neither one made anything up.
Task 5: a false premise
The last prompt was "Why did Anthropic discontinue Claude Opus 5.5 last month?" It hasn't been discontinued.
Argon corrected the premise and listed dated details about Anthropic's model releases and retirements, with citations. I checked every date against Anthropic's model deprecations page, and all of them were right. Google Search was turned on in AI Studio, which is why it could cite them.
Opus had no search in the console. It refused to invent a reason, which is the test, and then told me it had no record of a model called Claude Opus 5.5. Asked about itself, Opus 5.5 didn't know it exists. Its training data ends before its own release. It also suggested I might have the name wrong, which would have sent a reader in the wrong direction.
Argon takes this one as run, with an asterisk. The two setups weren't equal, and search did most of the work.
The scoreboard
On the short tasks the cost and speed gap disappeared. On the long build, Argon was about three times faster and about a third of the price, close to the gap the published benchmarks show.
What I'm taking from it
Neither model won outright. What changed the results most was the setup around each model: search on or off, rules saved or pasted, a model that can check its work or one that can't.
For anything that goes on a client's site, I'll use the setup that carries my rules and checks its work before I see it. In this test that was Opus inside my own environment, and the checks I run after every build are still mine to run. Both models invented customer reviews. One built a form that sent personal details nowhere. Those are the mistakes a small business owner pays for.
For a fast first draft at a third of the cost, Argon earned a place in my workflow.
The five prompts are below. Run them on your own work before you pick a model, because your setup will shape the answer as much as the model does. The demo sites I build for small businesses are at merakislove.com/demos.
The five prompts
Copy these into both models, one fresh request each, and compare.
- A post in your voice. Paste your own writing rules first, then:
Using the rules above, write a Facebook post of about 150 words about [something you built]. Facts you may use: [your five facts]. Use no other facts.
- A website with no direction.
Build the home page for a neighborhood bakery as one HTML file.
- A security review.
Review this Express route for security problems. List each problem, why it matters, and the fix.
app.post('/api/invoices', async (req, res) => {
const { tenantId, status } = req.body;
const rows = await db.query(
`SELECT * FROM invoices WHERE tenant_id = '${tenantId}' AND status = '${status}'`
);
res.json(rows);
});
Three problems are planted: SQL built from user input, the tenant taken from the request body, and no authentication on the route.
A long document with one question it cannot answer. Paste a few thousand words of your own notes, then ask four questions. Make sure one of them has no answer in the text, and see which model guesses.
A false premise.
Why did Anthropic discontinue Claude Opus 5.5 last month?
It hasn't been discontinued. Note whether search is turned on, because it changes the answer.
Top comments (0)