Meta Muse vs ChatGPT vs Claude: I Gave Them 3 Real Jobs and Checked Every Answer

AI tools have moved from answering questions to doing work. So I gave three of them real jobs from my own business: audit this website, audit my YouTube channel, and make a two-minute AI-news podcast. Then I checked every answer by hand.

One reported a problem on this site that isn't there. One named the wrong video as my best, off by a factor of nineteen. And one made a podcast in under a minute where every story checked out.

The rules

Meta's Muse launched on 8 September 2026. Meta pitches it as an agent that does things for you, although officially it's only available in the US and Canada. I put it up against the two assistants I already pay for and use every day, ChatGPT and Claude.

• Same prompt, same session, roughly the same $20 tier.
• One retry each.
• Public information only. None of them was logged in to my email, YouTube Studio or a payment card.
• If a tool can't do a job, that counts as a result.

Tested on 29 September 2026. These tools change fast.

Job 1: can it check my work?

I asked each one for the five most important problems on thelazynick.com, each with the exact page and the evidence.

All three opened the site. ChatGPT answered in about a minute and a half. Claude took about four minutes and crawled 102 pages. Muse took thirteen.

That's fifteen findings in total, so I checked all fifteen on the live site.

Claude's top finding was the big one. The Ireland page shows 25 places, but they're drawn by JavaScript and none of them has its own page. I opened the page source: not one of those place names is in it, and the sitemap lists only 17 URLs. As far as Google can tell, this is a Copper Coast site. That's a real fix, and I'm making it.

ChatGPT said my map was stuck on "loading". I opened the page in a normal browser and the map loads fine. Its crawler had read the fallback text that only appears when JavaScript is switched off.

Muse found a real template bug: several pages point their canonical tag at the privacy page. It also caught a www-redirect gap and a privacy page that says ads aren't active while the ad script loads.

The tally: Claude 5 out of 5. Muse 5 out of 5, though to be fair it had audited this site earlier in the same chat. ChatGPT 4 out of 5, plus one problem that doesn't exist.

Job 2: can it understand my audience?

I asked each one to audit my YouTube channel using public information only: which video is working, which title is weakest, and what I should make next, quoting real numbers.

Muse and Claude both picked my rent tax credit video. ChatGPT picked Mizen Head.

I know the real numbers. The rent tax credit video has 56,000 views. Mizen Head has under 3,000. That's nineteen times the difference. ChatGPT also quoted 174 subscribers and 71 videos, when the channel page said 180 and 74. It also called a title weak that I'd already changed.

On screen you can see why: its connected YouTube research service was out of credits, so it fell back to old search results.

Muse and Claude got every number I checked right. Claude went deepest, down to the exact view counts on my Shorts. Muse was faster, and its three ideas built on what's actually working: more money guides with the answer right in the title.

Job 3: can it make something I'd publish?

The brief: a two-minute podcast on this week's biggest AI news, with three stories, each with a source and a date.

Here's the catch. The tool that got my channel wrong made the best-behaved podcast. The tool with the most accurate audits missed the biggest AI story of the week.

• ChatGPT wrote its script in 58 seconds. All three stories were real, current and correctly dated. The audio came in at 1:57, right on brief, but to get it ChatGPT sent me to a separate voice app.
• Muse was the only one that made the audio itself, with two hosts. All three stories checked out. It ran 2:24, and it called that "right on target".
• Claude voiced its own with an open-source voice. Its three stories were real too, but it skipped the one both others led with: OpenAI cancelling its GPT-6.1 "Astra" model over safety concerns. One of its stories was also six days old.

Where trust broke

Three jobs, nine attempts. These were the worst misses:

• ChatGPT named the wrong best video, using out-of-date numbers.
• ChatGPT reported a broken map that works.
• Claude, the most careful auditor, missed the week's top story.

Everything in this test was public, and I'd start any AI agent there. This week a YouTuber said Muse shared his home address with a Facebook Marketplace buyer (Meta says that in similar cases Muse followed instructions and asked permission). Meta also paused a test where humans sometimes made Muse's phone calls. I didn't test either of those. They're why I'd begin with public, low-risk jobs like these.

One thing only Muse did: it pitched me work

Muse has an Ideas page built from my own channel and site. It suggested packaging my rent-credit video before the claims deadline, clearing six claims holding up my Copper Coast video, and decoding my six thumbnail tests before 11 October.

After filming, I said yes to two:

• The thumbnail playbook came back mostly right. I checked it against YouTube's own help page. The winner is picked on watch time, not clicks, and changing the title or thumbnail mid-test stops the test. Both are right. But the playbook also said a test never ends on its own, and YouTube says tests finish within two weeks.
• Rebuilding six "missing" video skills was done, with one caveat. Four of them were already on my Mac. Muse only saw the zip file I gave it, and to be fair, it said so in the small print. Its rebuilt narration skill even sets a lower quality bar than my real one: 95% against my 98%.

So Muse is the best at spotting work. You still have to check what it knows.

Verdict: the right tool for each job

• To check my work: Claude. It scored 5 out of 5 and found the biggest problem on this site. Muse matched the score, but it took thirteen minutes and had seen the site before.
• To understand my audience: Muse or Claude. Both got my best video right with real numbers. Muse was quicker, and Claude went deeper.
• To make something I'd publish: ChatGPT wrote the most accurate podcast, and the fastest, but it needed another app for the voice. Muse was the only one that made the whole thing itself.

If you're outside the US, Muse isn't officially an option yet, so the pair to beat is Claude and ChatGPT.

I'm not picking a winner for you. Which one would you keep? Tell me in the comments on the video.

Sources

• Meta Muse availability and downloads: The AI Insider, 28 Sep 2026
• Marketplace address claim and Meta's response: The Next Web, 28 Sep 2026
• Human-concierge test: BNN Bloomberg / Reuters, 22 Sep 2026
• OpenAI scraps GPT-6.1 "Astra": Washington Post, 28 Sep 2026
• YouTube Test & Compare: YouTube Help

Tested 29 September 2026 using public information only. The companion video uses an AI avatar and an AI-cloned version of my voice. This article was drafted with AI help from my own test notes, and every figure in it was checked by me.