Claude Cowork vs ChatGPT Operator vs Perplexity Computer: I Used All Three for a Week

I spent a week putting three AI agents through the same set of real tasks. Not benchmarks. Not spec sheets. Real work: research a competitor product, update a spreadsheet, draft a blog post outline, book a flight, organize local files, and generate a presentation.

Here is what I found.

How I Tested

I gave each agent the same six tasks and tracked: did it complete the task, how long it took, how much I had to intervene, and whether I would use it again for that task.

The three tools:

• Claude Cowork — Anthropic's desktop agent. Runs on your local machine. $20/mo (Pro) or $100/mo (Max).
• ChatGPT Operator — OpenAI's browser-based agent. Runs in a cloud sandbox. $20/mo (Plus, 40 tasks) or $200/mo (Pro, 400 tasks).
• Perplexity Computer — Multi-model orchestration agent. Cloud-based. $20/mo (Pro with credits) or $200/mo (Max, 10,000 credits).

All testing done on a MacBook Pro M3 over seven days in July 2026.

Task 1: Research a Competitor Product

The ask: "Research [competitor]'s pricing page, extract all plan tiers, feature differences, and write a one-page summary."

Claude Cowork: Gave it the URL and asked for a summary. It opened the page, extracted the information, and wrote a clean markdown file to my desktop. Took about 90 seconds. No intervention needed. The file was well-structured with a comparison table.

ChatGPT Operator: Opened the website in its browser sandbox, navigated through three pages, and returned a summary in the chat window. But it stopped mid-way to ask if I wanted it to continue to the enterprise pricing page. Slightly slower (~2 minutes) because of the confirmation pause.

Perplexity Computer: Routed the task across multiple models. Returned the competitor pricing plus context from three other review sites, a Reddit thread about user complaints, and a GitHub discussion about missing features. The cross-referencing was genuinely useful, but it took almost 4 minutes. The output was verbose — I had to trim it down.

Winner: Claude Cowork for speed and clean output. Perplexity for depth if you have the time.

Task 2: Update a Spreadsheet

The ask: "Take this sales data CSV, add a column calculating Q2 growth %, highlight rows above 20%, and save as .xlsx."

Claude Cowork: Accessed the file on my desktop, wrote a Python script, executed it, and saved the output. Took 45 seconds. The file was in the right format with the right calculations.

ChatGPT Operator: Cannot access local files. I had to upload the CSV. It processed the data in its sandbox and offered a download link. Worked, but the extra upload/download steps added friction.

Perplexity Computer: Same limitation as Operator — cloud-only, no local file access. Uploaded the file. It processed correctly but the output had a different column order than what I asked for. Had to redo it.

Winner: Claude Cowork. Local file access is a real advantage for data work.

Task 3: Draft a Blog Post Outline

The ask: "Create a 5-section blog post outline about AI agents for small businesses. Include key points for each section and suggested headlines."

Claude Cowork: Returned a detailed outline with research notes attached. It had browsed three small-business blogs for context before writing. The outline was practical, not generic.

ChatGPT Operator: Generated a solid outline from its training data alone. It was correct but generic — no external research, no recent statistics. Fine for a starting point, but I would add the research myself.

Perplexity Computer: Best output here. It pulled from current articles, competitor blogs, and recent surveys. The outline included 2026-specific data points that neither Cowork nor Operator had. The extra model routing paid off for research-heavy tasks.

Winner: Perplexity Computer for depth. Claude Cowork for structure.

Task 4: Book a Flight (Web Transaction)

The ask: "Find a round-trip flight from New York to London, departing August 15, returning August 22, under $800."

Claude Cowork: Opened a browser, went to Google Flights, searched, and hit a CAPTCHA. Blocked. I had to complete the CAPTCHA manually, then it continued and found options. Partial success — it needed human help for the security check.

ChatGPT Operator: This is Operator's core strength — browser automation in a sandboxed VM. It navigated Google Flights, Kayak, and Skyscanner. Found a $743 option on the third try. No CAPTCHA issues (OpenAI manages the browser environment). Took about 3 minutes but completed fully autonomously.

Perplexity Computer: Searched via its research mode (not a live browser). Returned a summary of typical prices and recommended airlines but could not search live pricing or complete a booking. Not designed for this type of task.

Winner: ChatGPT Operator. This is exactly what it was built for.

Task 5: Organize Local Files

The ask: "Clean up my Downloads folder: group files by type into subfolders, delete duplicates, and create a log of what was moved."

Claude Cowork: Full local file access. It scanned the folder, identified 3 duplicate PDFs, created subfolders (PDFs, Images, Archives, etc.), moved files, and wrote a log file. Ran in about 20 seconds. Done.

ChatGPT Operator: Cannot access local files. Could not complete this task at all.

Perplexity Computer: Same limitation. Cloud sandbox only. Could not help.

Winner: Claude Cowork by default — the only one with local file access.

Task 6: Generate a Presentation

The ask: "Create a 6-slide deck about AI agent adoption trends. Source data from 2026. Output as a PDF."

Claude Cowork: Researched the topic, created slides in markdown, then offered to generate them. Could not create a native PPTX — I had to convert the markdown manually. The content was good, the format was incomplete.

ChatGPT Operator: Generated a text outline for slides but no visual output. Would need a separate tool (Gamma, Beautiful.ai) to create the actual deck.

Perplexity Computer: Created a structured document with charts described in text. Same limitation — no native slide generation.

Winner: None of them. All three handled the research but none could produce a finished presentation file. You still need a dedicated presentation tool for this.

Where Each One Shines

The Pricing Reality

All three end up costing $20/mo for light use and $100–200/mo for heavy use. The subscription math is similar. The difference is what each actually does well.

What I Would Keep

If I had to pick one: Claude Cowork for most days. Local file access is the feature that actually saves time — not generating chat responses, but doing work on your actual files. The $20/mo Pro tier is enough for daily use.

But I would also keep a ChatGPT Operator subscription for web transaction tasks and a Perplexity Computer session for deep-dive research weeks. They complement each other more than they compete.

The takeaway after a week: none of these agents replace the others. They each own a specific type of work. The right answer in 2026 is "use all three for different jobs" — which is a frustrating conclusion, but an honest one.