We get asked this once a fortnight by founders who can only stomach paying for one AI subscription for their VA. The honest answer is “you want both”. The less-honest-but-useful answer is “if forced, Claude”.

Here is the side-by-side, scored against 14 real tasks our VAs run every week.

The 14 tasks, scored

Task	Claude	ChatGPT	Winner
Customer support reply drafts	9/10	7/10	Claude
Long-form blog drafts	9/10	7/10	Claude
Product description writing	8/10	8/10	Tie
Klaviyo flow email copy	8/10	8/10	Tie
Inbox triage + summarisation	9/10	8/10	Claude
Spreadsheet / CSV wrangling	9/10	8/10	Claude
Image generation	5/10	9/10	ChatGPT
Image-based input (read a screenshot)	8/10	9/10	ChatGPT
Current-web research	5/10	9/10	ChatGPT
Voice-to-action-list (audio transcripts)	8/10	8/10	Tie
Following complex multi-step instructions	9/10	7/10	Claude
Australian English consistency	9/10	6/10	Claude
Code-adjacent tasks (CSV, regex, JSON)	9/10	7/10	Claude
Speed of first response	7/10	9/10	ChatGPT
Average	7.9	7.7	Claude (just)

The aggregate is close. The shape of the difference matters more than the average.

Where Claude wins

Long-form customer-facing writing. Claude defaults to a less marketing-y tone, fewer exclamation marks, fewer cliches (“elevate your business!”, “revolutionise”). For an AU brand voice, that is closer to native. ChatGPT requires more taming.

Following complex instructions. When we ask for “8 customer reply options, the first 3 in 50 words, the next 3 in 80 words, the last 2 in 120 words, all Australian English, no em-dashes, signed off Cheers”, Claude gets it right first try roughly 85 per cent of the time. ChatGPT lands around 65 per cent.

Code-adjacent tasks. Anything involving JSON, CSV, regex, schema validation, or running scripts via Claude Code. ChatGPT can do these but Claude is fluent.

Australian English defaults. Tell Claude once to use AU spelling; it sticks. ChatGPT slides back to US conventions every 3-4 prompts.

Long context. Claude handles 200k tokens of context (your whole SOP folder, plus a client history) without quality degrading. ChatGPT degrades more visibly past 60-70k.

Where ChatGPT wins

Image generation. DALL-E 3 and the newer GPT image models are strictly better than Claude’s image output (currently no image output). For product mockups, social-post visuals, or quick concept art, ChatGPT.

Reading screenshots. Both can do this. ChatGPT is slightly faster and slightly more accurate at extracting structured data from a screenshot (an Instagram post, a UI screen, a printed document photo).

Current-web research. ChatGPT’s web search is more aggressive and more current. For “what’s the latest interest rate” or “what did Treasurer X say yesterday”, ChatGPT.

Speed. ChatGPT’s first-response time is consistently 1-2 seconds; Claude often takes 3-5 seconds. For high-volume task switching, that adds up.

Ecosystem. GPTs (custom assistants), Code Interpreter, the wider tooling ecosystem. Useful for specific advanced workflows.

What we actually pay for

Per VA, the stack costs $40 AUD/month:

Claude Pro ($20/mo): writing core, customer-facing drafts, content QA, long-form research, structured data work.
ChatGPT Plus ($20/mo): image generation, current-web research, screenshot input, ecosystem tools.

For specialist roles (heavy content creation, design-adjacent work), we add Perplexity Pro ($30/mo) for deep research. That brings the stack to $70/mo per specialist VA.

The ROI calculation: each tool saves the VA 30-60 minutes a week on tasks they would otherwise do manually. At $15/hr blended rate, 30 minutes a week is $32/mo in saved billable time. The subscription pays for itself before the second week of the month.

The single tool case

If you genuinely cannot pay for both, pick Claude.

Reasons:

Customer-facing writing quality is the highest-stakes output your VA produces. Claude wins there.
AU English consistency matters more than image generation for most placements.
Claude Code (the CLI) is included and outperforms ChatGPT’s Code Interpreter for file-touching tasks.
The long-context advantage compounds over months of placement.

You lose: image generation, fast current-web, the GPT ecosystem.

Workarounds: free Microsoft Designer or Canva for images, free Perplexity for current-web research, the ChatGPT free tier for the occasional image task.

Switching cost

Effectively zero for prompts. The prompts in our prompt library work on both models with minor adjustments. We have moved VAs between the two tools mid-week without retraining.

The real switching cost is muscle memory. A VA who has been on ChatGPT for 6 months will need 1-2 weeks to develop the same flow on Claude. Plan for it.

Where this goes in 2026-2027

Both companies are racing. Claude is gaining ecosystem breadth (Claude Code, web search, image input). ChatGPT is gaining writing-quality refinement. The gap between them on any given task is closing.

The bet we are placing: fewer, deeper relationships with tools, not more relationships with more tools. Our VAs are deep on Claude, working knowledge of ChatGPT, and fluent on the 6-7 SaaS tools that matter to each client. We have explicitly avoided training VAs on 12 different AI tools because tool-switching costs more than the marginal capability adds.

What’s next

For the working Claude Code playbook, see How we run our VA team on Claude Code.

For the working stack and what to pay for, see The AI-augmented VA stack.

For the prompts that work on either model, see Prompt library for Australian VAs.

If you want a placement that comes with both tools already trained in, book a discovery call.