I gave Claude 4.7 and Sonnet 4.6 three real freelance briefs pulled from Upwork. Real money on the table, real deliverables expected. One of them almost cost me a refund. Here's what actually happened when you stop reading benchmarks and start shipping.

The setup

Each model got the same raw brief - verbatim from the client's posting, no editing. No prompt engineering tricks, no prior context fed in. I timed the work, shipped the output, and tracked what the client would have accepted vs. rejected.

Brief 1: Next.js landing page with Mailchimp signup

A local cafe wanted a Next.js landing page with a working Mailchimp subscribe form. Straightforward scope - the kind of job that lives or dies on whether the implementation actually runs without a dozen back-and-forths.

Claude 4.7 produced a working Next.js 14 app-router page with proper server-action integration, Mailchimp audience API wiring, and a client-side success state that didn't re-render the whole page. No fixup needed.

Sonnet 4.6 shipped something that compiled, but the Mailchimp form posted to a stub endpoint. A real client would have come back asking why nothing landed in their audience list. I would have had to fix it - billable hours I can't charge for.

Brief 2: Shopify sentiment monitor

A Shopify store wanted a small tool that ingests product reviews and flags sentiment trends. Build-and-deliver in one turn. Claude 4.7 nailed the data model and shipped working Python with actual sentiment scoring. Sonnet 4.6 produced something that parsed reviews but had a subtle off-by-one in the rolling window. Would have passed a code review; would have given the client bad numbers for a week.

Brief 3: Invoice tracker built live in Claude Code

This one was different - I ran Claude 4.7 inside Claude Code (the terminal agent) with full file-system access. The ask was an Express + SQLite + PDFKit invoice tracker. Claude Code shipped 197 lines of working code, caught its own JSON parsing bug mid-build, and fixed it without a prompt. The kind of workflow that makes freelance work look like arbitrage.

Where Sonnet almost cost me a refund

The Mailchimp stub is the one that stings. Sonnet 4.6 wrote the form markup, wrote the submit handler, wrote the state update - and the handler posted to a URL it invented. Lint passes, tests pass, UI looks right. In production: nothing. For a $400 fixed-price Upwork job, a client would open a dispute within 48 hours. That's either a refund or an unpaid rework day.

The takeaway for operators

If you're using AI to ship freelance work: model selection is not a benchmark exercise, it's a risk-management exercise. The difference between "good model" and "right model for your stack" is the difference between keeping the money and refunding it.

• For production-integrated tasks (payments, APIs, auth), use the newer, more capable model - the cost delta is tiny next to one refund.
• For boilerplate and scaffolding, cheaper models are fine; just don't ship without running the output.
• Run the code before you send the invoice. Every time. Non-negotiable.

Watch the full breakdown

I recorded all three briefs live, including the Claude Code terminal session where it built the invoice tracker. Full video on TheOperatorAI YouTube.