Last week I asked GPT to polish a client email and it wrote "I hope you feel our sincerity" — I lost it

Last week I asked GPT to polish a client email. It wrote back: "I hope you feel our sincerity." I couldn't tell whether to laugh or cry. Three weeks ago, with the same prompt, it could still churn out a decent opener.

Ever feel like a prompt that worked great last month suddenly flops this week? You're not getting dumber — the model might have been quietly downgraded.

What's actually going on? Who's already hit this wall

A recent technical blog (w4g1.dev) dug into it: OpenAI, Anthropic, Google — these big names will quietly route your request to a smaller, cheaper model during peak hours, or slowly degrade output quality. The reason is simple — running big models burns cash, and saving money beats keeping quality.

My friend Xiaomin runs a Xiaohongshu (Little Red Book) coaching business, single-handedly handling 80 clients. Last month she vented to me: "I asked AI to write 50 viral headlines — the first 10 were decent, the next 40 were total garbage." I had her compare against a domestic model. The gap was so big she almost switched tools.

Truth is: you think you're using "GPT-4." Server gets crowded, and you might silently get switched to GPT-3.5 — or even smaller.

What you can do today: 0 yuan, 30 minutes

Cost: 0 yuan.

Time: 30 minutes today.

Technical barrier: None. If you can copy-paste a prompt, you're set.

First step: Open your usual AI (pick one: ChatGPT / Claude / Kimi / Doubao), grab the 3 work questions you ask most often, and run each one through two different tools. The gap jumps out at a glance.

I also messed this up once: I had one AI write all my client emails. A critical email got "downgraded" too — almost lost the deal. Since then, for anything important, I cross-check with at least two tools.

If you don't try it now, that's fine — this article isn't pushing you to act. It's so next time your AI output tanks, you blame yourself one less time: it might have cut corners, not your prompt.

Advice by stage: no one-size-fits-all

Just starting out (no clients yet): If you're just dipping your toes into AI tools, free domestic ones (Doubao, Kimi, Wenxiaoyan) are plenty. Once you land your first paying client, then think about upgrading to a paid tier — by then you'll actually know what you need.

Got 1-2 clients: I'd suggest you do a "dual-model comparison test" today. Throw the same work prompt at two tools and see which one stays consistent. Keep monthly AI spend under 100 yuan. Don't get suckered by "Pro" plans yet.

Scaling up (5+ clients or team mode): AI quality swings hit your deliverables directly. Two things I'd suggest: 1) Use paid tiers for critical workflows (more stable, pricier), free versions for exploratory work; 2) Log every "AI output disaster" — at month end, check if certain time slots are the worst offenders.

One last thing: knowing models might be "downgraded" isn't meant to make you anxious. It's so that when AI output tanks, you doubt yourself one less time — it might have cut corners.