I Used Four AI Models to Write a 10,000-Word Article in One Night. Here’s What It Cost Me.

At 11:07 p.m. last Tuesday, I pasted a 10,000-word brief into my writing tool and hit go. Half an hour later, a red bar appeared: quota almost gone. I scrolled the draft. Correct. Boring. I deleted a sentence. Typed it again. Deleted it. The cursor blinked. I made coffee. Forgot it.

I closed the tab.

The next morning I tried a different split. One model would plan. Another would draft. A third would polish. A fourth would do titles and summaries.

I started a new session. Loaded the source material. Set the reasoning level to medium. I asked the planner for an outline, character relationships, emotional arc, and structural anchors. It gave me a clean skeleton. I saved it. I did not ask it to write the article. Its prose is over-polished and hollow. It costs more for the same script. It is a planner.

I closed that session. Opened a new one. Gave the drafter the outline. Set the reasoning level to medium. It wrote the full draft. It traced back through settings. It kept logic consistent across tens of thousands of words. The draft had fewer of those over-polished sentences that sound like a textbook. In many places, it was good enough to polish, not rewrite. It also benefits from large cache-input discounts. That matters when the context keeps growing.

Then the editor. I copied the draft into a new session. Set the reasoning level to light. Asked it to scan for banned words. Replace vague statements with concrete actions. Turn “he felt” into “he crushed the cigarette against the rim of the cup.” It costs about half as much as the drafter. It is an editor, not a narrator.

The formatter stayed on the sidelines. Titles. Formatting. Extracting data. Converting files. Cheap and fast. Not for the core of a long article. It cannot hold a complex narrative together.

I made one mistake. I switched from the planner to the drafter in the same conversation. The quota bar dropped. I closed the tab. Now I have the planner generate a handoff document. Then I start a new session. The drafter reads the document. The context rebuilds in a controlled way.

There is also a long-context fee. Over a certain length, most models charge a higher rate for the entire context, not just the portion above the threshold. The planner has an official optimization that avoids this extra charge. The drafter and others do not. So I keep an eye on context length. If I need to explore a side idea, I open a side discussion without polluting the main thread.

Reasoning effort: planner medium for most planning. High if the structure is complex. Maximum costs about 25% more than medium. The marginal benefit drops quickly. Drafter: medium for drafting. High for deep argumentation or complex logical relationships. Do not switch frequently. Every switch can invalidate the cache. Editor: light or medium. It is doing editorial cleanup, not building the argument.

A 10,000-word Chinese article is about 15,000 output tokens and 50,000 tokens of context. The planner costs little. The drafter costs a medium amount. The editor costs little. If I use the planner for the entire article, the output cost alone is about 1.8 times the drafter. I also spend extra time removing the AI flavor.

My current setup is simple. Planner medium for the framework. Drafter medium for the first draft. Editor light for polishing. Formatter for titles and summaries.

That night, I wrote the last line at 3:14 a.m. The coffee was cold. I drank it anyway. Then I set the alarm for 7.