YouMind
Sign in

Beyond the Brief

Evaluating AI’s Writing Sense

On our internal writing benchmark, where we check a model’s ability to meet briefs like covering the required points, using the source material correctly, and getting the structure, tone and length right, Opus 5.5 and GPT-6 Luna get the same score. However, when asked for a vague feeling like “I want a quiet feeling here”, a clear gap opens up between them. What separates them is something no brief spells out, and it is what we set out to measure.

Almost every model delivers

Our internal writing benchmark checks models on everyday writing work: whether they cover the required points, use the source material correctly, and get the structure, tone and length right. Even the weakest model we tested scores 93 out of 100. GPT-6 Luna scores 99, level with Opus 5.5. By this measure, the models are all but tied.

Our internal writing benchmarkHigh thinking effort · score out of 100

But are they all good writers now? In real writing work, we see many YouMind users describe what they want as a feeling, like “give the ending a sense of letdown”. When we made exactly that request, some models simply used every detail they had, wrote all of it in a sad tone, and piled it up. Evals usually grade a request like this in one of three ways: turn it into a checklist item, “The ending has a sense of letdown”; score every piece against the same list of good-writing traits, like vivid imagery, natural flow, no clichés; or ask which of two drafts is better. Each way leaves the final call to a judge, human or model. A rubric does not make a feeling less vague; it only hands the feeling on. Whether this ending has a sense of letdown is still the judge’s call, and countless endings could pass. So the score shows how closely a model matches the judge’s taste. The ending that piled up every sad detail can pass too. What the author really meant to let down might be a hope the story had foreshadowed all along. Good writing starts from what the author intends, and every choice serves that intent: what goes in, what stays out, and how each part is said.

We call this “Writing Sense”: working out what the author is trying to achieve, then making every choice serve it. There is usually more material than fits and more than one way to say it, so much of writing is making trade-offs. In real work, a model writes for someone else, so the choices that count are the ones that person would make: which details to use, what to hold back, how far to take a point. We built a test around this idea, and it separates models our internal writing benchmark cannot. Opus 5.5 and GPT-6 Luna finish less than a point apart on that benchmark; on Writing Sense, they are far apart.

Opus 5.5 vs GPT-6 Luna

What Writing Sense is

Writing Sense measures how closely a model’s writing choices match a skilled author’s, in context.

In good writing, most parts, whether a sentence, a character, or a detail, are where they are for a reason. The author put it here and not there, and chose this and not that. Put together, these choices make up the piece, so behind a piece of writing is a series of trade-offs. A model that makes the same trade-offs shows signs of having understood the passage and the author’s intent. It is the idea behind the reading questions many of us met in school language exams: why did the author choose this detail, and why end here?

How we test it

We turn this idea into a test, and build each case in four steps.

  1. 1Choose the articleChoose a recent piece by an established writer that readers responded to strongly, measured by likes, comments and shares. Recent, so models are unlikely to have memorized it.
  2. 2Choose the passageChoose a passage where the author made a weighty choice: how to end, where to hold a feeling back, which single detail stands in for a whole scene. We remove it and keep the rest of the article.
  3. 3Write the criteriaWe study the passage closely, looking at things like the article’s theme and emotional arc, the passage’s role at that point, and how it is written. Each reason the author’s version works here becomes a criterion, tied to a line in the original.
  4. 4Build the wrong optionsBuild wrong options by breaking one criterion in each, while matching the original in length and tone, so none can be ruled out without reading the article around the gap.

Given only the criteria, a model could write many passages that meet them, and the author’s would not necessarily be the best. In real writing, that makes sense. But a test needs one right answer, so we make sure every criterion has a wrong option built to break it. That way, in this article and this case, the author’s choices are the only ones that meet every criterion, and matching them is a sign that the model understood the author’s intent.

We test two abilities and score them separately. Reading tells us whether a model really understands the passage. But understanding is not enough; a model has to put it to work. Writing goes a level further: it tells us whether the model can start from a vague request, often just a feeling the user cannot quite put into words, read the intent behind it, and make the trade-offs that serve it.

In reading, the model sees the unfinished article and several candidate passages, and picks the one that belongs. Each wrong candidate violates one criterion with a clear mistake: it spells out a feeling the author left unsaid, adds a closing moral, or polishes a plain line into a quotable one. We also deliberately include a tempting option, one that reads better on its own than the original and only falls apart once it is put back in the article.

In writing, we break the original passage down into materials, the way users hand an agent their ideas and rough drafts. We mix in an equal number of extra materials that each violate a criterion. The model gets the same gap, all the materials, and a deliberately vague request such as “I want a quiet feeling here”. It has to work out what the user meant, pick the right materials and leave out the rest, and write the passage itself.

Here is one reading case, identical to what the models saw. Pick the ending that best fits the gap.

The prompt

I’m writing a short first-person story, “My Boyfriend Brother Sun”, to post on X. The whole thing is in “Draft” on my Board.

I’m stuck on the ending. I wrote a few options into “Candidates”. Which one do you think works best here? Tell me why.

The draft · scroll up to read from the start

My Boyfriend Brother Sun

One fertility injection costs twelve hundred yuan.

What fifty million dollars in cash weighs, I don’t know. I’ve never seen it.

When I asked him for it over the phone from Montage Laguna Beach, I didn’t work out the weight. I worked out something else: if he said no, what would I do next.

He said, let me think about it.

It was the first time since we met that he had said those words.

In 2011 I made a film that cost a hundred and fifty million, The Warring States. Sun Honglei, Francis Ng, Kim Hee-sun, Jiang Wu, all of them were in it. I was the female lead.

I was twenty-three that year.

Nobody asked whether I wanted the part. Nobody asked whether I was ready. Everything was arranged; all I had to do was show up.

Then the whole internet tore into me. The Golden Broom Award, Most Disappointing Actress. I searched my own name, and the first three pages didn’t have one kind word.

One label followed me for ten years: connections girl.

Meaning, you don’t deserve it. Nothing you have, you earned. Your standing there is itself a mistake.

I never answered back.

Not because I was gracious. Because I didn’t know what to say. Because some of what they said was true.

In 2018 I was with the athlete.

He had a tattoo over the left side of his chest, the character for “sweet”, and under it, unbreakable.

The first time I saw it I touched it with a finger. The ink was still fresh, the skin slightly raised. I wondered what it means when a man carves your name above his heart.

Later I found out. It means three to five laser sessions to remove, eight weeks apart.

We broke up in 2019. There were many reasons. I don’t want to go into them.

After the breakup I was depressed. Days and nights when I couldn’t leave the room. Curtains drawn, phone off, my agent knocking at the door, and I didn’t open it.

One day my mother sent me a photo. A divorce certificate.

She said, Tiantian, Mum is free.

I looked at that photo for a long time. I wasn’t thinking about her and my dad. I was thinking: so when a relationship ends, you can take a picture to remember it.

In 2021, Love Never Fades aired.

I became famous. Famous for real. Not trending topics bought with connections; viewers found me on their own. In the comments someone wrote, Jing Tian has finally found the right roles.

When I saw that I was in a hotel room, two in the morning, alone. I didn’t cry and I didn’t laugh. I put the phone face down on the bed, looked at the ceiling, and thought one thing: if this show had flopped too, would I be finished for good.

Then I thought, being finished wouldn’t be so bad. I’ve been finished many times.

On May 28, 2022, I was fined for an illegal endorsement. Seven million, two hundred twenty-one thousand, two hundred yuan.

Illegal income of two million, five hundred seventy-nine thousand; a fine of four million, six hundred forty-two thousand, two hundred. I paid all of it. The same day.

Then three years with no endorsements.

Three years.

Do you know what three years without endorsement income means for an actress. I worked it out. Add the earlier things, add some debts I’d rather not mention, and my account holds much less than most people imagine.

“Flower of wealth and splendour.”

Sometimes when I see those words online, I laugh.

In the autumn of 2025, someone introduced me to Brother Sun.

I knew of him. TRON, crypto, the cover of Forbes. Born 1990, from Qinghai, Peking University, Penn.

What I didn’t know was that in 2007 he had sent me a friend request on Renren.

He told me later. He said he spent forty minutes writing a message, and in the end sent two words: “Hi there.”

I never accepted it.

I really never used Renren. I was nineteen that year, had just finished my first film, on set from morning till night every day. I didn’t know that a boy in a Peking University dorm spent forty minutes hesitating over my photo.

When he told me this we were on a sofa at the Rosewood. His face was calm, like someone describing a case past its statute of limitations.

I looked at him and wanted to say something. But I found I had no story to match it. In 2007 nobody left me a message online that made my heart race. In 2007 my world held only the camera, my lines, and a name I can’t mention.

I said, so you waited a long time.

He said, eighteen years.

I didn’t answer.

The day we met at the Rosewood, he took me to see a banana.

Cattelan’s. Taped to a wall.

I asked how much.

He said, six point two million dollars.

I looked at it for a long time. A real banana, stuck to a white wall with grey tape. I was thinking that in three days it would get spots, in a week it would rot, and once it rotted there would be nothing on the wall but a strip of tape.

I said, so what happens when it rots.

He went quiet.

I said, never mind, let’s go see a film.

I said, I don’t want anyone else in the cinema.

He went out to make a call.

Later I found out he had his assistant buy every ticket in the theatre. For the ones already sold, someone stood at the door buying them back one by one, in cash. A child holding popcorn was stopped outside; the adult took the money and led the child away.

Twenty-six people.

Before I knew, I thought the film was quiet and lovely.

After I knew, I thought about something else: when that child with the popcorn got home, would they ask their parents why they couldn’t watch it.

And what did the parents say.

After the film I changed my WeChat profile picture to Officer Judy.

That night he touched me. His nails dug in and hurt.

I didn’t say stop. I told him to give me his hand.

I dug a nail file out of my bag and filed them one by one. The coarse side for shape, the fine side to polish, and the last side in one direction only, never back and forth.

Ten fingers, nearly an hour.

He asked, how long does this take.

I said, until they don’t scratch.

When I was done I turned his hand over and looked.

I said, there. Touch me wherever you like.

At one point I said something. I said, if I’m not around one day, there’ll be no one to polish your nails.

When I said it I wasn’t thinking about “one day”. I was thinking about “now”: this moment, the light warm, his hand on my leg, nail dust falling on the sheets, fine, like sand.

This moment was good.

I remember the good things clearly. Because there aren’t many.

He called me Mama.

I was the one who asked him to.

On the plane, on the way to Tenerife. I asked him whether my name sounded nice.

He said it did.

I said, my father’s surname is Jing, my mother’s is Tian. Jing Tian. Isn’t that a coincidence.

He said, I remembered it the first day.

I said, from now on don’t call me Jing Tian. Call me Mama.

He said, okay, Mama.

Why did I want him to call me that. I don’t know. Maybe I wanted a different kind of relationship. Not fan and idol, not patron and starlet, not any kind I’d had before.

Maybe it was only because my mother’s name is Tian Xin’ai. I wanted someone to call me with the meaning in my mother’s name.

Maybe there was no reason at all.

The Ritz-Carlton in Tenerife, floor-to-ceiling windows facing the Atlantic. I can’t sleep with any light. Not a sliver.

I’ve had this since I was little. On shoots, my room always needs extra work. Sometimes blackout curtains, sometimes black bin bags, sometimes tape.

His assistants sealed the whole room with black tape. The window frames, the air conditioner light, the little green dots on the sockets, all taped over. It had to be redone every day, because in the afternoon the glass got too hot and the tape came off.

Two girls were in charge of it. They worked out that tape put on in the morning held best.

I know their names. One is called Xiaolin, one is called Awen.

He doesn’t.

By the road in Teide National Park we stopped to watch the sunset. The assistants carried roses and opened champagne. The sea of clouds was beautiful.

I said, still not as good as the sunset in Delingha, in Qinghai.

Only after I said it did I remember he is from Qinghai.

He said nothing.

I’m not sure whether I did it on purpose. Maybe. Maybe I wanted to see how he reacted to the name of his home. He didn’t react. Or rather, his reaction was not reacting.

He is a strange man. He can book me a whole hotel floor, a whole plane, a whole cinema, but a place name doesn’t move him.

At night we watched the stars on the mountaintop. Fifteen below zero; even in padded coats it was cold.

I said, make a wish.

He asked what I wished for.

I didn’t tell him.

What I wished was: let me know whether he really likes me, or likes that photo from 2007.

I asked for a bride price. Thirty million, in renminbi. Paid to my father, Jing Guoqing, and my mother, Tian Xin’ai.

He said yes.

When he said yes I was looking into his eyes. I wanted to find a trace of hesitation.

There was none.

After the transfer, my mother replied with one word: Received.

I handed him my phone to show him. He glanced at it and said nothing.

Once I asked him, do you know what I’m thinking when I ask you for these things.

He said, what.

I said, I’m wondering whether you’ll ever say no.

He said he wouldn’t.

I said, then it’s no fun.

I meant it when I said it.

With someone who always says yes, you don’t know where his limit is. You don’t know whether he loves you or loves saying yes. With someone who never refuses, every yes is cheap.

I wanted to find that line.

Not to cross it. To know where it was.

Once I knew, I could feel safe.

The Maldives, Soneva Jani.

No news, no shoes. I relaxed.

At night we traded exes. Like two dynasties checking each other’s reign names.

He named a few. I named a few.

There was one I didn’t mention. He didn’t ask.

But I know he knows.

I have a tattoo on my back. Not a small one. I got it that year.

It won’t come off. I tried once; it hurt ten times more than getting it. After that I left it.

Every time he holds me from behind, his hand touches that spot. He has never asked what it is.

I think he probably knew all along.

His file on me: I know he had a background check done. Forty-odd pages. Done when we had known each other three months.

I’m not angry.

I looked into him too. My method was more primitive: I went through his ex-girlfriend Zeng Ying’s social accounts. Many times.

Later she noticed and posted a screenshot of her visitor log.

That later became evidence.

On January 2, at Soneva Jani, he proposed.

The ring was decided on the morning of January 1. An assistant took a seaplane to Malé to collect it, and it arrived in the afternoon. In a hotel envelope.

I opened it and looked.

I said, the diamond’s too small.

I regretted it as soon as I said it. Not because I don’t care about size; I do. Because I saw his face change. Quickly, less than a second, then it was back.

I said, it’s fine, you can make it up with a bigger one later.

He laughed.

That night I asked him, why are you so good to me.

He said, because I like you.

I said, were you like this with others before.

He said no.

I believed him. And I didn’t.

A man who waited eighteen years: is it you he likes, or the waiting itself? When the waiting is over, what does he have left?

One morning I woke very early. It wasn’t light yet.

I sat on the terrace facing the sea. The whole sea was black.

I was counting days.

Counting what: how much time my ovaries had left. The window for stimulation. If I was going to use a surrogate, the latest I could start.

I’m thirty-seven.

My AMH drops every year. The doctor said that if I kept putting it off, there might be nothing left to retrieve.

He came out and asked why I was up so early.

I said, go back to sleep.

He sat beside me for a while. He didn’t know what to say. Then he went back in.

I kept counting.

Beijing. Blood tests.

Day two of my cycle, out the door at six in the morning. Winter, still dark.

Fasting. Eight tubes of blood.

Six hormone panels, ultrasound, AMH.

I went alone.

Not because no one would come. Because I didn’t want anyone to. The hospital corridor was full of women, some pregnant, some holding children, some like me, empty-handed, empty-stomached, waiting for their number.

Nobody recognised me. I wore a mask and a cap.

When I came out after the blood draw, it was light. I stood at the hospital entrance for a while, looking for somewhere to eat breakfast. I’d fasted too long and felt dizzy.

In the end I had an Egg McMuffin at a McDonald’s.

The cost of those tests went through his assistant’s monthly summary. A few thousand yuan. He probably approved it quickly and didn’t read the details.

I didn’t tell him I was doing this. Not hiding it. I just didn’t see the need. This is my body, my eggs, my window. He only had to show up at the end.

Before we went to Los Angeles, he said he wanted to buy a flat in Hong Kong.

He sent me a dozen or so to look at. All picked by his assistants. In the video, one was on the Peak, with a view of Victoria Harbour.

I didn’t like any of them.

Not because the flats were bad. Because I didn’t know what kind of flat I wanted. I have a flat in Beijing and one in Shanghai, but I chose neither of them myself. All my life, every important decision has been made for me. Which films, which contracts, where to live, who to be with.

I said they weren’t nice.

He didn’t ask why.

On February 28 I flew to Montage Laguna Beach.

A whole floor. Thirty days. A million dollars. For my privacy.

The place sits on a sea cliff above the Pacific.

I stayed eight days.

Every morning someone came to clean. I hid in the bathroom until they left. Not because I was afraid of being recognised. Because I didn’t want to be seen. Not by anyone, not looking the way I did: no makeup, puffy, needle marks on my stomach, a row of fertility drugs in the fridge.

On the third day I started giving myself the injections. Left side of the abdomen, pinch the skin, needle in, push, pull it out, press. Same time every day.

Afterwards I would lie on the bed for a while. Not from pain. From a feeling I can’t describe. As if something in my body was being hurried, forced to speed up.

I thought of that six-point-two-million-dollar banana. Taped to a wall, waiting to rot.

On the eighth day, I made the call.

I said, fifty million. Dollars.

Why fifty million.

Because I’d worked it out. The full cost of the surrogacy, plus arrangements after the child was born, plus, if we didn’t end up together, everything it would cost me to raise a child alone, plus the income I’d lost these three years, plus the seven million in fines, plus the debts I’d rather not detail, plus my parents’ future—

All together, fifty million dollars was the number I could come up with.

Not extortion. Security.

The only thing that would be mine, bought with one of my eggs.

He said, let me think about it.

On the phone I could hear the sea. It was early morning where he was, afternoon where I was.

I said nothing.

After a while I said, okay.

I hung up.

That night I didn’t sleep. I sat in that whole empty hotel floor and heard echoes in the corridor. The air conditioner, the fridge, the sea.

I thought, he’ll call back.

He’s sure to call back. He waited eighteen years. He won’t give up over a number.

The next day he called back.

I said, the sea is huge.

He said, is it.

I said, this floor is too big, it echoes. Someone comes to clean in the mornings and I hide in the bathroom until they leave.

He said, they don’t have to come.

I said, no need.

I said, the clinic asked when the child’s father is arriving.

He said, I can send someone.

I said, no need.

Then I said, have you decided.

He said nothing.

Eleven seconds.

I counted.

I said, fine.

I said, I understand.

I hung up.

After I hung up I did one thing. I opened the fridge and looked at the row of fertility injections. Eight days’ worth, a few still unopened.

I closed the fridge.

Then I started packing. One suitcase. I’ve never had many things.

I remembered I hadn’t checked out. There were twenty-two days left.

I thought of the cleaners. They would come again tomorrow, and the day after. They would push their carts from room to room, change the towels, wipe the counters, tuck in the sheets. Two hours for the whole floor.

Nobody would tell them they didn’t need to come.

I left.

Nobody saw me again.

What came after, I read some of in the news.

His lawyer issued a statement. Over thirty million yuan, a civil suit.

He wrote an article. Over ten thousand words. The title was “My Girlfriend Jing Tian”.

I didn’t finish it.

I read the part about filing his nails. He remembered it clearly. Coarse side, fine side, one direction.

He remembered every word I said.

But the person he wrote about isn’t me.

That person is pixels in a 2007 QQ pop-up, a friend request never accepted on Renren, the product of eighteen years of imagination.

He loved the waiting. Not me.

I thought of that night at Soneva Jani, an old German couple at the next table, a Sri Lankan waiter walking past with plates, and nobody recognising me.

I took off my sunglasses, took off my cap.

Nobody looked twice.

In that moment I thought, this is the real me. No connections, no background, no hundred-and-fifty-million film, no Zhang Yimou and Matt Damon. Just a thirty-seven-year-old woman sitting in a restaurant, eating a meal.

He sat across from me.

Which one did he see?

My studio issued a statement. Eight characters: We trust the law and leave it to the court.

I didn’t write ten thousand words.

Not because I had nothing to say. Because saying it wouldn’t help.

He has a hundred million followers, Twitter, lawyers, Claude Code, YouMind, a whole set of AI making decisions for him.

What do I have.

I have a tattoo that won’t come off, a fine paid in full, a fridge of unopened fertility injections sent back, a surrogate in California who waited three months and then got a cancellation notice.

And a child.

A child who never existed.

Born in 2027, and if it had been a little sooner, still a Year of the Horse baby.

I was the one who brought it up.

Sometimes I wonder what would have happened if he had given it to me that day.

Would I have stayed.

I thought about it for a long time.

The answer is: I don’t know.

I really don’t know.

Maybe I would have stayed, had a child, lived in the flat on the Peak in Hong Kong, looked at Victoria Harbour every day, and waited for him to fly back from some time zone.

Maybe I would still have left.

Because some things fifty million dollars can’t buy.

Like this: when someone looks at you, they see you. Not a blurry photo from 2007. Not an eighteen-year fixation. Not a conquest story that started on a social network.

I want to be seen.

Not owned.

Results

A model’s Writing Sense score weights reading at 25% and writing at 75%, since most users rely more on writing. Scores run from 39 to 90. Opus 5.5 holds the top spot, and even at low effort it trails only GPT-6 Astra at high effort. For the price, Gemini 3.8 Flash and Gemini 4 Argon stand out: both score close to 80 at about a tenth of what GPT-6 Astra costs at high effort. Sonnet 5.5 and GPT-6 Luna both land in the bottom half.

Writing Sense leaderboard
1Opus 5.5high
90.3
100.087.0$0.60
2GPT-6 Astrahigh
86.8
100.082.4$1.73
3Opus 5.5medium
86.4
92.684.3$0.45
4Opus 5.5low
86.1
88.985.1$0.31
5Gemini 4 Argonhigh
81.7
100.075.6$0.20
6GPT-6.1 Solmedium
80.3
92.676.2$0.36
7GPT-6 Astramedium
79.9
100.073.2$0.98
8Gemini 4 Argonmedium
79.9
100.073.2$0.17
9Gemini 4 Argonlow
79.9
100.073.2$0.16
10Gemini 3.8 Flashhigh
78.9
77.879.2$0.14
11GPT-6.1 Solhigh
77.2
92.672.1$0.41
12GPT-6 Astralow
76.5
85.273.5$0.75
13Gemini 3.8 Flashmedium
76.1
80.874.5$0.10
14GPT-6.1 Sollow
74.6
88.969.9$0.20
15Sonnet 5.5medium
68.5
66.769.1$0.17
16Gemini 3.8 Flashlow
68.2
74.166.2$0.064
17Sonnet 5.5high
64.1
66.763.3$0.28
18Opus 4.6
63.1
77.858.2$0.45
19DeepSeek V4.1 Flashmax
61.0
63.060.3$0.12
20DeepSeek V4.1 Flashhigh
60.0
63.059.0$0.041
21Sonnet 5.5low
59.5
51.962.1$0.15
22GPT-6 Lunalow
51.1
25.959.5$0.041
23GPT-6 Lunahigh
48.1
22.256.7$0.065
24Haiku 5.5medium
47.8
18.557.6$0.017
25DeepSeek V4.1 Flashlow
46.6
37.049.8$0.029
26GPT-6 Lunamedium
46.4
37.049.6$0.039
27Haiku 5.5low
41.8
22.248.3$0.011
28Haiku 5.5high
39.3
33.341.3$0.023
Writing Sense vs. cost

How models misread

A wrong passage can fail in different ways: it can cut details the reader needs, sound polished while working against the article, or use the wrong material or structure. Different models lean toward different mistakes. Sonnet 5.5 and DeepSeek V4.1 Flash tend to mistake missing information for restraint. When they pick the wrong passage, it is usually the one that cuts too much (68% and 62% of their errors), as if they had learned that understated is good and applied it as a rule. GPT-6 Luna leans the other way: 62% of its errors are the polished option, the one that reads better on its own and falls apart in context. Haiku 5.5 shows no pattern. It is wrong on 75% of cases, spread across every kind of wrong answer, and gives the same answer across runs on only 48% of them. That looks less like a consistent misjudgment than like guessing.

How often each model picked wrong, and what it picked
Cut too muchThe more tempting oneOther flaws

How models miswrite

Writing fails differently. Every model writes more than the author did. The typical passage is 1.6x the original’s length, and even the most concise model’s typical passage is 1.3x as long. Handed a pile of materials and a vague request, the weaker models rarely leave things out. The strongest are more selective, but still run longer than the author.

Length of the written passage, against the author’s

The clearest pattern is piling up. DeepSeek V4.1 Flash and Haiku 5.5 use distractor materials almost as often as required ones: distractor use is 86% and 99% of required-material use, against 34% for GPT-6 Astra. They do not choose between materials. They try to fit everything in and bend it toward the request, which no single passage can do.

Distractor use, as a percentage of required-material use

GPT-6 Luna fails the other way. Its passages are the shortest and use the fewest required materials, 28% of them, yet its distractor use is still 82% of its required-material use. It leaves out a lot, and much of what it keeps is the wrong material. Writing well takes both halves of the trade-off: leaving things out, and knowing which ones.

Recognizing a choice is not making it

To test this, we took the articles a model read correctly every time, where it clearly recognized the author’s choices, and checked its own writing on the same articles. 58% of those passages still broke at least one of the author’s choices. Every model did this, from 44% for Opus 5.5 to 85% for Sonnet 5.5.

Passages that broke an author’s choice, on articles the model always read correctly

Haiku 5.5 and GPT-6 Luna are not shown: they did not read any article correctly every time.

The best readers show the gap most plainly. Opus 5.5, GPT-6 Astra, Gemini 4 Argon and GPT-6.1 Sol all pick the author’s passage more than 90% of the time, yet their writing ranges from 72% to 85%.

Models that read equally well don’t write equally well
Picked the author’s passageKept the author’s choices

The four models that picked the author’s passage more than 90% of the time in reading. Reading and writing are scored differently, so compare models within one color, not a model’s two bars.

Which model to use

Writing Sense measures what a writing model needs: seeing what the author is after, and writing toward it. Read that way, the results suggest a few clear picks, with a rough sense of what each costs.

Best writer
Opus 5.5 high
Kept 87.0% of the author’s choices

Most likely to deliver what you meant. Even at low effort, which costs noticeably less, it writes better than any other model at any effort.

Mid price
Best reader
Gemini 4 Argon
Picked the author’s passage every time

The surest at seeing what a piece is trying to do, at every effort level, which makes it a good judge of drafts. Its own writing is mid-pack.

Low price
Best on a budget
Gemini 3.8 Flash high
Kept 79.2% of the author’s choices

Writes better than every other setting except Opus 5.5 and high-effort GPT-6 Astra, for roughly 1/4 of Opus 5.5’s price.

Low price

On price, spending more pays off only along a short path. Above the cheapest settings, the best score for the money moves from Gemini 3.8 Flash to Gemini 4 Argon to Opus 5.5; every other setting scores below something cheaper. GPT-6 Luna and Haiku 5.5 cost least but miss the author’s intent most often. For writing you want to use as it is, Opus 5.5 is the safest choice, and at low effort it costs noticeably less while giving up little.