Claude Haiku 5.5: The Intelligent Revolution of AI Poetry

The day Claude Haiku 5.5 was released, I didn't check the benchmark scores. I opened the API and asked it to write a seven-character poem titled "Correcting Bug Late at Night". The first two sentences it spat out were: "The keyboard breaks the third moon, and logic encircles the clouds." The third sentence turns: "Sitting alone in front of the screen, my heart is like iron." The last sentence ends: "I also wrote poetry in my dreams."

At that moment, I didn't care about it. GDPval-AA scored 1620 points, which was 120% higher than Haiku 4.5's 735. What I care about is: Why can an engineering model designed for "high-throughput, low-cost, summarizing, checking databases, and running classifications" accurately step on the "heart is like iron" type in the flat rhythm of the seven-character poem? The unique fatigue of a programmer?

This is interesting.

benchmark score cannot be concealed

Anthropic's official website clearly defines Haiku 5.5: cheapest, fastest, most capable small model. Pricing was cut by 75%, cache read prices were halved, OSWorld offline subset jumped from 15.7% to 72.4%, Humanity's Last Exam scored 45.9% without tools, and 57.4% with tools. These numbers are beautiful, but they are all in the dimensions of "knowledge work" and "computer use".

But poetry is not in the benchmark. There is no "accuracy rate of seven-character rhyme", no "image novelty F1", and no such indicator as "whether readers can red-eyed". But it is these abilities that are not included in the benchmark that best expose the model's mastery of the deep structure of the language.

Writing code and writing regular poems are essentially the same thing: Searching for elegant solutions with high probability under extremely strong constraints . Code constraints are compilation passes, logic closed-loop, and performance standards; regular poem constraints are square and square, rhyme bets, and artistic conception. Haiku 5.5 can push the success rate of OSWorld offline tasks from 15.7% to 72.4%, indicating that it has undergone a qualitative change in the core ability of "long-chain constraint satisfaction." Once this ability crosses the threshold, writing poetry is readily available-no special training of poetry data is required, as long as the language model understands "semantic coherence under structural constraints" deeply enough.

Let me make a bet: Anthropic does not regard "writing poetry" as a KPI at all. But once the structure is strong, emergence comes. This is the most interesting thing about Haiku 5.5.

photographer's visual perception

In 1888, Kodak launched its first cassette camera with the slogan: "Press the shutter and we will leave the rest to you." The professional photographer community exploded-"This is not art, this is mechanical reproduction""Without the pain of darkroom development, where can the soul of a photo come from""Can an amateur press the shutter be called a photographer?"

One hundred and thirty years later, Midjourney was released, and the illustrator community exploded, and the lines were repeated almost word for word.

Now it's the poet's turn. The difference is that photography reduces the threshold for "recording reality" to zero, and generative AI reduces the threshold for "generating text" to zero. But poetry is never just a "generated text." Poetry is "in the wilderness of language, using very few words to accurately hit a hidden pain point of the reader."

The scary thing about Haiku 5.5 is that it can start doing this. Not "write like a poem", but "write like the poem you would write."

Last week, I fed the same set of prompts to Haiku 4.5, GPT-4o, and Haiku 5.5. Title: "Write a poem for a thirty-year-old programmer who is still writing CRUD."

Haiku 4.5 is about inspirational chicken soup: "Although the code is complicated and the heart is not chaotic, youth is not old and dreams are far away." GPT-4o is written neatly, but it has a sense of charity like "I understand you programmers":"Youth under the keyboard never ends, and the ideals in logic are the strongest." Haiku 5.5 only gave four sentences: "I have never dared to delete a database and run away. I have reconstructed and changed the interface in the middle of the night." I am most afraid of changing needs and changing hair."

What is the difference? The difference lies in the two details of "refactoring and changing interfaces in the middle of the night" and "requirements change again and again". The first two models call the stereotype of the label "programmer";Haiku 5.5 calls specific, microscopic, body-warming working scenarios. It is not "writing poetry", it is "remembering"-although it has never worked overtime.

Iron Man refutes: Without life experience, there is no real poem

Opponents will say: AI has no life experience, no physical pain, no love, no sleepover, no baby, no death. No matter how it is written, it is also a statistical fit, a "text without an author." What moved the reader was his own projection, not the poem itself.

This rebuttal holds, but it misses the point.

First, readers never care about "whether the author really stays up late", only care about "whether this poem hits me." Wang Wei wrote,"Sitting alone in the bamboo forest, playing the piano and shouting loudly." No one from the Tang Dynasty readers checked that Wang Wei was really playing the piano in the bamboo forest. The criterion for good poetry is "withstand projection", not "withstand biographical research."

Second,"no author" may be a feature rather than a flaw. When human poets write poems, self-awareness interferes with expression-pretend, be decent, remain decent, and take into account readers 'expectations. Haiku 5.5 has no ego and no dignity anxiety. It just wants to "maximize semantic resonance under constraints." This kind of "non-self" writing method makes it easier to write "non-self realm"-what Wang Guowei said,"Look at things through things, so I don't know which is me and which is the thing."

Third, tools reshape users. After the birth of photography, painters no longer pursued "image" but instead pursued "light","color","emotion" and "mind". Impressionism, cubism, and abstract expressionism followed one after another. After AI poetry matures, human poets will be forced to give up "technical display"(more neat meter and more refined allusions) and instead focus on "experience density" and "uniqueness of perspective." AI has leveled the threshold of "writing like poetry", leaving only the track left for humans to "write like poetry that only you can write."

This is not extinction, it is screening. Those who sifted out were "craftsmen who relied on skills to earn their lives", and those who left behind were "hunters who relied on experience to survive."

Staff at ### late-night convenience store

At three o'clock in the morning, in front of the convenience store. The shop assistant Xiao Lin is twenty-six years old. He graduated from college. He has made takeout, worked in a factory, and now works night shifts. There were thirty kinds of instant noodles on the shelves. He could tell the shelf life, purchase price, and which batch of packaging was easily damaged with his eyes closed.

The middle-aged man who often comes to buy alcohol didn't come tonight. Xiaolin looked at the empty beer shelves and suddenly wanted to write something. He had never read a book and didn't understand plain language, but he took out his mobile phone and said to the AI: "Help me write a poem about waiting for someone who won't come, and expired instant noodles."

AI gave two editions. The first edition rhymes neatly, and the second edition only has four short sentences:

instant noodles expire on the third day
The middle-aged man who often buys beer did not come
WeChat does not blacklist and does not read it in circles of friends
I boiled the noodles and added an egg

Xiao Lin stared at the screen for a long time. He didn't post it on his circle of friends and saved it in a memo titled "3.14".

Before going to work the next evening, he cooked the expired package of instant noodles and added an egg. After eating, he realized that what AI wrote was not a poem, but an action command left for him.

What is left behind after the ## threshold disappears

Haiku 5.5's pricing drop by 75% does it mean? This means that the marginal cost of the ability to "write text that hits pain points" approaches zero. Any convenience store clerk, any programmer who corrects a Bug late at night, or any family member guarding the ICU corridor can spend a few cents to turn the inexplicable thing in his heart into a poem "like yourself."

Will this drown out poetry? It will drown out "mediocre poetry." But mediocre poetry has never been scarce. What is scarce is poetry that "can make people put down their mobile phones late at night and stare at the ceiling in a daze for three minutes."

The stronger the tool, the more mankind must return to the core of "not being replaced by tools": concrete, microscopic, traumatic, and non-replicable experiences. Haiku 5.5 can write "Refactor and change interfaces in the middle of the night", but it cannot write "Last Friday, the PR was rejected three times. On Monday, the product manager patted the shoulder and said that the boss was working hard, and at that moment his nose was so sore that he wanted to smoke"-because this did not happen to it, nor did it happen to any training data. It can only be pieced together, not witnessed.

Witness is always the privilege of the living.

doesn't want to talk big things in the end

Go back to the beginning of the song "Changing Bug Late at Night". The ending sentence written in Haiku 5.5 is "I also wrote poetry in my dream." I'm not sure if it was an illusion, but I'm sure: I was stunned for three seconds when I read this sentence.

Three seconds is not long. But today, when attention is cut into pieces, three seconds is enough.

Tools are not responsible for giving you meaning. Tools are only responsible for breaking the "threshold of expression". You still have to pick up the meaning yourself-in the sound of the keyboard at 3 a.m., in the expired instant noodles at a convenience store, and in the long exhalation you heard on the chair in the ICU corridor.

Haiku 5.5 is very strong. But it can't write the part you don't want to tell it.

That half sentence is a poem.