Opus 5.5 causing an identity crisis is real

A repeat of Piotr Migdał's one-shot Claude Code experiment building a three.js visualization of Italo Calvino's 55 Invisible Cities cost $14.98 at medium effort versus $264.64 at max effort with ultracode, a 17.7x difference, according to results published by blogger Ktalik. The max-effort run used 5,326,576 Opus output tokens (3,655,322 of them thinking), 242 web searches and 16h 29m of summed API time across about 11 parallel city-builder agents, adding 31,387 lines of code versus 7,746 at medium effort, yet never got mouse input working without errors. The author says Migdał's same-parameter result was slightly better visually but broadly similar, and released all source code and agent transcripts in a public repository.

Piotr Migdał is right in his post opens in a new tab https://quesma.com/blog/invisible-cities-one-shot/ . In his experiment, he uses a short, one-shot prompt: Make a three.js pnpm visualization of all Invisible Cities by Italo Calvino. Don’t ask questions, it is a one-shot task. You have 6h of work, use it until it becomes a masterpiece. Invisible Cities opens in a new tab https://en.wikipedia.org/wiki/Invisible Cities is a postmodern novel by Italo Calvino, made of short, poetic descriptions of 55 imaginary cities, each one complex and often very abstract. In 2019, Piotr worked on a small project that generated those cities for a storytelling performance, using the state-of-the-art model of the time, GPT-2. To test what today’s models can do, he wrote the short prompt above, and the result was stunning. Looking at the output, Piotr, a data vis specialist opens in a new tab https://p.migdal.pl/ , asks himself: What is my place in creating interactive media? And, after following his call to action see post , I have one simple answer: yes. Here are my results from repeating Piotr’s experiment. Click on the picture to go to the interactive version Detailed results of my Claude Code + Opus 5.5 runs in the two configurations: | Metric | medium effort | max effort + ultracode | Ratio max / medium | |---|---|---|---| | Total cost | $14.9800 | $264.6400 | 17.7x | | Opus 5.5 cost | $14.9800 | $261.4700 | 17.5x | | Haiku 5.5 cost | $0.0001 | $3.1700 | - | | Metric | medium effort | max effort + ultracode | Ratio max / medium | |---|---|---|---| | Opus input tokens | 402 | 637,693 | ~1,586x <sup 1</sup fn1 | | Opus output tokens | 313,168 | 5,326,576 | 17.0x | | …of which thinking | 69,578 | 3,655,322 | 52.5x | | Opus cache read | 27,899,390 | 297,377,614 | 10.7x | | Opus cache write | 392,035 | 18,013,330 | 45.9x | | Haiku input / output | 1,215 / 23 | 5,668,132 / 374,003 | - | | Web searches | 0 | 242 | - | | API time summed | 0h 53m 4s | 16h 29m 10s | 18.6x | | Wall time active | 2h 46m 45s | ~1h 50m 12s | 0.66x | | Lines added | 7,746 | 31,387 | 4.1x | Some notes: - Wall time: the run hit the session limit on Oct 9 at 14:19 UTC and resumed on Oct 13 at 12:06 UTC. The ~1h 50m adds the two working stretches prompt to limit: 1h 4m 2s; resume to final answer at 12:53: 46m 10s and leaves out the ~94h pause. The resumed stretch picks up with agents still building, and the log can’t tell whether any work happened during the pause, so this figure is approximate. - API time: it adds up all parallel agents about 11 city builders plus research agents . That’s why it is far higher than the wall time. I hope you like the visualizations. I’m not an Invisible Cities expert myself, and definitely not a data vis specialist. So, I’ll leave the review of the details of both artifacts to people who know more about those areas. What I can say for sure is that Piotr’s result while using the same parameters turned out a little better visually than mine, but in principle the results are very similar. The max effort + ultracode version is even more stunning, but interestingly it never got mouse input working without errors. If you want to test it out, you need to stick to keyboard controls - see the Use chapter in the project’s README file. And hey, if you are interested, the whole source code is in the repository, with the agents’ transcripts included Claude Opus 5.5, medium effort, no ultracode: \ interactive\ opens in a new tab https://ktalik.github.io/invisible-cities-2026.10/cc-2.1.296-opus-5.5-medium-no-ultracode/ , \ code\ opens in a new tab https://github.com/ktalik/invisible-cities-2026.10/tree/master/cc-2.1.296-opus-5.5-medium-no-ultracode/run-1 Claude Opus 5.5, max effort, with ultracode: \ interactive\ opens in a new tab https://ktalik.github.io/invisible-cities-2026.10/cc-2.1.296-opus-5.5-max-with-ultracode/ , \ code\ opens in a new tab https://github.com/ktalik/invisible-cities-2026.10/tree/master/cc-2.1.296-opus-5.5-max-with-ultracode/run-1 But that’s not all. If you find this interesting, take up Piotr’s call to action: run the prompt on your own models and agents, and share the results Of course, the title of this post is catchy. There is no real proof of Opus 5.5 causing an identity crisis for anyone. It’s just a real phenomenon that GenAI, in general, is causing it for some people already, and it started well before 2026. I don’t know if Piotr identifies with the crisis, but I know that identity crisis is not a joke, Jim . The picture is a funny meme, but the identity crisis is a serious thing. In my bubble IT , the issue of some of the programmers being upset with LLMs entering the game is pretty well known. However, not everyone realizes that what we observe is an identity crisis. I watched numerous YouTube videos and read many social media posts and blog posts by programmers about a profound sense of alienation, loss of human dignity, and anger over stolen independence. And you know what? You can find the same in texts from the 18th and 19th centuries - the Industrial Revolution. Back then, the term identity crisis didn’t yet exist, but the feelings and psychological problems were the same. There was no social media, no YouTube, but you can read autobiographies of working-class people. Historians have analyzed hundreds of autobiographies written by working-class citizens during the Industrial Revolution such as those collected by Emma Griffin in her book Liberty’s Dawn opens in a new tab https://www.econlib.org/workers-lives-during-the-industrial-revolution/ . Additionally, workers frequently petitioned governments regarding how machines stripped away their identities as skilled artisans. So we’ve already been there. Maybe, from history, we can learn how to deal with the status quo and cope with the situation? Eventually, everything should be fine. The technology will be more powerful, and the job will still be there, it’ll just look different. But some people are still anxious about the future. Today, programmers are not the only group falling into that area of anxiety. I’ve written before about mathematicians already relating to the identity crisis of programmers opens in a new tab https://lnkd.in/p/d9R6Y-3B . But, sooner rather than later, it’ll apply to all scientists and white-collar workers. While some scientists find autonomous AI agent swarms “extremely damaging” opens in a new tab https://www.science.org/content/article/how-ai-math-breakthrough-ignited-controversy , others use armies of AI agents to “solve physics” opens in a new tab https://lnkd.in/p/dpkrcNqh . But the future looks brighter than that. We already know what AlphaGo did to the Go players opens in a new tab https://lnkd.in/p/d3FXeswC , and I tend to expect the same, for every skill on which AI will have influence. This is the first post of my new blog. After almost 10 years of this site, I’ve decided now is the time to have my own work-related blog, just like a few of my colleagues. Expect occasional posts, but concise opens in a new tab https://x.com/Konrad Talik/status/2113227811284492706 ones: interesting data, insights, latest trends. Hopefully I did well this time Follow me on LinkedIn opens in a new tab https://www.linkedin.com/in/ktalik or Twitter opens in a new tab https://x.com/intent/user?screen name=Konrad Talik , where I will sometimes link my blog posts. 1. Input tokens count only the uncached part of the prompt. Almost all input was served from the cache see the cache rows , so the medium-effort baseline is tiny and this ratio overstates the difference in workload. ↩︎ fnref1