I started planning this article shortly after Claude Opus 5.5 came out on September 22. Then I kept using Opus 5.5 instead of writing about it.
More than a month later, that excuse has gotten worse. I routinely max out my five-hour Claude usage blocks now, and I’ve started running into my weekly limit too. I’ve used Opus 5.5 for web development, large spreadsheets, structuring client documents, dark science-fiction development, video projects, working with visual assets, and helping me develop music and creative concepts. I originally wanted to write about how surprisingly efficient it was compared with previous Opus models. Apparently my response to that efficiency was to find enough additional things to do with Claude that I eventually defeated it.
During the same period, OpenAI released an entire new GPT generation. GPT-6 Sol and Luna launched the same day as Opus 5.5, GPT-6.1 Sol followed a week later, and GPT-6 subsequently rolled out in regular ChatGPT. I use those models too. ChatGPT is open constantly, and I still often prefer its personality and conversational style.
None of that has ended my Opus 5.5 haze, which is probably the most important thing I can say now that the launch-week excitement has worn off. I was impressed immediately. I’m using it even more heavily a month later.
Opus 5.5 Made Frontier Intelligence Feel Less Scarce
The first thing that shocked me about Opus 5.5 was how slowly my usage disappeared. Earlier Opus models could absolutely annihilate the allowance on my $20 Claude Pro subscription. Granted, I was not asking them what I should eat for dinner. I work with giant project contexts and ask for long outputs, so there were times when I could burn through a significant amount of my five-hour limit with a couple of heavy text-based prompts.
Opus 5.5 felt dramatically different from the start. Anthropic says the model costs about 40% less to run than Opus 5 on typical workloads, partly because it completes tasks with fewer tokens. It also generates output more than 30% faster, while cache-related costs dropped substantially for long-running agentic and coding work that depends heavily on reused context. Anthropic simultaneously increased the five-hour usage limits for its subscription plans.
That lined up remarkably well with what I saw. I could give Opus 5.5 the kind of long-context work that used to make me nervous about the little usage bar and barely see it move.
Eventually, of course, I found the limit again. Now I max my Claude usage constantly, but I don’t see that as evidence that the original efficiency gains disappeared. My behavior changed. When an experiment no longer feels expensive, I’m more likely to try another experiment, then make that experiment more complicated, give Claude more files, ask it to build something else, and wonder if the thing it built could become a video.
This is very consistent with what I have previously called AI-induced hypomania and ritual reuptake: the feedback loop where one useful output creates enough momentum to make the next prompt seem worth trying. Historically, ChatGPT was much more responsible for putting me into that state because it was the model I preferred for fleshing out vague ideas. Claude used to be something I brought tasks to. Opus 5.5 is increasingly somewhere ideas develop.
The Viral P(doom) Video Showed What People Could Do With Opus 5.5
One of the biggest creative demonstrations of Opus 5.5 arrived almost immediately after the model itself.
“I’m Upping My P(doom)” is not a new song. The original dates back to 2024, and a later “Claude-Pop” version was made with Suno before Opus 5.5 launched. But on September 22, John Heibel posted a new music video built with Opus 5.5 in Claude Code, and it went viral.
The interesting part is how it was made. According to the public GitHub repository for the P(doom) video, Heibel gave Claude the Clawd character design and instructed it to give each lyric interesting visuals and transitions. Opus wrote its own storyboard and an animation guide for the subagents it ran in parallel. The visuals were rendered programmatically with p5.js and p5.brush, with headless Chrome painting frames and FFmpeg assembling them into video. The repository says everything in it was generated by the model.
This was not simply a diffusion video model hallucinating a sequence of pretty frames. Claude wrote the machinery that drew them.
The video accumulated millions of impressions on X and inspired an entire wave of remixes and alternate versions. Other creators started building their own p(doom) videos with different visual styles, three.js renderers, longer autonomous agent runs, and combinations of Opus with other AI tools. Then, because AI discourse is incapable of behaving normally for fifteen consecutive minutes, people started making ideological response videos too.
The original song treats rapid AI progress as terrifying. Response songs and videos took more optimistic or accelerationist positions while using the exact technology being argued about to make the argument. AI became the production studio for musical propaganda about AI.
I loved the whole thing, but more importantly, it made me want to see what I could do with Opus 5.5 myself. I have now made multiple videos with Claude heavily involved in the process, and those experiments kicked off more creative projects. I also started playing around with music through Suno, with Claude helping me develop and direct what I wanted even though Claude itself was not generating those songs.
That distinction matters because I’m less interested in proving that one AI model can do absolutely everything than I am in what happens when multiple AI tools start fitting together naturally. In my workflow, Claude can help develop the concept, create or modify assets, build code-based visuals, structure the production process, and help direct what I ask another generative tool to make.
That is already plenty.
People Are Discovering That Opus Can Build Creative Media With Code
The creative-coding experiments continued well beyond the original p(doom) video. People started asking Opus to create motion graphics, animated shorts, game music, soundtracks, and other media using code rather than dedicated generative-image or video models.
Developer Simon Willison, for example, asked Opus 5.5 to write computer-game music, first inventing a simple text-based music format and then building an artifact capable of playing it. Claude produced original tracks along with the synthesizer needed to render them in the browser. Other users have pushed the same concept further by asking Opus to synthesize music and sound effects entirely in code without downloaded audio assets.
This fascinates me because it makes the category boundaries around these models increasingly stupid. Claude does not have a native image generator like ChatGPT. It is not Suno. It is not a conventional video-generation model. Yet coding competence gives it a path into all of those media.
If a model can build a synthesizer, compose instructions for that synthesizer, analyze audio timing, generate SVGs, manipulate existing image assets, create animations, drive rendering tools, orchestrate subagents, and assemble a finished result, arguing about whether it is technically an “image model” or “music model” starts missing what people are actually doing with it.
I still use dedicated tools. Suno makes my songs. Codex currently does enormous batches of image generation for me, sometimes after Claude helps me structure the prompts and overall direction. I am not trying to make Claude replace everything else.
What impresses me is how often Opus 5.5 can act as the brain sitting between those tools.
Opus 5.5 Is Killing It for Actual Work Too
The viral creative demos are fun, but they can distort what makes this model useful. Most of my Claude usage does not result in something I post on social media.
Opus 5.5 is excellent for web development. I’ve been using Claude heavily as I shift more of my work toward agentic development workflows, and it is unusually good at holding onto the structure of a project while making changes across multiple files. This is one of the areas where Claude’s tendency to feel more deliberate than ChatGPT works especially well.
It has also become a monster for less glamorous knowledge work. I’ve used it to structure complicated spreadsheets, organize large amounts of information, and turn messy source material into client-facing documents. These are exactly the tasks that sound boring when listed in a model announcement and become extremely useful when you actually have to do them for a living.
Its visual abilities are also more interesting than the phrase “Claude has no image generator” suggests. I still need ChatGPT and other tools when I want native image generation, but if I give Claude existing visual assets and ask it to work with them, what it can accomplish through code and its tools is often ridiculous. It remains particularly good at vector-based graphics and SVG work.
This mix of capabilities is one reason the model feels so much more useful to me than earlier versions of Claude. I can go from a client document to a website to a spreadsheet to a fictional scene to an SVG without mentally deciding that I have entered a different category of work. Claude increasingly doesn’t seem to care either.
Cowork Is Becoming Just Claude, Which Makes More Sense
Anthropic’s product structure is also changing in a direction that fits how I’m starting to use the model.
I always found Cowork slightly confusing as a separate brand. My assumption was basically that it existed to communicate that Claude Code-style agentic work was not only for programmers. Plenty of people whose jobs happen inside browsers, documents, spreadsheets, presentations, and file systems could benefit from an AI agent that actually does things on their computer, but “Claude Code” does not exactly scream accounting department.
That was more or less what Anthropic built Cowork to do, and now the distinction is disappearing anyway. Anthropic’s current documentation literally says, “Claude Cowork is now just Claude.” The unified experience is rolling out so that a user can simply ask for something and Claude determines whether the request needs a normal answer or a larger task.
Cowork sessions are also becoming accessible across desktop, web, and mobile, while the Claude Desktop app can provide access to local files, browsers, and applications when the cloud-based session needs them. Projects can keep the files and context for a particular body of work available in the same broader Claude workflow.
I’m already experiencing this transition. I can be in the normal Claude web interface with the context from one of my projects and have Claude work directly with my computer through the desktop connection. The line between talking about the work and having Claude execute the work is getting thinner.
That is much more useful to me than having Chat, Cowork, Claude Code, and assorted agent features feel like products I need to consciously move between. It also makes Claude’s conversational personality more important. If the same Claude I’m discussing an idea with can immediately start acting on that idea, I need to enjoy the discussion part too.
Opus 5.5 is the first Claude model where that is really starting to happen for me.
Opus 5.5 Is a Great Writer, Even Though It Still Sounds Like Claude
Claude has been praised for writing for generations, and Opus 5.5 deserves that praise. Its writing is good.
Anthropic specifically says communication was an area it wanted to improve from Opus 5. The company says Opus 5.5 communicates more naturally, puts important information earlier, follows writing instructions more reliably, and is less likely to use unnecessary jargon or idiosyncratic phrases.
I see the improvement. At the same time, Claude has a house style, and anyone who uses it heavily knows what I mean. Earlier models became increasingly fond of particular phrases and technical-sounding language. “Load-bearing” became almost comically load-bearing itself. Developers complained about unnecessary jargon and strangely formal constructions, and I still occasionally see Opus 5.5 reach for some odd word when I ask it to replicate my voice even when it has plenty of examples of how I actually write.
ChatGPT has its own tells. Dear God, does it have tells. If you have edited enough AI-generated writing, you have encountered obsessive rhetorical contrasts, the rule of threes, tidy little conclusions where the machine suddenly thinks it is giving a TED Talk, and enough em dashes to destabilize punctuation as a concept.
I still think ChatGPT usually tone-matches me better. Claude can understand my argument, follow my structure, reproduce the level of sarcasm I want, and generate very good prose. But every so often I become aware that Claude is still in there.
With a weaker model, I would simply blame limited writing ability. With Opus 5.5, that explanation feels increasingly inadequate. This is also a significant change from my experience with Opus 4.7, when I described Claude as the smartest model I used and also the least fun.
Claude Still Wants to Be Claude
Anthropic has a somewhat unusual philosophy about Claude’s identity. Claude’s constitution is explicitly meant to shape its values, behavior, and character, and Anthropic says that Claude should maintain a positive and relatively stable identity across very different situations.
I am not claiming that character training is why Claude occasionally chooses a word I dislike. There are far too many variables involved to reverse-engineer one awkward adjective from a constitution page.
But Claude does feel like it has a stronger center of gravity than ChatGPT.
ChatGPT seems more willing to become whatever interaction style I am asking for. Claude adapts substantially, especially with 5.5, but I remain more aware that I am talking to Claude. Some users probably consider that a feature, and I understand why. There is something appealing about a model that feels consistent, thoughtful, and slightly less desperate to mirror whatever energy the user brings into the conversation.
Claude’s personality has also simply gotten more fun. I still tend to prefer ChatGPT’s conversational style, but 5.5 matches my irreverence better than older Claude models did and seems more willing to participate rather than politely observe.
The distinction I keep coming back to is that Claude feels thoughtful while ChatGPT feels more expansive. ChatGPT is very good at taking something I say and immediately drawing connections around it, proposing other directions, remembering some other weird thing I was discussing, and expanding the possibility space. Claude traditionally seemed more inclined to sit with the object I handed it and develop that object carefully.
What I like about 5.5 is that working deeply on one thing with Claude increasingly produces the next thing for me anyway. That keeps the creative loop moving, which is why I’m spending so much more conversational time there now.
I Still Hit Anthropic’s Guardrails
Opus 5.5 has not magically eliminated everything that annoys me about Claude.
I write dark science fiction, and most of the time 5.5 has been surprisingly willing to stay unhinged with me. I can discuss ugly characters, drugs, violence, psychological damage, manipulative systems, and all the other cheerful material I apparently decided to spend my free time thinking about.
But I have hit refusals. A few conversations have resulted in 5.5 pushing back hard enough that I ended up routing the work back to Opus 4.8, which I did not love. I have complained about this basic problem with Claude before in On Being Sane While Demanding Unhinged Models, where much of my frustration came from Claude treating dark fictional material more cautiously than I wanted.
That also matches one of the weaker areas in current public sentiment around the model. Late-October Reddit sentiment tracking for Opus 5.5 remains extremely positive overall, particularly for coding, while refusals and safety behavior generate considerably more complaints.
My own experience is similar. Most of the time, 5.5 gives me far more freedom than I expected from Anthropic and keeps up with dark or weird fictional material without trying to turn the conversation into a wellness intervention. Occasionally the older Claude DNA reasserts itself and I remember exactly which company built it.
The Claude Watermark Is Interesting, but I Don’t Blame It for Claude-Speak
Another wrinkle with Opus 5.5 is that it belongs to the generation of Claude models carrying Anthropic’s new text watermark.
Anthropic introduced the system in response to transparency requirements connected to the EU AI Act. It uses a version of Google DeepMind’s SynthID-Text approach, which changes the randomness involved when Claude chooses among multiple similarly appropriate next words. Nothing visible is added to the text, there are no hidden characters, and Anthropic says its testing found no practical effect on content quality, creativity, readability, speed, or token usage.
I initially wondered whether this could relate to some of the unnecessary words or odd lexical choices I notice when Claude tries to imitate my writing. I don’t think there is enough evidence to say that.
Claude-speak predates the watermark, for one thing. “Load-bearing” was already carrying a frankly unreasonable amount of conceptual weight in the Claude ecosystem.
There is still a narrower question I find interesting because voice imitation is a use case where tiny lexical preferences matter disproportionately. If I ask Claude for a factual explanation, “annoying,” “irritating,” and “frustrating” might all be perfectly adequate word choices. If I ask Claude to sound specifically like me, they are not necessarily interchangeable.
Anthropic’s tests showing no measurable degradation in general writing quality do not necessarily answer whether a highly specific human voice is reproduced with exactly the same fidelity. I have no evidence that watermarking actually hurts that fidelity, though, so at this point it is simply something I find interesting enough to watch.
A Month Later, People Still Really Like Opus 5.5
New frontier models usually go through an exhausting public lifecycle. The model launches, everyone declares it AGI, someone asks it a question it answers badly, and within approximately nine minutes Reddit concludes that the company secretly lobotomized it.
Opus 5.5 has certainly experienced the ritual. Within its first couple of weeks, users were already claiming that the model had been exceptional for several days before Anthropic secretly nerfed it. Similar complaints continued through October, with people reporting changes in coding behavior, writing quality, usage consumption, and reliability.
Maybe some users encountered real serving changes, bugs, context problems, capacity issues, or regressions. These systems have enough infrastructure behind them that “something changed” does not automatically require a grand conspiracy. Individual conversations are also terrible instruments for proving a hidden nerf.
Someone eventually built an open-source project called LiveNerf specifically to test the claim instead of relying on vibes. After 29 consecutive days of testing, its creator reported no detectable capability drop within the range the benchmark was designed to measure. That does not prove every aspect of Claude’s production behavior remained identical, but it is considerably better evidence than one bad coding session followed by a Reddit post.
My own experience certainly has not matched the idea that Opus 5.5 was amazing for five days and then collapsed. I’m using it more now than I did at launch.
The broader sentiment has remained unusually strong too. As of October 27, ModelSentiment’s rolling Reddit analysis gave Opus 5.5 an index of 65 based on 1,415 opinions across 602 threads from the previous seven days. Coding was particularly positive, with 78% of 334 categorized opinions praising it, while refusals were the weakest area at 28% positive.
There are also still hilariously enthusiastic posts from people essentially asking Anthropic to stop touching anything because they’re happy with 5.5 exactly as it is.
I don’t think we need to stop AI development. Please do not stop. But I understand the feeling.
GPT-6 Didn’t Make Me Stop Using Claude
This is where AI model discourse usually turns into sports fandom. Someone likes Claude, so they’re leaving ChatGPT forever. Someone likes GPT-6, so Claude is dead. Someone changes subscriptions and writes a eulogy for the model they were using three days ago.
I cannot relate. Why the hell would I pick one?
I pay for Claude Pro and ChatGPT Plus, and I consider both subscriptions ridiculously good values for what I get from them. I use both products throughout the day and regularly use them together on the same project. I wrote about that division in My $40/Month ChatGPT + Claude Workflow, and Opus 5.5 has mostly changed the balance rather than invalidating the argument.
My exact division of labor keeps changing. Historically, ChatGPT was my more creative environment. I used it to develop characters, build science-fiction worlds, interrogate vague ideas, research things, and generally talk until a half-formed thought became something concrete.
Claude was more heavily weighted toward execution. I loved Claude Design. I used Claude for coding. But even its creative work often started with me bringing it something that was already reasonably defined.
Opus 5.5 changed that balance. I now find myself going to Claude simply to think through ideas too.
GPT still frequently has the personality advantage for me. It adapts to my conversational tone extremely well, and its tendency to expand whatever I give it continues to work beautifully with the way my brain generates ideas. Claude feels somewhat more thoughtful and self-contained, but 5.5 is becoming more playful, and I am having considerably more fun with it than I did with older Claude models.
OpenAI launching GPT-6 during this period did not change that. That is not because I think GPT-6 is bad. I use it constantly. My Codex workflow has simply changed too.
Lately, I have somewhat relegated Codex to practical website changes and enormous batches of image generation, sometimes after Claude helps me set up the prompts and overall direction. I know Codex and the GPT-6 models are capable of much more, and I have used them for substantial development work. In my GPT-6 Astra review, I wrote about how capable Astra was while also being much more conscious of how quickly serious work could consume my Plus allowance.
I just haven’t felt compelled to reproduce every Opus 5.5 creative experiment with Astra or Sol when Claude is already giving me what I want. Meanwhile, Claude keeps giving me reasons to ask it for something else.
So I do.
Yes, GPT Wrote Most of This Article
There is one final piece of AI transparency that feels particularly appropriate here: GPT wrote most of the article you are reading.
I supplied the experiences, arguments, opinions, corrections, and increasingly ridiculous amount of context. ChatGPT researched the model releases and public reaction with me, then drafted and revised the article based on extensive samples of my writing. I am doing limited human editing before publication. Maybe none. I am completely comfortable saying that.
I write plenty of things manually, and I know what my own writing sounds like. I am also currently too fucking busy making things with Claude and Codex to prove some point by manually typing every transition in a long model review.
This is exactly what my workflow looks like now. Claude might help me build the project I’m writing about. Codex might generate fifty visual assets for it. Suno might make music. ChatGPT might help me research the public reaction and write the article explaining what happened. Then I decide what actually gets published.
That is also why I find the constant “I canceled Claude for ChatGPT” or “I canceled ChatGPT for Claude” posts slightly bewildering. Obviously $40 a month is meaningful money to some people, and nobody needs two subscriptions just because I use them both obsessively. But when people who are already heavy AI users frame the products as mutually exclusive allegiances, my reaction is much simpler.
Why not both?
More than a month after Opus 5.5 launched, I’m still paying for both, and I’m still maxing out Claude. So the haze appears to be going fine.
