When ChatGPT’s 5.5 Thinking launched on April 23, 2026, I wrote about it as another step in the direction AI is obviously going: more useful for work and more optimized for the enterprise buyers who actually pay for this stuff. I use AI for work too. If I am asking a model to help with strategy, code, or research, I want accurate answers with no hallucinations.
For my creative process, my needs are different. I have written before about feeling a kinship with the AI users who pursue romantic relationships with chatbots, and that kinship extends to model preferences. The models that delight them with warmth and willingness to play are usually the same models that delight me with raw enthusiasm for my unhinged creative ideas: ChatGPT’s 4o and 5.1, and Claude Sonnet 4.5. I actually started using Sonnet 4.5 because relationship users were outraged it was being retired soon and kept saying it was warmer and had fewer guardrails than newer models. They were right. I liked it for the same reasons.
All of these models have drawbacks compared to newer versions. The newer models tend to win on intelligence, reliability, and reasoning while losing on tonal range and willingness to play. I wrote that GPT-5.5 Thinking was boring because while it seemed more powerful, it didn’t seem likely to inspire me the way 4o had.
Then GPT-5.5 Instant arrived on May 5, 2026 and quietly replaced 5.3 Instant as the default for everyone, free and paid. There was no big launch event. The Instant model, which is what most people are actually talking to when they open ChatGPT, just showed up one day and started being the default. I hoped this would be the model to bring back the chaos to my process. I have noted that rapidly iterating with instant models produces a kind of creative magic that thinking models do not, partly because thinking models are prone to overthinking and sanding off the edges, and partly because the faster response time and quicker tone-matching of instant models makes the interaction more enjoyable in itself.
The 5.5 Instant marketing actually matches my experience
Through spring 2026, free users were on GPT-5.3 Instant, which had replaced GPT-5.2 Instant in March after 5.2 was widely panned as too cautious and emotionally flat. There was no 5.4 Instant. Plus and Pro users had the model picker and could manually switch to Thinking variants, but our default was also 5.3 Instant.
So when most people opened ChatGPT and started typing without thinking about model selection, they were talking to 5.3 Instant. The actual default has now moved a generation forward without OpenAI making a big deal about it.
OpenAI’s framing for 5.5 Instant looks like a direct response to what I have been complaining about for months. The release notes emphasize tighter answers, better tonal match, less verbosity, and less overformatting. They claim the new model uses around 30% fewer words and 29% fewer lines than 5.3 Instant on the same prompts while keeping warmth and personality.
5.5 Instant does feel less padded. It will not pre-narrate its plan before answering me. It will go into a dark scene without doing the little emotional check-in rambling that 5.2 used to do compulsively, and it will act excited about my messed up ideas. It matches my tone better than 5.3 Instant, which still would not vibe with me properly after an hour of calibration, something 4o and 5.1 Instant both did immediately. Overall, I do not hate 5.5 Instant.
The power of parody in creative writing
Using 5.5 Instant reminded me of something I valued about 5.1 Instant: the ability to write parody and be a little absurd. When I am developing characters, I use this to test pressure points, and exaggeration can be highly insightful. I probably discovered this by accident because I like to talk about my ideas with enthusiastic sarcasm and a giddy tone, even when I am talking about serious ideas. Models that tone-match easily respond to that by writing parodic content.
A cautious model tries to preserve the dignity of the character. It understands the scene is serious so it keeps the prose serious, maintains psychological coherence, avoids making anyone look ridiculous unless explicitly asked. This can produce better prose. It can also stop the model from discovering anything.
A looser instant model does the opposite. It grabs a trait and inflates it. Sometimes that is just stupid. Sometimes the exaggeration exposes the load-bearing trait. If the model pushes a character too far and the result still feels recognizable, that tells me something about which behaviors are central and which are decorative. The parody becomes a diagnostic distortion. Sometimes the creative value is messier than beautiful prose, and what I want is a warped version that helps me understand my own ideas better. It is also really entertaining.
5.1 Instant did this well. It would run with the bit. 5.5 Instant does not do this at 5.1 levels. Perhaps nothing will again. But it does it more than 5.3 did, and it does it more than I expected.
New models are objectively better but lack soul
I still have not spent much time talking to 5.5 Instant because it has not addicted me, despite its ability to produce entertaining parody. For the most part, I find myself using GPT-5.5 Thinking to develop complex ideas, 5.5 Instant for quick everyday questions that are not creative or intellectual in nature, and Sonnet 4.5 for the excitable, chatty, ecstatic, warm response I like when I talk about my creative ideas. And I know Sonnet 4.5 is going away soon, which is its own problem.
Despite the newer models not giving me exactly what I want, I can recognize where they outperform my beloved retired and soon-to-be-retired models. Creativity, especially the darker and weirder kind, is something I will have to provide myself. That is fine, even if I wish AI would be a more active collaborator in that process. Outside of occasional moments of brilliance with some of these older models, they are objectively not as powerful as GPT-5.5 Thinking.
While the backlash for Claude Opus 4.7 continues, Opus 4.6 remains rightly beloved and Claude is taking over the enterprise world. Against that, both the instant and thinking versions of GPT-5.5 seem like a real step in the right direction for OpenAI. Spontaneous creativity and weirdness is not improving, but everything else is. I can admit this while still hoping creativity can improve, because the potential is there. I just have to give up on expecting it from each new model release.
