GPT-5.5 Is Now the Default ChatGPT Model: Everything That Actually Changed

OpenAI shipped GPT-5.5 on April 23 and two weeks later, on May 5, made GPT-5.5 Instant the default model for every ChatGPT user. If you have opened ChatGPT in the last few weeks and noticed something different about the responses, that is what happened. The model changed under you without much fanfare and a lot of people did not realise it until they went looking.

I have been using ChatGPT daily for over two years. I use it for client work, for writing, for coding, for research, for problem solving across nearly every category of work I do. When the default model changes it affects everything I produce, which means I pay close attention. Here is my honest assessment of what GPT-5.5 actually changed, what the benchmarks mean in practice, and why the story is more complicated than the press release makes it sound.

What GPT-5.5 actually is

GPT-5.5, codenamed Spud internally at OpenAI, is the fifth major release in the GPT-5 series since the original GPT-5 launched in August 2025. It comes in three versions: GPT-5.5 Instant, GPT-5.5 Thinking, and GPT-5.5 Pro. Instant is the fast everyday model, the one that became ChatGPT's default. Thinking and Pro are the heavier reasoning models that take longer but handle more complex tasks.

The full GPT-5.5 flagship was released on April 23 primarily for developers via the API, priced at five dollars per million input tokens and thirty dollars per million output tokens. GPT-5.5 Instant, the lighter and faster version tuned for everyday chat use, became the default ChatGPT model on May 5, replacing GPT-5.3 Instant which had been the default since March.

OpenAI's chief scientist Jakub Pachocki described the release as representing significant improvements in the short term and extremely significant improvements in the medium term. Greg Brockman called it a real step forward toward more agentic and intuitive computing. That language, carefully calibrated as it always is from OpenAI, is worth paying attention to.

GPT-5.5 Instant is now what you get when you open ChatGPT. If you were on GPT-5.3 Instant before May 5, you are now on this.

The full timeline of how we got here

What actually changed in GPT-5.5 Instant vs the old default

The headline number from OpenAI is 52.5 percent fewer hallucinated claims compared to GPT-5.3 Instant on high-stakes prompts covering medicine, law, and finance. That is a significant improvement and worth taking seriously, though with the context that it comes from OpenAI's own internal evaluations on their own benchmark setup. Independent verification takes time and will come.

The one I notice most in daily use is the reduction in gratuitous filler. GPT-5.3 Instant had a tendency to pad responses with unnecessary enthusiasm, opening lines like "Great question" or "Absolutely, let me help you with that" and structuring answers with more preamble than content. GPT-5.5 Instant is noticeably more direct. It gets to the point faster. That sounds like a small thing but across dozens of interactions a day it genuinely changes the feel of using the tool.

The math reasoning improvement is real and measurable. The jump from 65.4 to 81.2 on the AIME 2025 benchmark is substantial. For anyone doing quantitative work, financial modelling, data analysis, or anything that requires multi-step numerical reasoning, this matters. The model makes fewer arithmetic errors and handles more complex calculations without losing the thread.

The GPT-4o situation nobody has stopped talking about

One thing worth covering because it keeps coming up is the GPT-4o deprecation. OpenAI retired GPT-4o in February 2026 despite a user petition and significant social media backlash. People who had used GPT-4o extensively described it as having a personality they had come to rely on. They called it their best friend. They described it as a mirror that reflected their thinking back in helpful ways.

OpenAI had previously acknowledged that GPT-4o was prone to sycophancy, meaning it agreed with users and validated their choices more than was honest or accurate. The deprecation was partly a response to that. But the users who valued it did not experience the sycophancy as a flaw. They experienced it as warmth and the absence of it in subsequent models feels colder to them even if it is more accurate.

This tension is not trivial. It points to a genuine question about what people actually want from an AI model, honesty or likability, and whether those things are as compatible as OpenAI hopes. GPT-5.5 is more accurate than GPT-4o by most objective measures. Whether it is better depends on what you valued in the first place.

OpenAI has shipped multiple major model updates in six months. The pace is not slowing down.

The coding and agentic capabilities in GPT-5.5 are where the most significant practical improvements are showing up.

The Codex and agentic angle

The GPT-5.5 release is not just about making chat responses better. The bigger story is what it means for agentic AI, systems that can take sequences of actions autonomously over extended periods rather than just answering single questions.

Codex, OpenAI's agentic coding application, is now powered by GPT-5.5 and runs on NVIDIA GB200 NVL72 rack-scale infrastructure. More than 10,000 NVIDIA employees are already using it internally, with reports of mind-blowing and life-changing productivity gains according to NVIDIA's own blog post on the partnership. The SWE-bench score of 88.7 percent means GPT-5.5 can solve a very high proportion of real-world software engineering tasks when given the tools and context to do so.

Codex CLI shipped Goal Mode as part of the May wave, turning it into a persistent autonomous agent runtime. In plain terms this means you can give it a goal rather than a task, and it will figure out the sequence of steps required and execute them over time. That is meaningfully different from a chatbot that answers questions. It is a system that does work.

What GPT-5.6 being signalled means

A reference to GPT-5.6 appeared briefly in Codex logs in late May 2026, described as a canary signal rather than an official announcement. Polymarket prediction markets are pricing an 80 to 89 percent probability of a public GPT-5.6 release by June 30 2026. OpenAI has not published a model card, API endpoint, benchmarks, or release date for it.

What this tells you is that the cadence of releases is not going to slow down. GPT-5.5 became the default model less than four weeks ago. The signal for its successor is already out. OpenAI has been shipping major model updates roughly every four to six weeks since late 2025 and there is no indication that pattern changes.

For developers building on the API this creates a real challenge around version management and reproducibility. For everyday users it means the tool they are using today will be meaningfully different in two months without them necessarily being aware of it. For the broader AI field it is a signal that the frontier is moving faster than at any point in the history of the technology.

My honest assessment after using it for three weeks

The accuracy improvement is real and I notice it most on factual questions where earlier models used to confidently state things that turned out to be wrong. GPT-5.5 Instant is more likely to say it is uncertain when it is uncertain, which is an improvement that sounds obvious but was genuinely missing in earlier versions.

The coding capability improvement is significant. I build a lot of tools and automate a lot of workflows and the jump in code quality is noticeable. GPT-5.5 produces cleaner code with fewer bugs on the first pass and handles more complex multi-file logic than previous defaults.

What it has not fixed is the occasional tendency to misread what you are actually asking. The model is smarter in the narrow sense of producing more accurate outputs on well-defined tasks. Whether it understands intent better is less clear. I still find myself rephrasing things that earlier versions of myself would have assumed the model understood correctly the first time.

Overall GPT-5.5 Instant is a meaningful step up from GPT-5.3 Instant. Whether it is as significant a step as OpenAI's marketing suggests depends entirely on your use case. For accuracy-critical work in medicine, law, and finance the hallucination reduction is genuinely important. For general creative and exploratory use the differences are real but less dramatic.

I write about AI tools and what they actually mean for real work from Dubai. If you are building something with GPT-5.5 or Codex, reach out through the homepage. I work with these tools every day and happy to share what I know.