Chat GPT 5.5. Reviews. The King Returns to the Throne.
On April 23, 2026, OpenAI released its latest artificial intelligence model, GPT 5.5. The launch took place just seven weeks after the debut of GPT 5.4, which says a great deal about the pace at which the AI industry is conducting its race. OpenAI continues its quiet-launch approach, releasing the latest model by surprise in the hope of achieving a “wow” effect, without risking disappointed expectations.
Was the intended effect achieved in this case? Absolutely. The model performs better at coding and reasoning, especially with longer texts and more complex tasks. It also does a better job of “guessing” what the user wants, which is a significant improvement over previous versions. Integration with other tools has also been improved; the model handles the generation of tables and documents much better and, in this respect, has begun to match Claude.
However, the leap in programming quality is only apparent when ChatGPT 5.5 works in Codex or another program-writing tool. Code written directly in the application has quite a few problems.
Also noteworthy is the improved graphics generation—the graphics module has finally been integrated with the reasoning model (similarly to how it works with Nano Banana Pro), so infographics can now be generated in ChatGPT, and it does this better than Gemini. Text writing is the new model’s weakest area. Although the model of course handles it quite well, there is no visible leap in quality here like in other fields, where GPT 5.5 is the leader or ranks second, just behind another top model.
Table of contents
-
ChatGPT’s overall advantage over the rest (the most complex application)
-
Generating graphics and infographics. GPT has surpassed Gemini.
-
Where does GPT 5.5 come from, and how does it differ from its predecessors?
OpenAI returns to the leadership position.
Since the launch of ChatGPT 5 (August 2025), OpenAI has not been on a good streak. The famous fifth version had essentially been announced since the launch of GPT-3.5 (the first publicly available LLM, which sparked the current AI revolution). GPT-5 was announced as a gigantic revolution that was supposed to change the world, but it turned out to be merely a slightly better model than OpenAI’s previous models.
In the meantime, Gemini and Claude managed to overtake GPT in many areas. New releases in the GPT-5 series did not save the situation. The model performed well when coding and handled precise instructions effectively, so it was valued by programmers and specialists, but ordinary users were gradually switching to other models. Although ChatGPT was of course still a popular model, it was slowly losing the champion position it had held since the beginning of the current AI boom.
GPT 5.5 reverses this situation.
Main advantages of the new model
The greatest improvement in quality can be seen in image generation (although this is due to the new GPT image 2.0 model, not GPT-5.5 directly). There is also a huge leap in quality in coding, reasoning, and document creation.
ChatGPT, Claude, Gemini, Grok – which one to choose, and how does the release of ChatGPT 5.5 affect the choice?
The release of the newer version means that ChatGPT has caught up with Claude in terms of coding, reasoning, writing, and research. Particularly significant is its ability to work with tables and documents (Claude has had this feature for a long time).
Personally, I believe Claude is still better at these things. In Claude, we can immediately preview such documents, whereas in ChatGPT, preview is available only for tables; the result of generating a document can only be seen after downloading the file. However, ChatGPT has the advantage of being able to generate graphics (and infographics). Claude does not have this feature at all (that is, it allows graphics to be generated using SVG code, but this has its limitations).
Claude is unable to place images in such graphics (e.g. model logos, unless we spend additional time generating the logos in SVG).
Infographics serve not only an aesthetic function; they are also a good way to summarize text and make the subject easier to understand. Therefore, it is worth using them, even if you ultimately do not use such infographics publicly.
ChatGPT 5.5 and Gemini
ChatGPT has overtaken Gemini when it comes to generating infographics. For a long time, this was the main advantage of the Nanobana Pro model available in Gemini. When it came to text generation, the differences were minimal, and even if one model was better, it was possible to achieve a similar result in another using the right prompts. However, if you needed infographics, Gemini was the only option. When it comes to generating infographics, ChatGPT 5.5 has significantly overtaken Google’s applications.
Video Generation: Gemini and Grok Outperform ChatGPT Here
The paid version of Gemini allows you to generate 3 videos per day, while the paid version of Grok allows you to generate a larger number of videos. It is still more of a toy than a proper tool, but it is already possible to create useful videos with it. And given the development of the technology, it is a good idea to know the basics of video generation.
OpenAI made the Sora video-generation app available on its paid plan for a long time, but recently discontinued it to focus on its main model.
However, neither Grok nor Gemini covers the entire topic of creating AI-generated videos, so if, beyond just having fun, you want to create a film lasting a few minutes, it is necessary to use other tools.
ChatGPT’s Overall Advantage Over the Rest (the Most Complex Application)
ChatGPT simply has the most features. So if you don’t see a difference in the quality of the language model itself, GPT’s additional features (such as advanced personalization, agent mode, etc.) may be an added advantage for you. These aren’t essential features; in many cases, they’re simply minor conveniences.
Generating Graphics and Infographics. GPT Has Surpassed Gemini.
For a long time, NanoBanana Pro was the undisputed leader in generating simple infographics. Nano Banana 2 is the first widely available graphics model combined with a multi-step reasoning model that performed an analysis before generating the image itself. This made it possible to create infographics at a useful level of quality. A significant leap in quality came from improved text generation, as graphics models generally struggled with this by default.
GPT has currently surpassed Gemini in this regard. The graphic above was created by ChatGPT after a few follow-up prompts.
For better comparison, it shows what result ChatGPT and Gemini produced from the same prompt: “Create a graphic showing the timeline of language models. Arrange it so that the company name (model family) is at the top: ChatGPT, Gemini, Claude, Grok, DeepSeek. Each of the five companies should have a vertical line and the release dates of its models. Go from the bottom upward, so that the present is at the top.”
Below is ChatGPT:
And below is the result of Gemini’s work.
The difference is colossal. Admittedly, both models made some mistakes with the first prompt. ChatGPT did not provide newer models, whereas the graphic itself is cleaner and, above all, does not look like a typical AI-generated creation.
However, the graphic from Gemini is full of elements characteristic of artificial intelligence. We therefore have duplicated dates on the main timeline, and each timeline additionally has a duplicated mini-timeline with different dates. Another problem is the decorative elements, which are generally unnecessary. In this case, the background does not look bad, but such things are often problematic. Admittedly, these issues can be easily fixed, but it is nice that GPT does what is expected by default.
GPT’s advantage is evident when it comes to enhancing graphics. Using OpenAI’s tool, I had no trouble creating a fully usable graphic (the one at the beginning of the article). Most importantly, I didn’t notice the model “breaking” anything even once. This is important when generating things with AI. When a subsequent command ruins what was created earlier, it disrupts the workflow, because it creates the impression that AI is simply unable to handle the task at all.
Generating infographics is one of the better applications of graphic models.
It is true that graphics generation is not handled directly by the GPT 5.5 model, but rather by GPT Image 2.0. However, it premiered just two days before the launch of the main model, so from the user’s perspective, the topic is quite closely related.
Importantly, the new graphics model is also available in the free version.
ChatGPT Images 2.0.
The latest image-generation model in ChatGPT is ChatGPT Images 2.0 (based on the model
gpt-image-2
), which OpenAI announced just a few days ago, on April 21, 2026.
Images 2.0 is OpenAI’s most advanced image generation model, with improved text rendering capabilities, multilingual support, and advanced visual reasoning.
Possibilities of “thinking”
The model has reasoning capabilities that allow it to search the internet, create several images from a single prompt, and verify its own results. As a result, Images 2.0 can create marketing materials in various formats as well as multi-panel comics.
Quality and resolution
Images can have a resolution of up to 2K and be generated in many aspect ratios. OpenAI claims that Images 2.0 can be used to create entire magazines, and the model also supports non-Latin languages.
Two modes are available: Instant (fast) and Thinking (reasoning).
Text rendering
The model can create materials such as restaurant menus ready for immediate use, without the spelling errors that plagued graphics models for years. Earlier models were based on diffusion architecture, which reconstructs images from noise and handled text poorly.
Tables and documents
From the perspective of an average user, the ability to generate a useful table in AI is often more important than a given language model’s advanced reasoning and programming capabilities. This feature existed in older versions of GPT, but until now it was only possible to create simple tables.
ChatGPT 5.5 can immediately generate complex workbooks consisting of several logically connected tables. And, most importantly, the model can modify them based on subsequent commands.
In this case, I asked for a table of programs and applications based on artificial intelligence. In addition to the main table, I immediately received sheets with a summary of the table, a chart, a list of tags, a glossary, a table description, etc., even though I hadn’t mentioned them (and these additional elements are quite useful).
From the very beginning of the ChatGPT 3.5 release, language models have been a great help in creating tables, and their assistance in writing formulas has been highly useful. However, everything had to be copied manually. The function for generating an entire table at once significantly simplifies the task. A table created in this way can then be downloaded and modified manually.
At the moment, there is no convenient official OpenAI plugin for modifying spreadsheets directly in Excel (Claude has this feature), but it seems to be only a matter of time before it is introduced. One possible solution is to download the ChatGPT or Codex app and work with files from your own drive.
Generating flyers and documents
And below is the result. ChatGPT created a double-sided version of the brochure for printing. At this stage, it requires many corrections, and in fact, everything needs to be done manually, but we already have a solid foundation to start working from. It should be noted that this is the result of a single command, so there is potential to achieve better results with more work.
And below is the version created as a graphic, rather than as a document.
Program Creation
When I started writing this text, I tested the ability to generate code that runs directly in the application. Unfortunately, I was disappointed. The capabilities were only marginally better than in older versions, while the program often would not run at all. Problems like these appear in Gemini or Claude only after 6–10 revisions, whereas in ChatGPT errors occurred with the first or second command in roughly 40% of cases. Of course, this is not a problem for programmers: by using better prompts, errors can be avoided. However, the feature for running code directly in the application is intended mainly for people who cannot program.
This feature can be used, for example, to quickly create a website prototype, prepare an interactive dashboard presenting data and information in an attractive format, or have fun creating simple games, which motivates users to download more advanced tools.
GPT-5.5 and Codex. A quality improvement is evident when the model operates in a programming environment.
One of the biggest beneficiaries of the new model is Codex, OpenAI’s agentic coding tool. In Codex, GPT-5.5 can now interact with web applications, test workflows, click through pages, take screenshots, and use them to create better versions of what it sees until the task is complete.
One of the tested features stands out in particular. After GPT-5.5 built an application with a bug, Codex opened a browser session, navigated the application visually, identified the problem, and fixed it autonomously. The model used its own cursor, clicked through the features, visually verified the fixes, and reported when all tests had passed successfully.
Over 85% of OpenAI employees use Codex every week across various departments, including engineering, finance, communications, marketing, data science, and product management. This figure demonstrates that GPT-5.5 in Codex is no longer a tool exclusively for developers.
How Does ChatGPT 5.5 Handle Programming?
The capabilities of OpenAI’s new model are well illustrated by the video below.
The best example from the film is the digital twin of the Earth, meaning a program displaying a map of the entire world. ChatGPT used ready-made libraries here, but it is still impressive.
Another interesting example is a three-dimensional game created based on two prompts.
Where does GPT 5.5 come from, and how does it differ from its predecessors?
To understand why GPT-5.5 is more than just a routine update, it is worth looking back. The models in the GPT-5.x family released since late 2025—5.1, 5.2, 5.3, and 5.4—were improved versions built on the same base model. GPT-5.5 is the first thoroughly retrained base model since GPT-4.5. Its architecture, training data corpus, and objectives focused on agentic action have been redesigned from the ground up.
The model is natively omni-modal, meaning it processes text, images, audio, and video within a single, unified architecture. It is no longer a text model with the ability to understand other media added on. All modalities are treated as first-class inputs from the very beginning.
Internally at OpenAI, the model was codenamed “Spud.” It was publicly presented as a response to Anthropic’s Claude Opus 4.7, which had appeared exactly one week earlier and taken the crown in programming. The timing of GPT-5.5’s launch was no coincidence.
What Can GPT-5.5 Do? General Capabilities and Applications
The key word describing GPT-5.5 is agentic capability. The model does not wait for the user to guide it through a task step by step. Instead of managing every stage of the work in detail, you can give GPT-5.5 an unstructured, multi-part task and trust it to plan its own actions, use tools, check its work, handle ambiguities, and continue until completion.
This means that instead of writing: “Analyze this file, then prepare a summary, and then create a chart,” you can simply say: “Here are six months of sales data. Find anomalies and propose corrective actions.” The model will independently establish a plan, select the tools, and deliver the final result.
Particularly strong progress can be seen in agentic programming, computer use, office work, and preliminary scientific research—in other words, areas where progress depends on reasoning within a broad context and taking action over longer periods.
OpenAI provides concrete examples of internal implementation. The company’s finance team used Codex with GPT-5.5 to review 24,771 K-1 tax forms totaling 71,637 pages, completing the task two weeks faster than the previous year. An employee in the sales department automated the generation of weekly business reports, saving five to ten hours of work per week.
Earlier teams testing the model could verify code created using the “vibe coding” method, review thousands of additional documents, and save up to ten hours of work per week.
BenchmarksGPT 5.5.
The numbers are impressive, but they require context. GPT-5.5 does not win everywhere, and a fair assessment requires looking at the bigger picture.
Model strengths:
On Terminal-Bench 2.0, which tests complex command-line workflows requiring planning, iteration, and tool coordination, GPT-5.5 achieves a score of 82.7%, setting a new record among publicly available models.
The leap in understanding long contexts is particularly impressive. On the MRCR v2 test with a one-million-token context, the score improved from 36.6% in GPT-5.4 to 74.0%, representing more than a doubling in effectiveness. This is a breakthrough change for applications requiring the analysis of entire codebases or extensive document collections.
On GDPval, which tests agents across 44 real-world occupations, from finance to legal research and product management, GPT-5.5 achieves 84.9%. On Tau2-bench Telecom, which tests complex customer service workflows, the score is 98.0% without any prompt tuning.
In bioinformatics and genetics, the model also sets new standards. On BixBench, it achieves 80.5% compared with its predecessor’s 74.0%, while on GeneBench it improves the score from 19.0% to 25.0%.
Where GPT 5.5 Loses to the Competition.
Not all benchmarks favor GPT-5.5. Claude Opus 4.7 scores 64.3% on SWE-Bench Pro compared with GPT-5.5’s 58.6%, meaning Anthropic’s model is better at solving real-world problems from GitHub repositories. Claude also leads on MCP Atlas, which tests the orchestration of multiple tools, scoring 79.1% versus 75.3%, while in multilingual natural language understanding, Claude and Gemini clearly outperform OpenAI’s new model.
Among developers, the reaction is measured rather than enthusiastic. Some point out that the model is prone to more confident errors—so-called hallucinations—which in agentic workflows can be more dangerous than stopping and asking a question.
Availability and Pricing
GPT-5.5 is available to paying OpenAI subscribers, including users of the Plus, Pro, Business, and Enterprise plans in ChatGPT and Codex. The GPT-5.5 Pro version is limited to Pro, Business, and Enterprise plans.
The pricing is one of the more controversial aspects of the release. The API price increased from $2.50 and $15 to $5 and $30 per million input and output tokens, respectively, representing the largest one-time price increase in the entire GPT-5.x series.
OpenAI argues that this increase is offset by greater token efficiency. GPT-5.5 uses significantly fewer tokens to achieve results comparable to GPT-5.4. The company is actively absorbing part of the gains from token efficiency to maintain value for subscribers, which means that Plus plan users can accomplish more work in a given month, even if the model is more expensive on the API side.
Real costs for teams remain a topic of discussion. At 10 million output tokens per month, GPT-5.5 costs $300, while Claude Opus 4.7 costs $250. However, if GPT-5.5’s superior agentic performance means 25% fewer iterations are needed, the cost difference disappears.
Safety
GPT-5.5 is the first OpenAI model to undergo the company’s full new suite of safeguards designed for the era of agentic models. The model was evaluated against the complete set of safety and preparedness frameworks, tested with the participation of internal and external red-teaming specialists, and subjected to targeted tests of advanced capabilities in cybersecurity and biology. Before its release, feedback was gathered from nearly 200 trusted early-access partners.
OpenAI classified the model’s capabilities in biology, chemistry, and cybersecurity as “high” according to its own risk assessment framework. At the same time, it made a special access pathway available to cybersecurity professionals who want to use the model’s capabilities for defensive work.
OpenAI’s Situation After the Release of GPT-5.5
It is difficult to assess GPT-5.5 in isolation from the broader market landscape. Its release just six weeks after the debut of GPT-5.4 highlights how intensely leading companies are competing for corporate customers and how their models are increasingly evolving through continuous updates.
The company also stated that Codex has 4 million active users and that ChatGPT has 9 million paying business customers. ChatGPT has over 900 million weekly active users and more than 50 million subscribers.
Industry analysts point out that the release pace is not accidental. The six-week cycle for launching new models means that businesses do not have time to configure one solution before the next one appears. It is a strategy that forces continuous dependence on the platform and makes migration to competitors more difficult.
Greg Brockman spoke openly about the “super app” project, which is intended to combine ChatGPT, Codex, and an AI browser into one integrated work environment.
ChatGPT 5.5 Reviews
In my view, a major advantage of the new model is its improved handling of tables and documents. Combined with the best infographic and image generation on the market (apart from applications dedicated to these tasks), ChatGPT 5.5 has a solid foundation for returning to the leading position.
Programming capabilities have also increased, but to notice them, you need a basic understanding of how programs work.
Recurring themes emerge among reviewers and users that are worth summarizing.
A definite plus is the leap in agentic capabilities. The model handles multi-step tasks much better without constant guidance. Its improved understanding of long contexts is also a leap rather than an evolution. Token efficiency is real and appreciated by developers paying for the API.
A downside is the API’s doubled price, which is particularly painful for smaller teams. Hallucinations in overly confident answers are a risk that experienced developers point out. In real-world programming tasks based on GitHub repositories, Claude Opus 4.7 is still better. Reviewers also note that the model performed worse with full-stack applications in practical tests.
Developers’ reaction is measured rather than euphoric. Some argue that the leap is greater than the version number alone suggests, because this is a full retraining of the base model, not another layer of post-training. But experienced practitioners are tempering the enthusiasm: Terminal-Bench is not the same as SWE-Bench.
An honest assessment of GPT-5.5 is this: it is a model that does exactly what it is designed to do—agentic work with tools in terminal and office environments. If your workflow involves long, multistep tasks requiring autonomous action, GPT-5.5 is probably the best option available. If you work with complex code repositories, solve bugs spanning multiple files, or need deep reasoning without scaffolding, Claude Opus 4.7 still offers stronger capabilities.
Both models define the current state of AI capabilities. Choosing between them is not a question of which is absolutely better, but of which is better suited to a specific application.