Skip to content
Back to all articles
By TJ SoftUpdated:

ChatGPT 5.4. OpenAI climbs out of its slump.

Chat gpt 5.4. Open Ai wychodzi z dołka.

Not long ago, the technology industry was beginning to embrace the belief that dark clouds were gathering over OpenAI. While Google was successfully developing its Gemini series, backed by the entire Google ecosystem, and Anthropic was setting new standards for AI usability with its Claude Opus model, OpenAI seemed to be resting on its laurels, releasing only minor updates.

The launch of GPT-5.4 breaks this streak, introducing a model designed from the ground up for advanced professional tasks. It marks a transition from simple chat to an autonomous assistant capable of carrying out complex engineering and analytical processes with minimal human supervision.

At the heart of the changes are “Thinking” mode and a context window of 1 million tokens (in the Codex version). This enables the analysis of entire documentation archives at once and explicit planning of actions before carrying them out, which drastically reduces the risk of errors (Gemini and Claude have had a context window of this size for quite some time). ChatGPT-5.4 operates as an agentic system—it can independently manage multi-step tasks and operate directly within the operating system, selecting the appropriate tools in real time.

Progress is confirmed by concrete data: 83% effectiveness in professional tasks (GDPval benchmark) and a 33% decrease in factual hallucinations compared with version 5.2. Thanks to improved reasoning and “learning humility,” the model is less likely to make up answers, becoming a reliable tool in sectors requiring the highest precision, such as finance and software engineering.

Table of Contents

ChatGPT 5.4 in a nutshell

It’s not just about the benchmarks, because we’ve become accustomed to the fact that reaching the top is nothing groundbreaking for the largest models, and often, after one model was released, another company would launch the latest version of its own model, which would once again take the top spot. The very nature of benchmarks is also a problem: if a given test is well known, model developers can optimize their solution for those benchmarks.

GPT-5.4 is a breakthrough model focused on professional work, integrating top-tier coding capabilities (from the Codex line) with advanced reasoning and native computer control, enabling agents to work autonomously in systems and applications. The key new feature in ChatGPT is the “Thinking” mode, which presents an action plan before providing the full answer and enables real-time model correction. Combined with a context window of up to 1 million tokens and an intelligent tool-search function, this dramatically increases efficiency while using fewer resources than version 5.2. The model sets new records in real-world task evaluations (achieving 83% performance relative to human experts in the GDPval test), while also offering 33% fewer hallucinations, significantly improved presentation aesthetics, and a powerful leap in spreadsheet modeling quality.

According to official data, the probability of errors in individual facts (hallucinations) decreased by 33% compared to version 5.2, while the overall number of incorrect answers decreased by 18%.

It is worth recalling that the launch of the “basic” version of GPT 5 was, for many, a moment of particular disappointment. Although GPT-5 was objectively a solid and efficient tool, it failed to bear the enormous weight of expectations that had built up around OpenAI following the successes of versions 3.5 and 4o, as well as the specialist models from the “o” series. The industry, fueled by visions of imminent AGI, expected a revolution, but received only evolutionary refinements that, when compared with the momentum of Claude and Gemini, seemed conservative. OpenAI drew conclusions from the previous launch, and the new model is being released almost out of the blue.

ChatGPT must be powerful and must have unique features and advantages to survive competition from Google, Claude, Grok, and Llama.

For a long time, ChatGPT had virtually no competition that posed a threat to it. The previously mentioned Gemini, Grok, and Claude have been creating products that rival ChatGPT for quite some time, but OpenAI had obvious advantages stemming from the fact that it was first to market. ChatGPT became synonymous with language models; the program also offered the most features (such as projects).

Google was unable to catch up with the competition for a long time, which came as a surprise to many—in the end, for two decades the company had been regarded as a leader in algorithms and AI. After all, it was Google Translate (launched in 2006) that was a groundbreaking tool, while the research paper “Attention Is All You Need,” published in June 2017 and developed by researchers from Google Research and Google Brain, laid the foundation for today’s transformer-based language models.

For about a year, Google has returned to the forefront, and for the past few months it has surpassed ChatGPT in many respects. Google has several key advantages:

  • Access to existing algorithms and a vast database (YouTube, Maps, Business, search engine).
  • Ecosystem – Google Drive, Google Docs, Gmail, operating system.

The first point offered a vision in which Google’s AI would leave the competition behind in the future, so it is worth implementing and investing in these systems. The second point means that many people and companies have access to paid AI “as a side benefit,” for example by using Google Cloud.

So if you can spend $20 a month and thereby gain access to Google Drive, Workspace storage, etc., there is little reason to spend an additional $20 on a slightly better ChatGPT.

The Rest of the Competition: Claude, Grok, Llama

For people who care about more ethical and creative AI, Claude is the default choice. Anthropic is a company founded by former OpenAI employees, in which Google has invested heavily. It is easier for a smaller company to remain authentic. OpenAI is no longer seen as “a cool startup that did something groundbreaking.” Exaggerated promises about the capabilities of new models, along with bizarre financial dealings with Oracle and Nvidia, have led the company to be viewed as yet another huge, soulless corporation.

Chart of ChatGPT and Claude app downloads.

Next, we have Grok, whose advantage is “Twixer” (formerly Twitter, now X), as well as Teslas and robots (Tesla Optimus). The final piece of the puzzle is Meta’s products, namely the more open models in the Llama series, as well as other open-source models developed by a number of smaller institutions.

Selecting a thinking mode and adding new commands while ChatGPT is working, without having to interrupt the session.

It is also worth mentioning the latest changes to the interface and operating logic of GPT-5.4, where four distinct “thinking” modes have been introduced. In this setting, the model does not generate a response immediately; instead, it independently initiates a series of commands, compares the results obtained, and only then constructs the final content based on the “mini-research” it has conducted.

The key innovation, however, concerns the UX (User Experience) layer itself. OpenAI has introduced a feature that allows the machine’s thought process to be adjusted in real time. If, while the model is working, we decide that something should be added or notice that the analysis is heading in the wrong direction, we can send a new instruction without having to interrupt the session or restart the generation. This dynamic clarification makes the interaction more seamless, although—as the rankings show—even such conveniences are not enough to silence criticism regarding the system’s ethics and overall performance.

ChatGPT 5.4 is a powerful tool for programming (vibe coding)

OpenAI’s products performed very well in programming, and the leap in coding quality was GPT-5’s main advantage (progress was not nearly as significant in other areas). Recently, however, Google has begun to undermine OpenAI’s position, mainly through the Antigravity application—an integrated source code editor for Gemini and Claude. Gemini 3 (along with newer versions) is particularly good at frontend development and the visual aspects of applications. Claude, in turn, has a reputation as an LLM that handles complex logical problems in code exceptionally well. Both Gemini and Claude also offer a much larger context window than OpenAI’s older models.

With version 5.4, however, we see another major leap in programming capabilities. When we combine this model with Codex (OpenAI’s programming agent), it can work on entire projects, creating and modifying multiple files at once. It is worth mentioning that a dedicated Codex app can now be downloaded – for a long time, GPT required a direct connection to a GitHub repository, and for many people, downloading a separate app may be easier than configuring GitHub. We can also connect Codex to advanced development environments such as Visual Studio.

A good mini-benchmark of coding ability is creating simple, aesthetically pleasing programs, especially those operating on three-dimensional objects. Such tests are easy for anyone to understand, even without programming experience. The following video provides a good illustration of the model’s real capabilities:

With just a few prompts, GPT-5.4 can create a visually polished program that displays meteorological data from around the world on an interactive 3D model.

Adding 3D previews of various cities is not a problem either:

Of course, GPT did not build the 3D engine and 3D objects from scratch, but simply used existing libraries and solutions; nevertheless, it is still an impressive achievement.

GPT 5.4 can independently work with 3D objects. Here, for example, it created a complex object based on a photo:

And here he created a mini-simulator in which objects can be moved and which handles lighting and reflections correctly. Simulating a mirror in 3D space is quite a challenging programming task, which is why many games forgo mirrors and realistic reflections.

ChatGPT 5.4 is adapted for working with Excel; a special plugin will be released shortly.

ChatGPT-5.4 is optimized for working with spreadsheets. Alongside the latest release, an Excel plug-in was launched, although for now it is available only in beta. This can be confusing, because for several years there have already been many “GPT for Excel” plug-ins created by third-party companies. Some of them are reliable, some do not work, and some may even be dangerous.

Thus, OpenAI is following in the footsteps of Anthropic, whose Claude model has had Excel integration for some time. Gemini’s integration with Google Sheets is rather limited; in spreadsheets, we can discuss our data with AI, get advice on formulas, and, at best—when English is set as the language—generate a simple table to insert into the spreadsheet. The situation is similar with Microsoft’s Copilot (Windows), although it is even worse here, as it is not even possible to insert simple tables.

GPT-5.4 represents a fundamental shift in OpenAI’s development strategy, moving away from general versatility toward narrow, professional specialization. The best evidence of this evolution is its performance on the GDPval benchmark, which tests artificial intelligence’s ability to perform real-world tasks across dozens of different professions. The model not only matches human experts but actually surpasses them, achieving better or equal results in an astonishing 83% of cases. In the world of finance, this means moving from generating simple definitions to creating advanced analytical structures that previously required hours of work from qualified employees.

Great help in investment analysis

One of the most impressive achievements of the new architecture is its performance on tasks typical of junior investment banking analysts. In tests simulating complex spreadsheet modeling, GPT-5.4 achieved an accuracy rate of 87.5%, representing a significant qualitative leap over previous versions. Thanks to native plugins for Excel and Google Sheets, the model can now navigate cells independently, understand complex relationships between formulas, and build growth forecasts directly within the professional’s work environment. This solution eliminates data-copying errors and enables automatic verification of the consistency of financial models with minimal user supervision.

Processing Gigantic Data Sets

The true power of GPT-5.4 in a business context becomes apparent when working with documentation spanning long time horizons. Thanks to its context window of up to one million tokens, the model can “read” and correlate years of financial report archives or thousands of invoices without losing the thread. It operates as an autonomous agent capable of extracting key performance indicators from hundreds of pages of documentation and synthesizing them into a professional PDF report. Crucially, the model performs these operations while consuming significantly fewer tokens than older versions, making the most complex financial audits not only more accurate but also less expensive and faster to complete.

Fewer hallucinations

In the new GPT-5.4 model, OpenAI has focused on deep architectural changes that directly address AI’s biggest pain point—hallucinations. Thanks to new technologies, the likelihood of errors in individual facts has fallen by as much as 33% compared with version 5.2, while the overall number of incorrect answers has decreased by 18%.

This success is built on several key pillars:

  • “Thinking” mode and an explicit reasoning plan: This is the most important process change. GPT-5.4 does not immediately produce text; instead, it first constructs an internal plan that is visible to the user. The model “thinks” about the structure of the response, allowing it to catch logical errors before formulating them. Moreover, as a user, you can intervene during this process, correcting the machine’s direction of thought at its source.

  • Agentic web research: The Deep Web Research system is no longer a simple link search engine. It now operates as an autonomous agent that verifies information across multiple sources simultaneously. As a result, the model is much better at identifying details hidden deep within source texts, which previously often led to “making up” missing facts.

  • Large context window: Increasing the memory to 1 million tokens in the Codex and Pro versions has radically improved factual consistency. Hallucinations often resulted from AI “forgetting” data provided at the beginning of a long session; with such a huge window, the model has constant access to the entire body of material, so it does not need to guess the context.

  • Confidence calibration and “learning humility”: Thanks to improved training (RLHF), GPT-5.4 can better assess the extent of its own knowledge. It has been programmed to admit when it does not know something more often, instead of generating merely plausible-sounding falsehoods. In extremely difficult situations, the model can also activate Extreme Reasoning mode, devoting more computing power to verifying complex scientific issues.

Interestingly, this greater substantive depth also made it possible to reduce unnecessary refusals. Previous models, afraid of making a mistake, often took the cautious route and refused to answer. GPT-5.4, thanks to a better understanding of its own limitations and data, can navigate facts more safely, less often hiding behind generic disclaimers.

It looks like OpenAI is finally teaching its child that “I don’t know” is sometimes the best answer—paradoxically making it much smarter in the process.

What About GPT 5.3?

Many users may be wondering what happened to version 5.3, which passed through the market without as much fanfare as the latest release. In reality, GPT-5.3 was a classic, straightforward update focused almost entirely on performance optimization and latency reduction. While version 5.2 laid the foundations and 5.4 brought powerful reasoning capabilities, 5.3 served as the engine designed to make interaction with AI almost instantaneous. It was then that OpenAI introduced the “/fast” mode, allowing developers and analysts to iterate on code at lightning speed without waiting for the model’s lengthy “thinking” processes.

That is precisely why the GPT-5.4 model is compared to GPT-5.2.

Current Problems with ChatGPT and OpenAI

The technology world was shaken by reports of OpenAI’s controversial decisions, which cast a shadow over the company’s ethical foundation. An increasing number of users are choosing to abandon ChatGPT, and the reason is extremely serious: a choice between humanity’s safety and defense contracts. While competitor Anthropic refused to cooperate with the Pentagon on the development of “killer bots” and mass-surveillance tools—and was consequently placed on a government blacklist—OpenAI stepped into its place without hesitation, establishing a close relationship with the military sector.

The backlash against OpenAI is gaining momentum under the banner of the #QuitGPT movement, which has already collected nearly 2.5 million pledges to cancel subscriptions. Users, disappointed with the tech giant’s stance, are migrating en masse to Claude. Anthropic’s app has remained at the top of the App Store, becoming a new home for those who value ethics and data security more highly than aggressive market expansion.

For many people, moving away from ChatGPT is not only a political statement but also a logistical challenge. Securing their existing work and transferring the workflows they have developed to new platforms is becoming crucial. Guides on “future-proof” AI memory architecture are gaining popularity, enabling users to avoid dependence on a single provider and securely manage their data, regardless of whether they ultimately choose Claude or another solution that upholds ethical standards.

Decline in Form: GPT Loses Ground in the Rankings

OpenAI’s image problems go hand in hand with a clear technological struggle, as confirmed by the latest LMM Arena rankings. In terms of quality, GPT models have ceased to dominate, ceding ground to the competition in nearly every key category. In the overall text (Text) ranking, OpenAI’s flagship models, such as gpt-5.4-high, occupy only distant positions, being overtaken not only by the Claude family but also by Gemini models.

The situation is even worse in specialized applications:

  • Coding (Code): The podium was completely dominated by Anthropic models (Claude Opus and Sonnet), while the highest-ranked GPT model is only in 6th place.
  • Document Analysis (Document): Here too, Claude Opus 4-6 leads the way, while OpenAI falls outside the top five, losing to solutions from Google.
  • Visual Tasks (Vision): Gemini’s dominance in this category is undisputed—Google’s models occupy the entire podium, pushing GPT into 4th place.

Of course, it is important to remember that the results of the LMM Arena should be approached with a degree of caution. Rankings based on crowdsourcing and users’ subjective impressions may be susceptible to temporary trends or specific prompting techniques. Nevertheless, when such a vast sample of voters consistently points to the superiority of Claude and Gemini, it becomes a clear market signal: OpenAI has lost its position as the undisputed leader in quality, and the “king of AI” is currently on the defensive.

Summary

The launch of GPT-5.4 marks the moment when OpenAI definitively ends its defensive phase and shifts to a determined counteroffensive. After a period in which competing ecosystems from Google and Claude’s precision from Anthropic began to seriously erode the leader’s position, the San Francisco giant unveiled a model that is no longer merely an “intelligent conversationalist,” but has become an autonomous professional partner. Its focus on professional work, integration with tools such as Codex, and native computer-use capabilities mean that GPT-5.4 sets a new benchmark for the usefulness of artificial intelligence. This is no longer just a race based on benchmarks, but a real battle over who can deliver a tool capable of independently carrying out complex, multi-stage tasks with minimal human supervision.

The greatest breakthrough brought by the new update is a dramatic improvement in substantive accuracy. A 33% reduction in hallucinations and the introduction of a “Thinking” mode with an explicit reasoning plan are changing the dynamics of working with the model. The ability to correct the AI’s actions in real time, without interrupting its work, is a revolution in user experience (UX), eliminating the frustration associated with misinterpreting instructions. Combined with a massive context window of up to one million tokens and powerful capabilities in programming and financial modeling, GPT-5.4 is becoming a system that can be entrusted with critical business processes. An 83% effectiveness rate compared with human experts in professional tasks is a clear signal that AI is reaching a level of competence previously reserved for qualified specialists.

A Return to Our Roots in a Modern Style

For many users who, like me, started looking toward Gemini because of its speed and integration with the Google ecosystem, GPT-5.4 is a powerful argument for returning to OpenAI. Although Google’s advantage in access to data and office tools is undeniable, the unique features of version 5.4, such as advanced coding, native Excel plugins, and agentic web search, offer an analytical depth that the competition still lacks. It seems that OpenAI has learned a lesson from its overly cautious launches in the past. The new model is no longer trying merely to charm us with fluent speech; instead, it is learning humility and precision, becoming a tool tailored to the specific, difficult challenges of today’s job market. OpenAI has not merely climbed out of its slump—it has just charted a new direction for the development of the entire industry.