ChatGPT 5 Reviews and News
A few days ago, the long-awaited ChatGPT-5 premiered. Since then, a few things have already changed. Don’t like the new model? Access to GPT-4o has been restored. However, we currently do not have access to the other older models. In addition, we have been granted access to GPT PRO.
Table of contents
What Was ChatGPT 5 Supposed to Be?
The first announcements of ChatGPT 5 said that it would change the way the world works. This model was supposed to reach the level of AGI (Artificial General Intelligence), meaning it would solve tasks like a human, relying more on reasoning than on the knowledge it possesses. Current models can sometimes “come up with something” or find a solution that was not described, but they are unable to do this on multiple levels or develop a given line of thought over a longer period of time. These announcements were made when we had witnessed a major leap in quality between GPT 3.5 and GPT 4o.
Later announcements were more restrained, as it turned out that the technology underpinning LLMs was hitting a wall, and simply increasing computing power was producing smaller gains than expected. More recent announcements about GPT-5 focused on it being “generally better” and replacing all previous models, which was particularly important because things had become rather chaotic. Choosing between 4o, 4.1, o3 Pro, o3 mini, and so on was not very convenient. The new model was supposed to detect how advanced an answer we expected and decide for itself whether to complete the task quickly or devote more time to it.
Main advantages of the latest model from OpenAI
In the remainder of the text, I will describe the biggest problems, but first, let’s focus on the advantages, which definitely do exist. The GPT-5 model is primarily faster (ChatGPT’s long response times were a problem in the past). Additionally, we have a larger context window (meaning we can have a longer conversation with the model). In quick tests, the difference is not very noticeable, but when actually using the model, it is a very significant change, especially when working on something more complex. Besides that, ChatGPT-5 is much better at coding, hallucinates less often, and performs better in many other tests. Automatically selecting the model’s complexity for a task is also a significant advantage. Further benefits include the fact that the GPT-5 model is cheaper to use, which is highly significant for anyone using the API.
Problems with ChatGPT 5
The new model was presented as a major revolution. In practice, it turned out to be a solid step forward; however, with its introduction, we lost access to older models in the ChatGPT app (older models are still available via the API, so if someone really cares about this, they can run them—in that case, instead of a monthly subscription, you pay for each prompt).
Automatic model selection does not work perfectly.
The problem lies in the mechanism for selecting the level of complexity itself (which does not always work well), as well as in the quality of the texts themselves, which is sometimes worse. We can control how much time and energy the model should devote to considering a problem by describing it in the prompt. If we ask it to do something simple but want it to think it through, we can write something like: spend a great deal of attention on this. On the other hand, sometimes it is more convenient to select the model by clicking. That is why we can currently manually choose a mode: Fast, Thinking, or Pro, which is very convenient.
Some people like the casual style of the 4o model
Additionally, the 4o mode has returned. Many people simply liked it a lot and had grown accustomed to its style. This also brings up the topic of the “human side of AI.” After the latest updates, the 4o model tended to flatter the user excessively. For many people, this was merely a minor issue, resulting in somewhat embarrassing texts (which could nevertheless be easily fixed). However, a significant problem emerged: ChatGPT reinforced people’s delusions. For example, if someone wrote that they were a god, ChatGPT would reinforce that belief. This problem was later fixed, but 4o still writes in a more “human, empathetic style.” The GPT-5 model has a more substantive, restrained style.
Random words, Chinese characters
Some people complained about random characters (Cyrillic, Chinese, or other Asian characters) being added to the text. This phenomenon is more commonly noticeable in Grok; it hasn’t occurred in ChatGPT for a long time.
The following example also illustrates the problem with understanding commands. This problem did not occur with the 4o model; when asked for H2 and H3 headings, it did not include the text “h2” and “h3” in the headings.
This also involves shortening the text. The GPT-5 model is capable of writing long texts (the context window allows it to do so), but in this case it mistakenly chose the fastest solution.
The write an article command is very useful for research and learning. Sometimes, it also provides a good foundation that can be used and expanded when writing longer texts.
Model Selection from the User’s Perspective
The most advanced users will learn the differences between specific models. This approach makes sense when someone relies heavily on AI or uses language models to build tools used by others. For obvious reasons, there are only a handful of such people, and they can use the developer version.
Apart from the most advanced users, there are many people who use AI consciously. In such a case, an approach in which ChatGPT chooses which model to use on its own is not the best.
The selection option is useful if it is logically organized and the model names are designed with users in mind (rather than resulting solely from technical considerations). If the models are arranged linearly, the choice is simple. There may also be two parallel lines.
In any case, a slider allowing users to specify how much time the model should spend completing the task would also be a convenient solution. The user is not interested in which specific solutions are used; the time and potential length of the text are more important.
Technical Aspects of ChatGPT 5
“Routed Duo” System: GPT-5 Main and GPT-5 Thinking
GPT-5 introduces an innovative “routed duo” architecture comprising two main models:
gpt-5-main
i
gpt-5-thinking
.
**gpt-5-main**
is a fast, general-purpose model designed to quickly provide answers to everyday queries.
**gpt-5-thinking**
is a model with deeper reasoning, designed for more complex problems requiring intensive thought processes and the ability to engage in internal deliberations within a “chain of thought.” The built-in router decides which model to use depending on the complexity of the conversation, the user’s intent, and tool requirements. This dynamic adaptation aims to balance efficiency with deeper reasoning, providing an optimal response in every situation. OpenAI plans to combine all these capabilities into a single, seamless model in the future, indicating an effort toward even greater unification and simplicity.
Large Context Window and Multimodal Capabilities
One of the most significant improvements in GPT-5 is the unprecedented expansion of the context window. In the ChatGPT Pro version, the model can process up to 400,000 tokens, representing a huge leap from the maximum of 32,000 tokens in GPT-4. This capability allows the model to analyze very large documents, extensive research papers, and even entire codebases within a single query—something that was previously impossible.
Furthermore, GPT-5 continues and expands the multimodal capabilities that debuted in GPT-4. In addition to processing text and images, GPT-5 can handle full video and audio files. This means that the model can interpret and generate responses based on a variety of inputs, which is crucial for creating more natural and context-aware interactions. Results from benchmarks such as MMMU and VideoMMMU confirm its outstanding performance in multimodal understanding, including visual, video-based, spatial, and scientific reasoning.
The development of multimodal AI is one of the hottest trends in 2025, and this market is expected to grow at a CAGR of over 30% between 2024 and 2032, underscoring the strategic importance of these capabilities.
Better Reasoning
GPT-5 is characterized by significant advances in reasoning and problem-solving, allowing it to perform at the level of a “PhD-level expert.” The model is far less prone to hallucinations—a common problem in previous versions, where AI provided inaccurate or misleading information. In real-world factuality evaluations,
**gpt-5-main**
demonstrates a 26% lower hallucination rate than GPT-4o, and
gpt-5-thinking
65% lower than o3 and 78% fewer answers with serious factual errors. This significant reduction in errors increases the model’s reliability and usefulness in critical applications.
This model also introduces the concepts of “slow thinking” and “deliberate reasoning” to further reduce hallucinations. Rather than relying solely on rapid response generation, GPT-5 can engage in a step-by-step reasoning process, using mechanisms such as tree search and self-evaluation to provide more reliable and accurate results. This ability to perform deeper analysis and self-correction makes GPT-5 a more reliable assistant in high-risk scenarios.
ChatGPT in Benchmarks (Comparison with Other OpenAI Models)
Test Category | Comparative Test (with tools/without) | GPT-5 (high) | OpenAI o3 (high) | GPT-4o |
Coding | SWE-bench Verified (pass@1) | 74.9% | 69.1% | 52.8% |
Aider Polyglot (pass@2) | 88.0% | 79.6% | 26.7% |
Mathematics | AIME ’25 (without tools) | 94.6% | 88.9% | – |
Multimodality | MMMU (pass@1) | 84.2% | 82.9% | 74.4% |
MMMU Pro (pass@1) | 78.4% | 76.4% | 62.7% |
VideoMMMU (max. 256 frames) | 84.6% | 83.3% | 60.9% |
Health | HealthBench Hard | 46.2% | 31.6% | 0.0% |
Instructions | Scale MultiChallenge (pass@1) | 69.6% | 60.4% | 54.9% |
COLLIE | 99.0% | 98.4% | 65.8% |
Long Context | OpenAI-MRCR: 2 needle 128k | 95.2% | 55.0% | 56.4% |
OpenAI-MRCR: 2 needles 256k | 86.8% | – | – |
In the area of coding, GPT-5 achieved 74.9% on SWE-bench Verified and 88% on Aider Polyglot, making it OpenAI’s strongest coding model to date. In mathematics, it achieved an impressive 94.6% on AIME 2025 without using tools. In terms of multimodal understanding, the model achieved 84.2% on MMMU and 84.6% on VideoMMMU, demonstrating its advanced capabilities in interpreting visual and video data.
In the field of healthcare, GPT-5 Thinking achieved 46.2% on HealthBench Hard, significantly outperforming previous models. The model also demonstrates substantial progress in following instructions and using agentic tools, achieving 69.6% on Scale MultiChallenge and 99.0% on COLLIE. Moreover, in long-context tests, GPT-5 maintains high accuracy, achieving 95.2% on OpenAI-MRCR with 128k tokens and 86.8% with 256k tokens.
GPT 5 Is Less Likely to Cheat.
ChatGPT-5 demonstrates significant progress in reducing deceptive tendencies and improving instruction following. The model
gpt-5-thinking
significantly less frequently exhibits deceptive behavior in agentic tasks with missing capabilities, broken tools, or missing inputs. For example, in coding deception tests, its score was 0.17 compared to 0.47 for o3, and in browsing tests with broken tools, 0.11 compared to 0.61 for o3. Monitoring the “chain of thought” (CoT) in production results in approximately 2.1% deceptive responses for
gpt-5-thinking
compared to 4.8% for o3.
OpenAI also reduced the model’s tendency toward sycophancy, that is, excessive flattery or validation of negative emotions, which was a problem in previous versions. Although GPT-5 is significantly better in this regard, a certain “yes-man” tendency may still lead to misinformation in sensitive contexts such as health or politics.
Models are trained to follow the instruction hierarchy: system > developer > user. Although the results for
gpt-5-thinking
are strong,
gpt-5-main
shows some regressions in certain tests, which are scheduled to be fixed. Resistance to “jailbreaks” (tricks intended to bypass safeguards) has been improved, but the model is not entirely infallible, and targeted, multi-step attacks may still work in individual cases. It should also be noted that, despite the progress, GPT-5, like previous models, may still perpetuate existing biases from the data on which it is trained, requiring ongoing attention and intervention.
How to prevent a language model from providing instructions that would enable someone to build a bomb.
OpenAI is approaching the potential risks associated with GPT-5 with great caution, particularly in high-risk domains. The model
gpt-5-thinking
was classified as “highly capable” under the preparedness framework for biological and chemical applications. Although there is no definitive evidence that the model could help a novice cause serious biological harm, OpenAI has adopted a precautionary approach.
A two-tier real-time monitoring system has been implemented, consisting of a fast topic classifier and a deeper reasoning monitor that operates based on a detailed taxonomy of biological threats. This is complemented by account-level enforcement and an API security identifier for developers, aimed at ensuring control over potential misuse. The ChatGPT agent feature has also been classified as high risk.
There are concerns that AI models such as GPT-5 could be used to create malware, generate phishing scams, or assist in the development of chemical or biological weapons. Although the model may partially assist with hacking tasks, it is not powerful enough to independently carry out serious attacks on well-secured systems. Nevertheless, combined with human expertise, it may pose a risk to poorly protected targets. The safety of image inputs is also monitored to prevent the creation of dangerous content in combination with text, although the system is not infallible.