{"id":13205,"date":"2025-08-07T22:05:48","date_gmt":"2025-08-07T20:05:48","guid":{"rendered":"https:\/\/haimagazine.com\/uncategorized\/gpt-5-more-intelligent-but-is-it-smarter\/"},"modified":"2025-08-08T09:42:34","modified_gmt":"2025-08-08T07:42:34","slug":"gpt-5-more-intelligent-but-is-it-smarter","status":"publish","type":"post","link":"https:\/\/haimagazine.com\/en\/ai-news-2\/gpt-5-more-intelligent-but-is-it-smarter\/","title":{"rendered":"\ud83d\udd12 GPT-5: More intelligent, but is it smarter?"},"content":{"rendered":"

The OpenAI summer offensive was like a meticulously planned military operation. First, in mid-July, ChatGPT Agent<\/strong> took the stage \u2014 a tool capable of roaming the internet independently, filling out forms and working with data from multiple sources. Shortly thereafter, Study Mode<\/strong> appeared \u2014 a mode in which ChatGPT guides users through the complexities of logical reasoning. In early August, the company revealed more cards: two large open-source language models, handing powerful tools to the community.<\/p>

However, all this was merely a prelude. The carefully built tension reached its climax on August 7, when GPT-5<\/strong> was unveiled. Instantly available to everyone via a web version and through an API, in three power variants (GPT-5, GPT-5 mini and GPT-5 nano), the new model is more than just a faster and more accurate text generator. While the official narrative focuses on the slogan “useful, smart, fast”, it’s worth examining how well the promises match the model’s actual capabilities, and what shifts in artificial intelligence usage standards GPT-5 brings.<\/p>

Why is GPT-5 significant?<\/strong><\/h4>

Where is the biggest change? In the mechanism OpenAI calls a “router”. It’s an internal dispatcher that analyzes the complexity of the task and decides whether a routine response suffices or if it needs to activate a deep processing mode \u2014 “think hard”. Users don’t have to ask for anything anymore. The model itself knows when to stop improvising and start “reasoning”. In practice, this means a qualitative leap in tasks requiring multi-stage analysis, writing complex code or crafting coherent arguments. At the same time, the model has gained a new valuable skill: it can admit ignorance when the input data don’t allow for a credible response.<\/p>

In its particular style, OpenAI coined this change with a catchy slogan: “reasoning at a PhD level.” It’s a marketing promise that should be taken with a grain of salt. A real PhD is not just about proficiency in processing information but also about awareness of social context, intentions, personal limits and the ability to reflect critically. GPT-5 undoubtedly takes a step towards deeper processing, but attributing it with an academic degree is more a clever metaphor than a measurable fact.<\/p>

The true significance of GPT-5 doesn’t lie in benchmarks or marketing slogans, but in the fact that it redefines the relationship between humans and machines. We’ve gained a partner who can not only think longer but most importantly knows when it’s necessary. Besides, it’s not afraid to say, “I don’t know.”<\/p>

What can the new model do?<\/strong><\/h4>

The most significant change in GPT-5, the aforementioned “router” mechanism, is the engine that drives all other innovations. The model no longer treats every task equally, but intelligently manages its computing power. What does this actually mean?Firstly, writing and programming are elevated to a higher level of collaboration<\/strong>. GPT-5 understands the text’s structure and the author’s intent much better. Whether it’s analyzing data, editing an official statement or creatively brainstorming over an essay, the model not only generates text but also suggests changes, inquires about purpose and context. The same goes for code. GPT-5 maintains context in complex projects, fixes errors and suggests optimizations more efficiently. Combined with an Agent that can independently launch a testing environment, it can become a digital partner in coding.<\/p>

Also, we gain more control over the style and form of interaction<\/strong>. Users can now choose from four conversation modes: precise, concise, creative or detailed. This simple tool allows you to tailor the nature of responses to your current needs. A cosmetic detail has also been added \u2014 the ability to change interface colors, signaling a move towards greater personalization.Finally, the voice becomes a tool for learning<\/strong>. The enhanced Voice Mode now includes a “Learn with me” feature. Conversations with AI no longer consist of a series of one-time questions. Instead, they resemble a session with an interactive tutor who guides you through the topic step by step, patiently explaining and answering additional questions. This feature, recently available only in written conversation mode, is designed for those who want to use AI to genuinely expand their skills and is also a powerful tool for learning languages, for instance.All these changes \u2014 deeper reasoning, better writing, more advanced coding and adaptive voice support \u2014 shift the focus from “generating responses” to a collaboration based on understanding the user’s goals.<\/p>

Measuring intelligence: what do the numbers tell us?<\/strong><\/h4>

Marketing claims are one thing, but the real measure of quality for large language models comes from independent benchmarks. They test AI in realistic, complex tasks where it’s not just the right answer that counts, but also resilience to errors and the ability to engage in logical reasoning. GPT-5’s results in these tests show a clear pattern.<\/p>

The progress in programming tasks is particularly impressive. In the SWE-Bench Verified<\/strong> test, which involves solving real-world problems from open-source repositories, it achieves an effectiveness of 74.9%<\/strong>. In Aider Polyglot<\/strong>, it correctly edits code in multiple programming languages in 88%<\/strong> of cases. Meanwhile, in SWE-Lancer<\/strong>, simulating a freelancer’s work, it dominates the competition, smoothly transitioning from the brief to implementation.<\/p>

GPT-5 also records the highest scores in competition tasks. In the AIME 2025<\/strong> test, it achieves 94.6%<\/strong> without additional tools and as much as 99.6%<\/strong> in “thinking” mode. In the prestigious HMMT (Harvard\u2013MIT)<\/strong> tournament, its effectiveness reaches nearly 100%<\/strong>. In the comprehensive Humanity\u2019s Last Exam<\/strong> covering expert questions from many fields, GPT-5’s dominance is overwhelming. With deep thinking mode enabled, it scores 42%<\/strong>, while GPT-4o only reached a level of 5.3%<\/strong>.<\/p>

\"\"
Source: OpenAI<\/figcaption><\/figure>
\"\"
Source: OpenAI<\/figcaption><\/figure>
\"\"
Source: OpenAI<\/figcaption><\/figure>

The model demonstrates equally high competencies in high-risk fields. In the HealthBench Hard<\/strong> test, which checks medical knowledge in complex diagnostic scenarios, GPT-5 achieves results comparable to those of a human specialist, while showing a low tendency to generate false information.<\/p>

In benchmarks, GPT-5 not only outclasses its predecessors in simple tasks but also gains a significant advantage where complexity increases. In tests that require a deep understanding of context and multi-stage reasoning, it clearly surpasses both GPT-4 and other competitive models. Although no test fully reflects real-world usage, the numbers don’t lie: GPT-5 is better tuned for tasks where there’s no room for error.<\/p>

Trust architecture: how secure is GPT-5?<\/strong><\/h4>

With such immense power, safety is no longer an add-on but becomes the foundation. OpenAI states that GPT-5 has undergone over 5000 hours of rigorous testing, where internal and external teams played the role of “adversaries” trying to break the model’s security.<\/p>

The key element is the safe completions<\/strong> mechanism \u2014 a built-in safety brake that limits the generation of risky content in areas such as health, finance or law. GPT-5 tends to remain silent or signal uncertainty rather than provide an answer that could be misleading much more often than its predecessors. Stricter rules have also been introduced for medical scenarios, where the model, instead of making diagnoses, refers to credible sources.<\/p>

What’s critically important from the user’s perspective and data security: GPT-5 doesn’t learn from our conversations in real time. As stated in the documentation, individual interactions don’t update the model’s weights or internal knowledge. It remains a static model, whose knowledge base is frozen in time. The optional “memory” (memory<\/em>) feature allows for storing context from previous conversations, but this isn’t a form of continuous learning \u2014 the model doesn’t change its “understanding” of the world. OpenAI notes that it may use aggregated anonymous data to improve future versions, but never based on a single session. Such architecture, standard in commercial implementations, intentionally limits the risk of accidental leakage of confidential information and ensures that the model behaves in a predictable manner.<\/p>

Despite these safeguards, it’s important to maintain a healthy distance. No model is 100% secure, and the ability to generate unpredictable responses remains a real challenge. GPT-5 is getting closer to knowing when it should remain silent, but the ultimate responsibility for critical decisions still rests with us people.<\/p>

Availability<\/strong><\/h4>

With the launch of GPT-5, OpenAI is changing the distribution rules. The model is immediately available to everyone, but the company is abandoning the manual version selection we knew from the times of GPT-3.5 and GPT-4. Now the system decides which computing power we’ll use.<\/p>

In the ChatGPT interface, the model is set to default and its variant depends on the plan you have:<\/p>

  • Free users<\/strong> gain access to the flagship model GPT-5<\/strong>, but once they exhaust their response limit they’ll be switched to GPT-5 mini<\/strong>, with natural performance limitations.<\/li>\n\n
  • Plus plan subscribers<\/strong> ($20 USD\/month) work on the full version of GPT-5<\/strong> with a significantly higher usage limit.<\/li>\n\n
  • Pro plan<\/strong> includes access to GPT-5 Pro<\/strong>, a variant optimized for the most demanding tasks and longer context.<\/li><\/ul>
    \"\"
    Source: X<\/figcaption><\/figure>

    Developers have three variants available through the API: from the lightest nano<\/strong>, through the balanced mini<\/strong> to the full-sized GPT-5<\/strong> for the most complex applications.<\/p>

    \"\"
    Source: OpenAI<\/figcaption><\/figure>

    Between promise and reality<\/strong><\/h4>

    The communication surrounding the launch of GPT-5 was skillfully managed. A series of announcements under the “OpenAI Summer Update”, the swift introduction of the Agent, learning mode and open-source models \u2014 all this built a narrative about a company that’s accelerating and opening up to users. But behind this fa\u00e7ade, there are several issues worth clarifying.<\/p>

    The most important of these is the already mentioned comparison to a “person with a PhD.” While catchy, this phrase intentionally blurs the line between machine processing and human expertise. The model doesn’t understand the world, lacks consciousness, and is incapable of autonomous thinking. It processes data, looking for patterns and correlations. If “PhD” is supposed to signify precision, depth and cognitive accountability, GPT-5 remains merely a tool, not an peer for discussion.<\/p>

    The ultimate test for GPT-5 won’t be meticulously selected benchmarks or a skillfully conducted marketing campaign. The real verification will come when the model reaches the masses \u2014 companies, schools and the screens of millions of users who will test its actual value in everyday tasks. For now, we’re left to observe a fascinating spectacle: on one side, promises of a revolution; on the other, the market’s first cautious reactions. What emerges from this picture is not just a portrait of new technology but above all a strategy of a company that is redefining the rules of the game in the artificial intelligence market.<\/p>

    What’s next?<\/strong><\/h4>

    Does the premiere of GPT-5 mark a new chapter? OpenAI is clearly changing its strategy: instead of giving users a choice between models, it shifts the decision-making burden to the system, which selects the appropriate power itself. At the same time, the company emphasizes safety, consciously opting out of real-time model learning. This is a key decision that prioritizes stability over dynamics.<\/p>

    What are the next steps? A natural direction seems to be further development of the agent architecture, deeper integrations with external applications and an emphasis on multimodality \u2014 the ability to smoothly process text, image and sound within a single system. There’s also increasingly advanced personalization on the horizon, which will allow the model to better understand the user’s context and needs.<\/p>

    GPT-5 is more of a milestone on the path to creating an architecture that increasingly resembles less of a tool and more of an operating system that enhances human thinking. How we leverage its potential and the limits we set for it \u2014 technological, ethical, and legal \u2014 depend not only on its creators but on all of us.<\/p>

    \"\"
    Source: YouTube<\/figcaption><\/figure>

    At the end of the OpenAI Summer Update, a nice Polish touch \u2014 Jakub Pachocki, Chief Scientist at OpenAI, thanked the entire team for their dedication.<\/p>","protected":false},"excerpt":{"rendered":"

    GPT-5 has arrived. We’ve checked out what it can do and why its ability to “think deeper” is crucial.<\/p>\n","protected":false},"author":6,"featured_media":12608,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"rank_math_lock_modified_date":false,"footnotes":""},"categories":[813],"tags":[707,830,829],"popular":[],"difficulty-level":[36],"ppma_author":[665],"class_list":["post-13205","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai-news-2","tag-ai-5","tag-chatgpt-2","tag-openai-2","difficulty-level-easy"],"acf":[],"authors":[{"term_id":665,"user_id":6,"is_guest":0,"slug":"ewa-chamczyk","display_name":"dr Ewa Chamczyk","avatar_url":"https:\/\/secure.gravatar.com\/avatar\/d8344b96a8892a3acc4586dbbf4e6175ac6982b50001b652b02295d2d7e0dd34?s=96&d=mm&r=g","first_name":"dr Ewa","last_name":"Chamczyk","user_url":"","job_title":"","description":"Co-founder & Head of AI Books @CampusAI, z pasj\u0105 pisze i m\u00f3wi o AI i kulturze"}],"_links":{"self":[{"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/posts\/13205","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/users\/6"}],"replies":[{"embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/comments?post=13205"}],"version-history":[{"count":1,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/posts\/13205\/revisions"}],"predecessor-version":[{"id":13206,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/posts\/13205\/revisions\/13206"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/media\/12608"}],"wp:attachment":[{"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/media?parent=13205"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/categories?post=13205"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/tags?post=13205"},{"taxonomy":"popular","embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/popular?post=13205"},{"taxonomy":"difficulty-level","embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/difficulty-level?post=13205"},{"taxonomy":"author","embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/ppma_author?post=13205"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}