The GPT-5 controversy

Dear Readers,

Hardly any other release has electrified – and polarized – the AI world as much as GPT-5. Within hours of the livestream announcement, superlatives and criticism flew back and forth like punches in a boxing ring: PhD-level intelligence, drastically fewer hallucinations, multimodal brilliance – and yet frustration over rate limits, “generic” responses, and faulty charts. The big question: Are we witnessing the first peak, where the pace of progress is slowing down, or just the teething pains of a huge technological leap?

Today, we take a deeper look behind the shiny façade: we break down the technical innovations, shed light on the downsides of the launch, and ask what the controversy reveals about the industry's maturity. Also in focus: the strategies behind OpenAI's mass market approach, the importance of expectation management in the AI industry, and why GPT-5 – despite all the criticism – is the first point of contact with a “thinking” assistant for millions of people. If you want to know whether this is the beginning of a revolution or an evolution, read on.

All the best,

On August 7, 2025, the moment had arrived: After months of speculation and an unprecedented hype cycle, OpenAI announced GPT-5 in a livestream – the model that was supposed to usher in the next generation of artificial intelligence. But just a few hours after its release, deep cracks began to appear in the seemingly perfect facade. While Sam Altman spoke of a model that worked “like a team of experts” and seemed “less like AI and more like a helpful friend with PhD-level intelligence,” a wave of criticism swept across social media. A Reddit thread titled “GPT-5 is horrible” garnered nearly 3,000 upvotes and over 1,200 comments within hours.

This discrepancy between promise and reality raises fundamental questions about the current state of AI development. GPT-5 not only represents OpenAI's latest technological leap forward, but also exemplifies the challenges facing an entire industry that is caught between groundbreaking innovations and the relentless laws of physics and mathematics. The controversy surrounding GPT-5 reveals a paradox: while the model achieves new best performances in benchmarks, it simultaneously disappoints those power users who have been pushing the limits of language models for years.

Are we at a turning point in AI development, where the era of exponential improvements is reaching its natural limits, or are these merely the teething pains of a fundamental breakthrough that will only reveal its true strength over time?

TLDR covers the best tech, startup, and coding stories in a quick email that takes 5 minutes to read.

No politics, sports, or weather (we promise). And it's read by over 1,250,000 people!

Subscribe for free now and you'll get our next newsletter tomorrow morning.

GPT-5 undoubtedly brings impressive technical innovations that clearly distinguish it from its predecessors. The most striking feature is its integrated reasoning ability—the model can now “think” before it responds, similar to what has already been implemented in the o-series models. These thought processes are visible to users and allow them to understand the model's decision-making processes. GPT-5 is “smarter across the board, providing more useful responses across math, science, finance, law, and more,” as OpenAI emphasizes in its official announcement. In other words, OpenAI has succeeded in further developing reasoning and achieving top scores in numerous benchmarks.

Particularly noteworthy are the improvements in hallucination reduction. GPT-5 shows “a sharp drop in hallucinations—about six times fewer than o3—marking a clear leap forward in producing consistently accurate long-form content.” This development addresses one of the most fundamental problems of large language models and could revolutionize their use in critical applications.

From a technical perspective, GPT-5 has a significantly expanded context window of up to 272,000 tokens input (including reasoning) up to 128,000 tokens output and up to 400,000 tokens in a combined context via API– a considerable increase over its predecessors. In long contexts, “GPT-5 outperforms o3 and GPT-4.1, by a margin that grows substantially at longer input lengths.” At the same time, the pricing structure has been optimized: At $1.25 per million input tokens and $10 per million output tokens, GPT-5 is positioned as a cost-effective solution for developers.

The architecture combines different modalities in a single model for the first time. GPT-5 can process text, images, audio, and code in an integrated workflow – a consolidation that previously required separate models such as DALL-E and Whisper. This unified AI approach could fundamentally simplify the development of AI applications and open up new fields of application.

Despite its impressive specifications, the launch of GPT-5 revealed significant weaknesses that quickly dampened the euphoria. The most obvious problem lies in what is known as “routing” – the automatic selection between different model versions depending on the complexity of the query. Users report inconsistent behavior, with simple queries being forwarded to computationally intensive reasoning models, while complex problems are handled by weaker variants.

Rate limits proved particularly problematic. Users complain about “way less prompts allowed with plus users hitting limits in an hour,” which significantly limits practical usability. These restrictions effectively negate the promised performance improvement, as intensive workflows have to be interrupted.

(some examples of the controversy)

Another point of criticism concerns the quality of responses. Paradoxically, many users complain that despite its technical superiority, GPT-5 shows “less personality” and produces more generic responses. The model is criticized for “generic responses” and a deteriorated “ChatGPT user experience compared to its predecessors.” This development points to a classic trade-off between security and creativity—the increased security training may have compromised spontaneous, lively interaction.

Basic mathematical skills, an area where improvements were particularly expected, continue to show weaknesses. The model “failed on a simple algebra arithmetic problem that elementary schoolers could probably nail, 5.9 = x + 5.11.” Such errors in elementary calculations significantly undermine confidence in the advertised “PhD-level intelligence.”

Admittedly, OpenAI reacted quickly, as Sam Altman announced on X one day after the release of GPT-5. Within a very short time, the rate limits for Plus users were increased, the routing problem was addressed, and the UI was redesigned.

However, it remains to be seen whether this will satisfy the community.

A particularly controversial aspect of the GPT-5 controversy concerns the way OpenAI presented the new model to the public. During the livestream, OpenAI used “seemingly straightforward charts for benchmarks” with a fundamental error: "the labels do not remotely match the bar heights. The bar for 69.1% is the same height as the one for 30.8%.“ This discrepancy between the visual representation and the actual values raised questions about the integrity of the presentation.

The misrepresentation was so obvious that it resulted in ”an apology from the company and a public acknowledgment of error by CEO Sam Altman." This oversight—whether intentional or not—reinforced the feeling in the community that OpenAI was presenting its results in a favorable light. At a time when the AI industry is already under suspicion of making exaggerated promises, such mistakes were particularly damaging to its credibility.

The flawed data visualization is symbolic of a larger problem: the entire announcement seemed “very odd. Countless plots were mislabeled, live demos had bugs, and the early rollout is doing some weird stuff.” This observation suggests that OpenAI may have been under external pressure to release GPT-5 earlier than would have been technically optimal.

The AI community's response to GPT-5 reveals a deep divide between different user groups. While casual users often report positive experiences, power users – those who work intensively with AI models on a daily basis – are significantly more critical. “Power users have been strikingly underwhelmed with OpenAI's GPT-5 so far, raising questions about diminishing returns.”

This discrepancy is no coincidence. Power users have spent months and years exploring the limits of GPT-4, developing workarounds for its weaknesses, and establishing specific workflows. Their expectations of GPT-5 are based on concrete use cases and measurable improvements. When these fail to materialize or are counteracted by other factors – such as rate limits or changed behavior – the disappointment is correspondingly great.

An interesting aspect of the debate concerns the question of diminishing returns in AI development. One user on Hacker News remarked, “it feels like the race has never been as close as it is now,” and speculated about “the hard take-off / winner-take-all mental model.” This observation suggests that the differences between the leading AI models from different providers could be narrowing – a sign that the industry may be reaching the fundamental limits of current approaches.

In short, the entire release of GPT-5 seems to be geared toward the mass market. It is no coincidence that OpenAI has repeatedly emphasized that up to 700 million users now use ChatGPT every week. For the vast majority of “casual” users who do not care about top scores in benchmarks such as MMLU, SWE, or GPAQ-Diamond, GPT-5 remains a huge leap forward despite all the controversy: thanks to routers, many will be testing a reasoning model for the first time ever.

The controversy surrounding GPT-5 is inextricably linked to the expectation management that OpenAI engaged in in the months leading up to its release. “OpenAI's CEO, Sam Altman, overpromised on GPT-5, and real-life results are underwhelming.” This assessment reflects a fundamental problem in the AI industry: the pressure to continuously deliver groundbreaking improvements leads to inflated expectations.

Months before the launch, rumors circulated about AGI-like capabilities, revolutionary advances in reasoning, and a whole new quality of AI interaction. Social media and tech influencers amplified these expectations until a veritable hype was created. As one analyst aptly noted, “Overpromising will always lead to some sort of underdelivering, but what we're getting is still phenomenal.”

This dynamic reveals a structural problem in AI development: the speed of technical progress cannot keep pace indefinitely with the expectations of investors, the media, and users. Objectively speaking, GPT-5 is a remarkable technical advance – but measured against the exaggerated expectations, it seems disappointing.

Nevertheless, other metrics, such as those recently published by METR, show that the excitement is not entirely unfounded. GPT-5 impressively demonstrates how much the exponential curve continues to rise in terms of the time horizon in SWE tasks.

The first few weeks after the GPT-5 launch paint a mixed picture. While the technical innovations—especially the reasoning abilities and reduction of hallucinations—undoubtedly represent progress, the model is struggling with implementation issues and is not meeting all of the high expectations. The reality lies somewhere between the promise of revolutionary improvements and the realization that “what we're getting is still phenomenal,” even if it doesn't live up to all hopes.

OpenAI has already responded: rate limits will be doubled, known bugs will be fixed, and the routing system will be overhauled. These quick adjustments suggest that many of the current issues are solvable and not due to fundamental architectural weaknesses.

The GPT-5 controversy marks a pivotal moment in the evolution of artificial intelligence – not as a technical failure, but as a reality check for an industry that must navigate between exponential expectations and the laws of scientific progress. The analysis shows that GPT-5 has achieved significant technical breakthroughs: the integration of reasoning capabilities, the drastic reduction of hallucinations, and the consolidation of multimodal functions represent genuine innovations that have the potential to fundamentally change AI applications.

At the same time, implementation problems – from inadequate rate limits and faulty routing algorithms to misleading presentation charts – reveal systemic weaknesses in OpenAI's product development and communication strategy. However, these problems are largely technical in nature and can in principle be solved, as the corrections already initiated show.

The true significance of the GPT-5 controversy lies in its role as a turning point for expectation management in the AI industry. It demonstrates that even with continuous technical improvement, the era of “magical” leaps may be giving way to a phase of incremental optimization. This is neither a sign of failure nor an indication of the end of the AI revolution, but rather a natural maturation process of a technology that is transitioning from exponential breakthroughs to sustainable, application-oriented development.

This realization presents both challenges and opportunities for the future of AI development. Companies will need to learn to communicate more realistic expectations, while users and investors will need to develop a more nuanced understanding of the complexity of AI progress. GPT-5 may not be the revolution many had hoped for – but it could be the building block for a more sustainable, application-focused AI future.

The crucial question is no longer whether the next generation of AI models will fulfill all our dreams, but how we can make the most of the impressive capabilities already available. This paradigm shift from expecting magical solutions to pragmatically applying existing possibilities could be where the real breakthrough lies – not as a spectacular leap, but as the methodical evolution of a technology that already has the potential to fundamentally change the way we work and think.

Sources:

🔗 https://www.oneusefulthing.org/p/gpt-5-it-just-does-stuff

🔗 https://cdn.openai.com/pdf/8124a3ce-ab78-4f06-96-49ea29ffb52f/gpt5-system-card-aug7.pdf

🔗  https://openai.com/de-DE/index/introducing-gpt-5/

🔗 Lambert, Nathan. "GPT-5 and the arc of progress." Interconnects, August 2025. https://www.interconnects.ai/p/gpt-5-and-bending-the-arc-of-progress

🔗 Tom's Guide. "Nearly 5,000 GPT-5 users flock to Reddit in backlash — it 'feels like a downgrade' and 'I feel like I'm taking crazy pills'." August 2025. https://www.tomsguide.com/ai/chatgpt/chatgpt-5-users-are-not-impressed-heres-why-it-feels-like-a-downgrade

🔗 Futurism. "GPT-5 Users Say It Seriously Sucks." August 2025. https://futurism.com/gpt-5-sucks

🔗 The Droid Guy. "Many ChatGPT Users Disappointed With GPT-5 Update, Calling It Sterile, Incomplete, and 'A Downgrade'." August 2025. https://thedroidguy.com/many-chatgpt-users-disappointed-with-gpt-5-update-calling-it-sterile-incomplete-and-a-downgrade-1269149

🔗 TechRadar. "ChatGPT users are not happy with GPT-5 launch as thousands take to Reddit claiming the new upgrade 'is horrible'." August 2025. https://www.techradar.com/ai-platforms-assistants/chatgpt/chatgpt-users-are-not-happy-with-gpt-5-launch-as-thousands-take-to-reddit-claiming-the-new-upgrade-is-horrible

🔗 Constellation Research. "OpenAI launches GPT-5, a system of models." August 2025. https://www.constellationr.com/blog-news/insights/openai-launches-gpt-5-system-models

🔗 IEEE Spectrum. "OpenAI's GPT 5: Vibe Coding Reaches New Heights." August 2025. https://spectrum.ieee.org/openai-gpt-5-agi

🔗 Microsoft News. "Microsoft incorporates OpenAI's GPT-5 into consumer, developer and enterprise offerings." August 2025. https://news.microsoft.com/source/features/ai/openai-gpt-5/

Keeping up with AI is hard – we get it!

That’s why over 1M professionals read Superhuman AI to stay ahead.

Sign up and start mastering AI

Looking at the GPT-5 controversy from a strategic perspective, a remarkable shift in OpenAI's focus becomes apparent, one that is often overlooked in the heated debate within the AI community. Paradoxically, the disappointment of power users could be an indication of OpenAI's wisest decision in the long run: a deliberate shift toward the mass market.

While the technical elite gets worked up about rate limits and routing issues, it loses sight of the fact that it represents only a fraction of the user base. For the overwhelming majority of ChatGPT users – an estimated 700 million people worldwide every week – GPT-5 will be their first reasoning model. These users have never pushed the limits of o3, never developed complex workflows, and are unfamiliar with the frustrations of inconsistent API responses. For them, GPT-5 is simply magic: an AI assistant that “thinks” before it answers, that thinks through complex problems step by step and makes it transparent how it arrives at its conclusions.This perspective changes the entire assessment of the supposed “controversy.” What power users perceive as regression—the more “generic” answers, the more conservative orientation, the more predictable responses—are precisely the qualities that create trust for mainstream users.

OpenAI has probably realized that the path to market dominance does not lie in satisfying a small but vocal group of experts, but in tapping into the enormous potential of those who have only used AI superficially so far. Democratizing a reasoning model for the mass market could prove to be one of the smartest moves in the history of AI – even if it provokes criticism from the established community in the short term.

In this respect, GPT-5 will most likely be a resounding success despite all the current controversy. The real triumph will not be seen in the discussion forums of AI enthusiasts, but in the usage statistics, retention rates, and the quiet revolution in the everyday workflows of millions of people who will experience the power of a thinking digital assistant for the first time.