What are the new features of GPT-5?
OpenAI describes GPT-5 as “the smartest and most useful model yet—with built-in reasoning that makes expert knowledge accessible to everyone” (OpenAI, 2025). According to the official GPT-5 system card, OpenAI describes GPT-5 as “a unified system with a fast model, a deeper reasoning model (GPT-5 Thinking), and a real-time router that quickly decides which model to use based on the type of conversation and complexity.” The “router” is an administrative layer or decision-making logic, not necessarily a standalone “hardware router” in the traditional sense. The model to be used is decided on a per-prompt basis. The model promises shorter response times and more stable performance, while also offering longer dialogue times and personalized responses. Here is an overview of the most important new features, based on OpenAI's release blog:
Improved accuracy and speed: more precise answers and faster responses than its predecessors, especially in the fields of mathematics, science, finance, and law.
Optimized programming capabilities: introduction of GPT-5 Codex as a specialized model for demanding programming and debugging tasks, a model that has been specifically trained for coding agent tasks; it is used in development environments and is also available to companies via the API. In general, the standard GPT-5 model has significantly improved programming capabilities compared to GPT-4o
Advanced creative writing: GPT-5 transforms raw ideas into compelling, impressive texts with literary depth and rhythm
Improved health literacy: HealthBench scores significantly higher than all previous models, based on realistic scenarios and criteria defined by doctors.
Advanced context processing: The extended token context window allows GPT-5 to engage in longer and more complex conversations.
New model variants: Specialized variants such as GPT-5 mini and GPT-5 nano are optimized for specific tasks
Enhanced security features: GPT-5 Codex includes both model-level mitigations (e.g., training against malicious tasks and prompt injections) and product-level protections such as sandboxing and configurable network access.
This makes GPT-5 more versatile for use in a wide variety of areas, from content creation and programming to business analysis. The following graphic demonstrates the system architecture and underlying models for GPT-5:
While companies and developers welcome this potential, critics are increasingly voicing concerns about data protection and social implications.
Versions and Usage of the GPT Models
Here an overview of the current models of OpenAI's GPT used:
Versions and use of the GPT models
GPT-5 represents the ongoing evolution of OpenAI’s AI models. GPT-4 laid the groundwork with advanced text and image processing and reliable accuracy in everyday tasks. GPT-4o, as a turbo-optimised variant, operated faster and more accurately, and also experimented with audio and video processing as well as empathetic and emotional tone. As the next generation, GPT-5 combines the highest accuracy with native multimodality for text, images, audio and video, an extremely large context window, and an integrated thinking mode. The choice of model depends on the required accuracy and capacity, depending on the use case.
The latest Frontier model is GPT-5.4. GPT-4o and other models were withdrawn from ChatGPT in February 2026. Within ChatGPT, GPT-5.3 Instant, GPT-5.4 Thinking and GPT-5.4 Pro are available. OpenAI is thus consistently focusing on expanding the new generation and further developing the GPT-5 model series.
Also new: ChatGPT Go introduces a new, affordable subscription tier between Free and Plus, significantly broadening access – strategically timed just ahead of the planned advertising rollouts. At the same time, OpenAI has officially confirmed for the first time the introduction of advertising in ChatGPT, specifically via ‘Sponsored Recommendations’ for Free and Go users.
GPT-5 vs. GPT-4/4o
Whilst GPT-4 already represented a major leap forward in AI-powered text and image processing, GPT-5 is now setting new standards in precisely this area of accuracy and performance. GPT-4 offers developers and businesses a powerful platform for applications in text processing, chatbots, content generation and much more. GPT-5 significantly expands these capabilities: whilst GPT-4 Turbo and GPT-4o in particular already supported multimodality, GPT-5 allows for audio and video content in addition to text and images, offers a significantly larger context window and has introduced the new thinking mode. Furthermore, the standard GPT-5 model has significantly improved programming capabilities compared to GPT-4o.
An overview of GPT-5 models
The GPT-5 model series comprises several distinct variants that differ in terms of computational complexity, depth of reasoning, response quality and primary use case.
GPT-5: The first version of this generation, it introduced the built-in ‘Thinking’ paradigm as well as an adaptive, routed architectural system. GPT-5 (Instant and Thinking) was withdrawn from ChatGPT on 13 February 2026; access via the API remains available for the time being.
GPT-5.1: An upgrade of GPT-5 featuring improved personality settings and instruction handling; the 5.1 models were removed from ChatGPT in March 2026, but remain available via the OpenAI API.
GPT-5.2: This version introduced significant improvements in structured reasoning and professional knowledge work. However, GPT-5.2 is scheduled to be decommissioned on 5 June 2026.
GPT-5.3: An evolution of GPT-5.2, focusing on stability and the reduction of hallucinations. Key improvements include
- More precise tool integration and agent control
- More robust context processing in longer sessions
- More consistent responses for multi-step reasoning tasks
- GPT-5.3 Instant as a fast and powerful standard model for everyday tasks.
Codex variants: GPT-5 Codex (version 5.3) is a specialised variant of GPT-5 designed for agent-based coding and optimised for complex programming and cybersecurity via Codex platforms, as well as Codex CLI/IDE integration.
GPT-5.4 (currently the latest model): The current Frontier model marks a structural evolution of the entire product line. It is the first mainstream model to directly integrate the Frontier coding capabilities of GPT-5.3 Codex. It features a context window of over one million tokens (922,000 input, 128,000 output) with support for text and image inputs, and is available in three variants:
- GPT-5.4 Standard: The default option for general and professional tasks.
- GPT-5.4 Thinking: A reasoning variant in ChatGPT with adjustable reasoning depth (Standard, Extended, Light, Heavy, depending on the plan).
- GPT-5.4 Pro: High-performance variant for the most complex enterprise use cases, optimised for step-by-step reasoning, instruction following and maximum accuracy.
Current debates surrounding GPT-5: Why is the new model so divisive?
The issues of security and transparency are particularly divisive, as are trust and sustainability in the context of GPT models. The disappointed reactions to the GPT-5 models appear to stem from overly high expectations and hype, which are being better managed by OpenAI’s competitors.
High expectations and criticism of mistakes
Users on platforms such as X and Reddit reported poorer performance compared to GPT-4o. Expectations of the new models proved to be too high, and the question “What has happened to GPT?” is frequently being asked. Some users criticised the new model as cold and distant, particularly in comparison to GPT-4o, which was especially known for its emotionally rich conversations. This reignited the debate about user engagement and the emotional aspects of AI models. Furthermore, Sam Altman, the CEO of OpenAI, admitted to blunders during the rollout of GPT-5: the bar charts showing model performance displayed in the launch livestream contained errors, which Altman attributed to human error caused by exhaustion. (Reddit, 2025)
Advertising and user trust in 2026
As recently as 2024, Altman described advertising as a “last resort” for ChatGPT; however, an official advertising trial has been running in ChatGPT in the US since 9 February 2026 and has been gradually expanded since then. Adverts are always clearly labelled as “Sponsored” and visually separated below the actual ChatGPT response. The so-called “Sponsored Recommendations” are displayed in line with the ongoing conversation. Advertising is not yet active in Germany and the DACH region, although the ChatGPT Go subscription page in German already displays the note “This plan may contain advertising”. Although the move sparked significant online criticism regarding user trust and data protection, OpenAI has so far reported no negative impact on trust metrics. (OpenAI, 2026)
Another point of criticism concerns the source and reliability of the data used by the models. Versions such as GPT-5.2 have repeatedly been found to cite an AI-generated online encyclopaedia that lacks human editorial oversight. The risks of hallucinations and misinformation have not been completely eliminated, and OpenAI must address this challenge strategically and make continuous improvements.
Rapid model development and fierce competition
The rapid succession of releases, from GPT-5.1 and 5.2 to 5.3 and 5.4, within just a few months is causing increasing confusion among users and developers. The discussions that have emerged have centred primarily on the following questions:
- Will AI models become more personalised in the future?
- Are smaller, specialised variants more useful?
- Are new generations of models being released and replaced too soon?
- What about the quality and ethics of the new models?
OpenAI’s market leadership is under pressure: a lack of transparency, distortion of competition and changes such as the introduction of advertising are leading users to increasingly turn to alternatives on the market, such as Google Gemini or Anthropic’s Claude. Anthropic is deliberately positioning its chatbot, Claude, as an ad-free alternative.
Assessment and Outlook
Despite impressive technical advances, the GPT-5 cycle shows that OpenAI is increasingly facing growing competitive pressure from major rivals such as Google and declining trust among its own user base. Nevertheless, GPT-5 has now established itself as an integral part of numerous enterprise and developer platforms, including Microsoft Copilot and Azure AI Foundry. This widespread integration optimises the use of GPT-5 in business processes and customer service applications. The introduction of a hybrid multi-model system enables more efficient task distribution and leads to improved performance when handling complex queries. Benchmark tests confirm higher accuracy and reduced error rates compared to previous models.
"GPT-5's strengths lie primarily in agentic applications, especially in the area of in-depth research and as an assistant in programming."
Patrick from moinAI, CEO & Head of Research
What about gpt-oss?
With less buzz in the general news world, we at moinAI are very excited about the open source model gpt-oss. This was introduced in the same week as the GPT-5 models. The advantage of the oss variant for moinAI as a SaaS provider is that it is very fast and performs similarly to the popular GPT-4o model. This makes it possible to offer the service at its current excellent level, hosted entirely on German servers at moinAI, without dependence on providers such as OpenAI or Azure. This ensures that customer data is never passed on to third parties.
What can we expect from future models?
Future models such as GPT-6 could therefore have better personalisation features to respond to individual user preferences, as well as further increasing the context window and offering even more accurate multimodality. However, exact values are not yet available, and the official announcement of GPT-6 and its specific features remains to be seen.
GPT-5 already marks a significant advance in AI development, but there are complaints about lower transparency and control compared to previous models, and responses are perceived as too formal and less appealing. OpenAI needs to address this feedback and incorporate it into future releases.
[[CTA headline="Start your company's AI journey with moinAI!" subline="Unlock the full potential of modern language models and AI technologies in your business." button="Try it now!"]]
