GPT-5: The next generation of AI for more accurate, secure and natural interaction

On 7 August 2025, OpenAI introduced GPT-5, a unified system that delivers intelligent responses by automatically switching between fast and deep thinking modes, promising expert-level performance. It is available via ChatGPT (free and subscription-based) and via the API.

Core technologies and variants

• Unified system with model routing
  GPT‑5 includes an automatic router that switches between gpt‑5-main, gpt‑5-thinking and mini/nano variants based on complexity, tool usage or user intent such as ‘think hard’ – for an optimal balance of efficiency and depth.

• API variants for developers
  Three versions are available via the API: gpt‑5, gpt‑5‑mini and gpt‑5‑nano, tailored to performance, cost and latency requirements.

Improved technical expertise

GPT-5 significantly outperforms previous models in areas such as programming, mathematics, writing, health questions and visual perception.

• Improved technical expertise – more reliable, helpful answers
  Thanks to new ‘safe completions’ training, GPT-5 can also provide helpful but safe answers in sensitive dual-use areas. It distinguishes between harmless, publicly available information and potentially misusable details. When the intent is unclear, the model provides abstract answers, explains rejections transparently, and suggests safe alternatives.
• Reduced hallucinations & improved instruction following
  OpenAI reports fewer false statements, better compliance with user specifications and so-called ‘safe completions’, which handle dangerous queries safely and thoughtfully.
• Personalisation & new UI features
  Users can choose personalisation options such as tone, chat styles (e.g. ‘Cynic’, “Nerd”, ‘Listener’) and UI colours; integration with tools such as Gmail or Google Calendar is possible.
• Extended context scope
  GPT-5 processes significantly more context – reports speak of windows between 256,000 tokens and 1 million tokens.
• Reduce flattery and refine style
  GPT-5 is less effusive and approving, largely dispenses with unnecessary emojis and formulates more subtly and thoughtfully than GPT-4o. The aim is to make the conversation seem more natural – more like with a helpful, academically educated person than with a machine.

Does GPT-5 really perform better?

More accurate answers to complex questions are among the greatest advances made by GPT-5. Compared to previous generations, there are significantly fewer hallucinations. Analyses based on realistic inputs show that GPT-5 makes around 20% fewer factual errors than GPT-4o and – in extended thinking mode – even around 70% fewer errors than the OpenAI o3 model.

To achieve this accuracy, OpenAI has invested specifically in improving reasoning for open-ended, complex questions and established new testing procedures. Factual accuracy was evaluated using public benchmarks such as LongFact (concepts and objects) and FActScore (fact-based statements). GPT-5 Thinking in particular shows a dramatic improvement here: the hallucination rate is reduced by a factor of six compared to o3. The result is sustainably higher reliability in the creation of longer, fact-oriented texts – a decisive step towards consistent and accurate AI-supported content.

Availability and access

• Free access & subscription benefits
  GPT‑5 is available to all ChatGPT users. Free accounts are limited to the mini version. Pro subscriptions offer unlimited use of GPT-5 and access to GPT‑5 Pro.
• Enterprise use & API
  Developers can access all variants via the API. GPT-5 is already being used in enterprise solutions, Copilot, Apple Intelligence and other tools.

User reactions and OpenAI response

Despite the technical advances, GPT-5 also met with criticism from users. Many complained that the new standard seemed less emotional than its predecessor, GPT-4o, and in some cases provided inaccurate or impersonal responses. In particular, the automatic model routing, which switches between different GPT-5 variants depending on the query, caused dissatisfaction as it occasionally failed to deliver the expected model quality. OpenAI responded quickly to the feedback: earlier models such as GPT-4o remain available, and the company announced improvements to the model switching logic. It also promised an increase in usage limits and a targeted optimisation of the response style towards a warmer, more personal tone. Overall, GPT-5 is a significant step towards intelligent, adaptive AI assistance.