OpenAI sets new standards in intelligence, reliability and versatility
OpenAI has unveiled GPT-5, its most powerful AI model to date, combining numerous improvements in intelligence, efficiency and safety. Users benefit from a significantly enhanced model quality, new features and broader application possibilities.
A system that adapts flexibly
GPT-5 brings together various approaches in a modular architecture: an efficient base model answers most queries swiftly, while a deeper reasoning model (“GPT-5 Thinking”) is deployed for more complex tasks. An intelligent router decides in real time which model type is best suited—depending on the nature and complexity of the conversation, as well as explicit user cues such as “think hard about this”. This router is continuously trained using real usage data and user preferences. If a usage limit is reached, a compact mini version of the models takes over further tasks. In future, these capabilities are set to merge into a single model.
Practical improvements for everyday life
Compared to previous versions, GPT-5 not only delivers faster and more precise responses, but is also far more useful for real-world, everyday applications. Reduced hallucinations, better adherence to instructions and less excessive agreement ensure greater reliability. Performance has been specifically enhanced in writing, coding and health.
- Writing & creativity: GPT-5 offers even greater support for users when drafting and refining texts. The model can handle complex literary forms such as free verse or demanding stylistic requirements, and helps with everyday tasks like writing reports, emails or speeches.
- Programming: The new model excels particularly in front-end development and debugging large repositories. With just a single prompt, it can generate attractive websites, apps or games, placing emphasis on design, typography and user experience.
- Health: GPT-5 achieves new top scores on assessment platforms like HealthBench and acts as an active advisor: potential risks are proactively addressed, and answers are better tailored to users’ context, knowledge and region. While the model does not replace a medical professional, it helps users better understand medical results and ask informed questions.
Comparative examples: GPT-4o vs. GPT-5
A clear comparison: while GPT-4o still tends to follow more traditional structures in poetry, GPT-5 impresses with vivid imagery, emotional depth and cultural context. The new model is thus able to solve tasks more creatively and with greater nuance—from poetry analysis to planning complex projects.
Outstanding benchmark results
GPT-5 sets new standards across a range of disciplines. The model achieves around 94.6% on the AIME-2025 mathematics benchmark (without aids), 74.9% on SWE-bench Verified (coding), 84.2% on multimodal benchmarks (MMMU) and 46.2% on HealthBench Hard. In the Pro version, GPT-5 scores even higher, for example 88.4% on GPQA, a particularly demanding science test. These advances are reflected in everyday use—from maths and coding to visual understanding and health queries.
Efficiency and reliability set new benchmarks
GPT-5 delivers more performance with less “thinking effort”: in tests, up to 80% fewer output tokens were needed for comparable tasks. The number of hallucinations has also been massively reduced. With web search enabled, the error rate was around 45% lower than GPT-4o, and GPT-5’s reasoning model was about 80% less error-prone than OpenAI o3. For open, fact-based tasks, “GPT-5 Thinking” produced around six times fewer hallucinations than older models.
Greater honesty and less deception
A key advance: GPT-5 communicates its own limitations and impossibilities more clearly. For tasks that cannot be solved or where important information is missing, the model openly admits this instead of guessing or pretending. In real-world tests, the deception rate dropped from 4.8% with OpenAI o3 to just 2.1% with GPT-5.
More safety with “Safe Completions”
Rather than simply refusing, GPT-5’s new safety architecture enables it to respond to sensitive or ambiguous queries as helpfully and thoughtfully as possible—without crossing safety boundaries. Transparent explanations and safe alternatives contribute to greater robustness and user-friendliness, especially in sensitive areas such as virology or chemistry. Extensive red-teaming tests and multi-layered protection mechanisms provide additional safeguards.
Less excessive agreement, more natural interaction
Compared to older models, GPT-5 is less ingratiating and overly agreeable. So-called sycophancy, or excessive compliance, has been significantly reduced through targeted training data and new evaluations—from 14.5% to under 6%. The result: communication feels less artificial and instead more natural, competent and helpful.
Personalisation: New personalities for ChatGPT
Thanks to improved controllability, users can now choose between four predefined personalities for ChatGPT: Cynic, Robot, Listener and Nerd. These can be easily adjusted in the settings and allow for a more individual interaction—ranging from factual to supportive to humourously sarcastic. All new personalities also meet high standards regarding sycophancy.
GPT-5 Pro: Even more power for demanding tasks
For particularly complex or large-scale tasks, GPT-5 Pro offers an even more powerful version. It uses additional computing resources to deliver even deeper analyses and responses. In external tests, experts preferred GPT-5 Pro in almost 68% of cases compared to the standard version, especially for challenging tasks in science, mathematics, medicine and coding.
How to access GPT-5
GPT-5 is now the default model in ChatGPT for all logged-in users, replacing GPT-4o, OpenAI o3 and other previous versions. The reasoning model is selected automatically depending on the task, but can also be specifically prompted, for example with “think hard about this”. Plus and Pro subscribers receive higher usage quotas and access to GPT-5 Pro, while Team, Enterprise and Edu customers will see the rollout within a week. Free users can also use GPT-5, but after reaching their usage limit, they are automatically switched to the slimmer mini version.
All details, further technical information and examples can be found here.