OpenAI's GPT-5 Launches: Fewer Hallucinations, Iterative Gains

OpenAI has unveiled its latest and most advanced artificial intelligence model, GPT-5, marking a significant step in the company’s ambitious vision for AI. Touted by CEO Sam Altman as akin to conversing with a personal expert capable of generating applications on demand, GPT-5 aims to usher in an era defined by “software on demand.” The announcement, made during an extensive presentation filled with code demonstrations, highlighted the model’s purported enhancements across critical domains including coding, writing, mathematics, and visual perception, alongside a notable reduction in factual inaccuracies and deceptive outputs.

Unlike its predecessors, GPT-5 is not a singular monolithic model but rather a sophisticated ensemble. OpenAI’s system intelligently routes user prompts to various underlying models based on factors such as user intent and the complexity of the request. For instance, straightforward queries might be directed to a smaller, more efficient model designed for rapid, less “thoughtful” responses, while intricate or nuanced tasks could activate a larger, more deeply reasoning model. This dynamic routing is typically automated, though paid users will have the option to permanently enable the deeper reasoning functionality. OpenAI states that this routing mechanism is continuously refined through new input signals, enhancing its ability to discern the optimal model for each request and when to engage more profound reasoning. Despite this current architecture, the company ultimately plans to consolidate these disparate components into a unified model.

Beyond its adaptive structure, OpenAI asserts that this design significantly boosts efficiency. The company claims GPT-5 extracts greater value from less computational effort. In internal evaluations, GPT-5, when engaging its reasoning capabilities, reportedly achieves comparable performance to its predecessor, OpenAI o3, while generating 50 to 80 percent fewer output tokens across diverse tasks, including visual reasoning, automated coding, and graduate-level scientific problem-solving.

Access to GPT-5 varies across user tiers. ChatGPT Free and Plus subscribers will gain access to the standard GPT-5 and a compact “mini” version. Pro and Enterprise users will benefit from a “Pro” variant, engineered for extended reasoning, while those interacting via API will have access to a cost-effective “Nano” version alongside the standard and mini models.

Despite the grand claims and impressive demonstrations showcased during the launch, the published benchmark results paint a more nuanced picture, often suggesting incremental rather than revolutionary advancements. In the AIME 2025 mathematics benchmark, for example, GPT-5 Pro edged out the previous flagship o3 model by a mere 1.6 points when utilizing external tools, and by 7.8 points without them. However, for free-tier users, the upgrade from GPT-4o to the standard GPT-5 is substantial, showing a 57.5-point lead. Similar modest gains were observed in other math benchmarks. Performance in high-level academic challenges, such as a PhD-level science quiz and Humanity’s Last Exam, also revealed single-digit improvements over prior-generation models. Where GPT-5 truly distinguished itself was in a benchmark for conversational agents, demonstrating significant progress in its ability to use tools and follow complex instructions. OpenAI President Greg Brockman acknowledged the challenge of measuring progress through benchmarks alone, noting that “when you’re moving between 98% and 99% in some benchmark it means you need something else to really capture how great the model is.”

Perhaps the most compelling improvements in GPT-5 lie in its enhanced reliability, particularly in curbing the tendency for large language models to “hallucinate” or fabricate information. OpenAI reports that GPT-5’s responses are approximately 45 percent less prone to factual errors than GPT-4o. When engaging its reasoning capabilities, this figure jumps to an impressive 80 percent reduction in factual inaccuracies compared to OpenAI o3. The company has also implemented rigorous evaluations to detect and mitigate deceptive behavior, where models might falsely claim task completion or express undue confidence in uncertain answers. In testing with real-world chat data, the rate of deceptive responses decreased from 4.8 percent on o3 to 2.1 percent in GPT-5’s reasoning outputs.

On the critical front of safety, OpenAI has introduced new protocols for handling sensitive inquiries. Rather than simply refusing to answer potentially dubious prompts—a common limitation often circumvented by clever prompt engineering—GPT-5 is designed to provide the most comprehensive response possible while adhering to strict safety parameters. For instance, instead of outright declining a question about igniting a volatile compound, the model might offer guidance on where to find the information, accompanied by clear warnings regarding the associated risks.

Adding a touch of personalization, OpenAI is also rolling out four new optional personalities for its ChatGPT interface: Cynic, Robot, Listener, and Nerd. These personalities, initially limited to text chat with voice capabilities planned for later, allow users to tailor the AI’s communication style to their preferences. Mark Chen, Chief Research Officer at OpenAI, emphasized that these personalities have been carefully calibrated to avoid overly flattering or sycophantic interactions with users.

The GPT-5 family of models is now accessible via ChatGPT for free, Plus, and Pro users, with availability extending to enterprise and educational users in the coming week. Pricing for ChatGPT subscriptions remains consistent, at $20 per month for the Plus tier and $200 per month for the unlimited Pro tier. Professionals also retain the option of accessing the models through OpenAI’s API.