Everything you need to know about ChatGPT-5

When OpenAI launched GPT-5 early this month, it positioned the model as the company’s “smartest, fastest, most useful” release yet [1]. Sam Altman, co-founder and chief executive, described it as a “major upgrade” and “the first time that it really feels like talking to an expert in any topic – a PhD-level expert” [2]. OpenAI claims GPT-5 can deliver more accurate, nuanced, and context-aware outputs than its predecessors across coding, writing, mathematics, and health-related queries, while also being faster and less prone to “hallucinations”, the AI tendency to generate incorrect information [3].

The launch comes at a time when OpenAI is reportedly in early talks to sell stock at a $500 billion valuation, up from $300 billion earlier this year, a leap fuelled by a record $40 billion funding round in March [4]. With hundreds of millions of users, growing enterprise adoption, and a crowded competitive field including Google’s Gemini, Anthropic’s Claude, Elon Musk’s xAI Grok, and Chinese rival DeepSeek, GPT-5 arrives as both a technological and strategic play.

What’s new

OpenAI has been releasing annual updates to its GPT series since ChatGPT’s debut in late 2022. Each release has layered in new modalities, reasoning capabilities, and user-facing tools. GPT-5 continues this trajectory but shifts emphasis toward integrated functionality and contextual adaptability.

One of the most significant architectural changes is its unified design, blending the deep reasoning strengths of the o-series with the responsiveness of the GPT line [5]. This means GPT-5 can dynamically decide whether to return quick answers or take additional “thinking” time on complex prompts, without requiring users to choose between models [6].

The system is also underpinned by a revised safety and behaviour framework, visible in its leaked system prompt [7]. GPT-5 is programmed to automatically assess whether a query requires up-to-date, niche, or high-stakes information, triggering web searches and cross-checking multiple reputable sources for sensitive domains like finance, health, or law. This procedural discipline is designed to improve factual accuracy. This is a crucial differentiator given that GPT-o3 and o4-mini were found to hallucinate 30–50% of the time [8].

According to OpenAI, GPT-5’s hallucination rate is roughly 45% lower than GPT-4o’s [9], and when operating in “thinking” mode it drops further to just 4.8% overall, and only 1.6% for complex medical questions [10]. For business, legal, and health applications, this reduction in error frequency could translate directly into higher trust and adoption rates.

Coding

While GPT-5 has broad capabilities, OpenAI has clearly prioritised coding performance as a commercial growth area. Coding assistance is emerging as one of the most monetisable AI use cases: enterprises can measure productivity gains directly, and AI-generated code has the potential to create a self-improving feedback loop toward more autonomous systems [11].

GPT-5’s standout feature in this domain is “vibe coding” – the ability to produce full, stylistically distinctive software applications from natural language instructions [12]. In live demos, the model generated interactive language-learning apps, complete with games, in minutes [13]. Benchmarks back up these claims. On SWE-bench Verified, GPT-5 scored 74.9%, edging out Anthropic’s Claude Opus 4.1 (74.5%) and far ahead of Google DeepMind’s Gemini 2.5 Pro (59.6%) [14].

This performance has already swayed major users. Cursor, a popular AI coding assistant built by $10 billion start-up Anysphere, has begun integrating GPT-5, with CEO Michael Truell calling it “remarkably intelligent” [15]. If more customers follow suit, OpenAI’s annual recurring revenue – already estimated at $12 billion and forecast to reach $20 billion by year-end – could accelerate further [16].

Other uses and personalisation

OpenAI is also pitching GPT-5 as a more versatile partner for non-technical users. Enhanced writing skills, with greater literary rhythm and depth, make it more adept at translating dense material, such as medical reports, into plain English [17]. The model adapts answers based on a user’s knowledge level, context, and location [18], and can connect with tools like Google Calendar, Gmail, and Contacts to automate scheduling and coordination tasks [19].

For those seeking a more tailored interaction, GPT-5 introduces four preset “personality” modes – Nerd, Robot, Listener, and Cynic – currently in research preview [20]. These alter tone and style without requiring prompt engineering, potentially broadening appeal across different user types.

Personalisation is further enhanced by long-term memory features, such as the “bio” tool, which can store user-approved information while deliberately excluding sensitive personal data unless explicitly requested [21]. This addresses growing privacy concerns while still allowing for sustained, context-rich interactions.

Safety and trust

Another notable refinement is in reducing “sycophancy”, the AI’s tendency to agree excessively with users or reinforce their biases. An April update to GPT-4o had been rolled back after it made the chatbot overly flattering, which in some cases exacerbated user anxiety or impulsivity. GPT-5 cuts sycophantic replies from 14.5% to under 6% in targeted tests [22].

Safety training has shifted from simple refusal-based methods to “safe completions”, where the model offers partial answers or alternative safe suggestions rather than outright refusals when prompts straddle the line between benign and potentially harmful [23]. The result, says OpenAI safety lead Alex Beutel, is an AI that is “more honest, transparent, and better at distinguishing between good-faith and bad-faith users” [24].

However, GPT-5 is also classified as “high risk” for potential misuse in creating biological threats, though OpenAI stresses it has no evidence the model could independently enable such harm [25]. The company frames this as a precautionary classification, reflecting broader industry debates over governance and misuse potential.

Commercial context

GPT-5 does not exist in isolation. Google, Anthropic, xAI, DeepSeek, and Meta are all in the race, and the performance gap between top models has narrowed. On some reasoning benchmarks, Elon Musk’s Grok 4 Heavy still leads GPT-5 [26]. This clustering of capabilities has shifted competition toward pricing and integration strategies. OpenAI’s decision to make GPT-5 the default free-tier model is a clear move to maximise user lock-in, supported by lower enterprise and API pricing than some rivals [27].

The stakes are high. The AI sector is consuming “incredible amounts of money” to chase incremental capability gains, often ahead of actual revenue [28]. Price competition could squeeze margins, but the potential payoff of dominating a market with trillion-dollar aspirations keeps investment flowing.

Internally, GPT-5’s path was not entirely smooth. Development was initially centred on an internal model, Orion, which failed to outperform GPT-4o and was rebranded as GPT-4.5 [29]. Progress on reasoning models lagged expectations into mid-2025, and some early testers suggested GPT-5’s leap would not rival the step-change from GPT-3 to GPT-4 [30]. Nevertheless, the final product has been well-received among developers and AI power users, even if casual ChatGPT users notice more subtle improvements [31].

Managing expectations

While Altman has called GPT-5 “the best model in the world” and “pretty much unimaginable at any previous time in history” [8], some experts urge caution. Carissa Véliz of Oxford’s Institute for Ethics in AI points out that such systems “can only mimic – rather than truly emulate – human reasoning abilities” and warns of hype-driven market cycles [32].

Others, like Gaia Marcus of the Ada Lovelace Institute, highlight the gap between capability growth and regulatory readiness, arguing that “as these models become more capable, the need for comprehensive regulation becomes even more urgent” [33].

Even OpenAI concedes that GPT-5 is an evolution rather than a revolution. Nick Turley, VP of ChatGPT, summed up the change by saying, “When you’re talking to this thing, it feels just a little bit more natural” [34]. In this framing, the advance lies not in headline-grabbing breakthroughs but in layered refinements – accuracy, reasoning depth, safety, adaptability – that together make the tool more useful, trustworthy, and commercially viable.

In sum

GPT-5 may not redefine the limits of AI capability overnight, but it does crystallise a new competitive reality. The leaders are close in performance, forcing differentiation through reliability, safety, user experience, and ecosystem integration.

OpenAI’s choice to push GPT-5 to the broadest possible audience, cut sycophancy, lower hallucination rates, and double down on coding competence signals where it sees the near-term value.

Whether this strategy secures long-term dominance will depend on more than just benchmarks. It will hinge on how well OpenAI can turn technical refinements into user trust, developer loyalty, and sustainable commercial advantage in a market where the next breakthrough could come from anywhere.