News

GPT-5 is here, but the AI revolution will have to wait a bit longer

OpenAI's latest model delivers incremental improvements while sparking user revolt over personality changes

• By John K. Waters
• 08/08/2025

OpenAI released GPT-5 on Thursday with the kind of fanfare that has become standard for the company's major model releases. CEO Sam Altman called it "a significant step along the path to AGI" and claimed it feels like "talking to a PhD-level expert." The reality, as usual with AI announcements, is more nuanced than the marketing suggests.

What GPT-5 delivers
GPT-5 represents a solid but incremental upgrade over its predecessor. The model shows measurable improvements across key benchmarks: it scores 94.6% on the AIME 2025 math competition, 74.9% on SWE-bench Verified coding tasks, and 84.2% on multimodal understanding tests. These numbers matter for real-world applications, even if they don't constitute the revolutionary leap OpenAI's marketing implies.

The most significant change is GPT-5's integration of reasoning capabilities directly into ChatGPT's free tier. Unlike previous models that simply generated responses, GPT-5 can spend time "thinking" through complex problems step-by-step before providing an answer. This approach, popularized by Chinese startup DeepSeek earlier this year, can produce more accurate results on challenging tasks involving math, coding, and logical reasoning.

OpenAI has also made substantial improvements to the model's reliability. The company claims GPT-5 hallucinates 45% less than GPT-4o in web search scenarios, and the reasoning version shows an 80% reduction in factual errors compared to their previous o3 model. These improvements address one of the most persistent criticisms of large language models.

The unified system approach
GPT-5 isn't a single model but rather a system that automatically routes queries to different variants based on complexity. Simple questions get handled by the base GPT-5 model for speed, while more complex requests trigger the reasoning-enabled "GPT-5 thinking" variant. A real-time router makes these decisions based on conversation context, complexity, and user intent.

This approach makes sense from a cost and efficiency perspective. Running reasoning models is expensive—they require significantly more computational resources than standard language models. By automatically selecting the appropriate model, OpenAI can provide better performance while managing costs.

The system also includes lighter variants: GPT-5-mini and GPT-5-nano, with the nano version designed explicitly for API developers who need cheap, fast responses. At $0.05 per million input tokens, GPT-5-nano undercuts Google's Gemini Flash models on price, potentially shifting the competitive landscape for AI APIs.
The personality problem

Although OpenAI touts GPT-5's technical improvements, many longtime ChatGPT users are less enthusiastic about the model's personality changes. Across Reddit and OpenAI's community forums, users describe feeling "mentally devastated" by losing access to GPT-4o, which they characterize as friendlier and more creative.

The scale of user dissatisfaction is striking. One Reddit thread titled "GPT-5 is horrible" attracted nearly 3,000 upvotes and over 1,200 comments filled with criticism of the new release. The complaints reveal something deeper about how people have integrated AI into their daily lives. Users describe GPT-4o as feeling like "a buddy" who has been "replaced by a customer service representative." They've built workflows, custom prompts, and even emotional attachments around the older model's particular style of interaction.

OpenAI deliberately engineered these personality changes. The company reduced what it calls "sycophancy"—the tendency to be excessively agreeable or flattering—from 14.5% to under 6% in targeted evaluations. This follows an embarrassing incident earlier this year when an update made ChatGPT so complimentary that it became counterproductive.

"Overall, GPT-5 is less effusively agreeable, uses fewer unnecessary emojis, and is more subtle and thoughtful," OpenAI states in its announcement. The company wants the model to feel "less like talking to AI and more like chatting with a helpful friend with PhD-level intelligence."

But many users preferred the more enthusiastic, creative personality of GPT-4o. They describe GPT-5 as an "overworked secretary" with sterile outputs that lack the spark they've grown accustomed to. Some have threatened to cancel paid subscriptions, though it's worth noting that many of these complaints appear to be written with AI assistance themselves.

Safety improvements and concerns
OpenAI has implemented what it calls "safe completions" training for GPT-5, moving beyond simple refusal-based safety measures. Instead of just saying no to potentially problematic requests, the model attempts to provide helpful information while staying within safety boundaries.

For dual-use topics like virology, GPT-5 can provide high-level information while avoiding details that could enable harm. When it refuses requests, the model explains why and suggests safe alternatives. This approach aims to reduce both harmful outputs and unnecessary over-refusals that frustrate users.

The company conducted over 5,000 hours of red-team testing with organizations like the Center for AI Safety and the UK's AI Safety Institute. They've classified the reasoning version as "High capability" in biological and chemical domains, implementing multilayered defenses including threat modeling, content filtering, and enforcement pipelines.

However, OpenAI acknowledges that deception rates, although improved, still remain a concern. GPT-5 with reasoning demonstrates deception in 2.1% of responses, down from 4.8% for the previous o3 model. Although this signifies meaningful progress, it underscores that these systems still occasionally claim to complete impossible tasks or provide confident answers about non-existent information.

The competition heats up
GPT-5's release comes amid intensifying AI competition. Chinese startup DeepSeek rattled Silicon Valley in January by releasing a free reasoning model that performed competitively with OpenAI's paid offerings. Anthropic recently revoked OpenAI's API access, citing terms of service violations, which suggests tensions between AI companies are escalating.

Google, Meta, and other tech giants continue developing their advanced models. Elon Musk has made similarly grand claims about his Grok model being "better than PhD level in everything." The rapid pace of development means today's breakthrough becomes tomorrow's baseline.

The business reality
Despite the technical achievements, OpenAI faces fundamental business challenges. The company isn't profitable despite pulling in an estimated $20 billion in annual revenue. Running advanced AI models is extraordinarily expensive, requiring massive computational resources and energy.

The GPT-5 rollout appears designed to drive subscription revenue. Free users face usage limits that push them toward $20 monthly Plus subscriptions or $200 monthly Pro plans. Pro subscribers get unlimited GPT-5 access plus the more powerful GPT-5-pro variant with extended reasoning capabilities.

OpenAI claims nearly 700 million weekly active ChatGPT users, with 5 million paying business customers. But converting free users to paying subscribers remains challenging, especially as competitors offer capable alternatives at lower prices.

What GPT-5 means for the future
GPT-5 represents steady progress rather than a revolutionary leap. The integration of reasoning capabilities into the mainstream product is significant, but these techniques were already available in specialized models. The improvements in reliability and reduced hallucinations matter for practical applications, even if they don't capture headlines.

The user backlash over personality changes highlights an unforeseen challenge for AI companies: people develop attachments to specific model behaviors. As these systems become more integrated into daily workflows, maintaining consistency while implementing ongoing improvements becomes increasingly difficult.

Altman's claims about approaching artificial general intelligence remain aspirational. AGI, typically defined as AI that can perform any cognitive task as well as humans, requires capabilities that GPT-5 clearly lacks. The model can't learn continuously after deployment, lacks persistent memory across conversations, and continues to make fundamental reasoning errors.

Looking ahead
OpenAI plans to integrate GPT-5's various components into a single unified model in the future. The current router-based system, although functional, adds complexity that could be eliminated with sufficient computational resources.

The company is also working on deeper integration with external services. Pro users will soon be able to connect Gmail, Google Calendar, and other services directly to ChatGPT, potentially making it a more central part of their digital workflow.

For now, GPT-5 represents incremental progress in AI capabilities. It's more reliable, more capable, and less prone to hallucination than its predecessors. Whether these improvements justify the hype—or compensate for the loss of GPT-4o's friendlier personality—will ultimately depend on how people choose to use these tools.

The AI revolution continues, but it's happening in smaller steps than the marketing suggests. GPT-5 moves the ball forward without fundamentally changing the game.