OpenAI's GPT-5 arrives with four tiers, reduced deception, and deprecates old models

OpenAI has officially launched its long-awaited flagship model, GPT-5, in four distinct flavors tailored to different use cases: GPT-5 for logic and multi-step tasks, GPT-5-mini for cost-sensitive applications, GPT-5-nano for low-latency needs, and GPT-5-chat for advanced enterprise conversations. CEO Sam Altman debuted the model during a webcast, highlighting major gains in code generation, math, physics, and healthcare. The model can dive deeper into problems without user prompts and is available immediately to free, Plus, Pro, and Team users, with Enterprise and Edu access coming August 14.

Safety and trust are central to this release. OpenAI researchers emphasized that GPT-5 is "significantly less deceptive" than prior models, achieved through reduced hallucinations and improved factuality. A new technique called "safe completions" allows the model to decline potentially harmful responses while remaining helpful. This move addresses growing concerns about AI misuse and aligns with broader industry efforts to make models safer.

In a notable break from the past, OpenAI is deprecating all previous GPT models. Taya Christianson, a spokesperson, confirmed that after 60 days, GPT-5 will become the default for all tiers, eliminating the need for users to choose which model to use. The release also introduces customizable chat interface colors and a research preview of chat personalities. This launch comes as the AI industry faces diminishing returns from scaling larger models, signaling a shift from raw performance benchmarks toward more nuanced, safety-focused improvements.