Article illustration 1

OpenAI has officially rolled out GPT-5, its most advanced language model to date, to all ChatGPT users, marking what CEO Sam Altman calls "a significant step along the path to AGI." During a high-stakes press briefing, Altman emphasized that while GPT-5 isn't yet artificial general intelligence—defined in OpenAI's charter as a system outperforming humans in most economically valuable work—it represents a leap in capability. "GPT-5 is the first time that it really feels like talking to an expert in any topic, like a PhD level expert," Altman stated, drawing a vivid parallel to the iPhone's transition from pixelated to Retina display. The model's deployment comes with bold claims: it's smarter, faster, and more accurate, with a dramatically lower hallucination rate, positioning OpenAI to further cement its lead in the generative AI arena.

A Multi-Tiered Revolution for Developers and Users

GPT-5 isn't a one-size-fits-all release. OpenAI unveiled three tailored variants to cater to diverse needs:
- GPT-5-mini: A lightweight option priced at $0.25 per million input tokens and $2 per million output tokens, available to free ChatGPT users.
- GPT-5-nano: An ultra-efficient API-only model costing just $0.05/$0.40 per million tokens—undercutting competitors like Google's Gemini 2.5 Flash for budget-conscious developers.
- GPT-5-pro and GPT-5-thinking: Reserved for the $200/month Pro tier, offering unlimited access, enhanced power, and extended processing for complex queries. Pro users also retain legacy model access.

The chat interface now intelligently routes queries to the optimal model based on complexity and subscription level, eliminating manual selection. For everyday users, free and Plus tiers gain access to core GPT-5 capabilities, with Pro features including Google integrations. Starting next week, Pro subscribers can connect Gmail, Google Contacts, and Google Calendar, enabling ChatGPT to proactively reference personal data—a move that blurs the line between assistant and autonomous agent.

Unleashing Unprecedented Capabilities

Benchmarks shared in OpenAI's system card reveal GPT-5's prowess. In coding, it aced the SWE-Bench Verified with 74.9% accuracy (up from previous models), showcasing its ability to fix bugs and handle multi-language tasks. During the demo, post-training lead Yann Dubois tasked GPT-5 with creating a French-learning web app; within minutes, it generated a polished, interactive site—proof of its "agentic" efficiency. "It's a great coding collaborator," said Michelle Pokrass, another post-training lead. "It executes long chains and tool calls effectively, follows detailed instructions, and provides upfront explanations of its actions."

Health applications also shine: GPT-5-thinking scored 25.5% on the challenging HealthBench Hard metric, a substantial improvement validated by physicians. Crucially, hallucination rates—defined as factual errors—plummeted by 65% compared to GPT-4o, thanks to over 5,000 hours of red teaming and "safe completions" that constrain harmful outputs. Safety research lead Alex Beutel noted, "We've significantly decreased the rates of deception," though the system card admits mitigations aren't foolproof.

Why This Matters: The AGI Horizon and Industry Ripples

With 700 million weekly users and 4 million developers on its API, OpenAI's rollout isn't just an upgrade—it's a strategic play to dominate AI's next era. The expanded 256,000-token context window allows deeper document and conversation analysis, while customizable personalities (Cynic, Robot, Listener, Nerd) hint at more human-like interactions. Yet, Altman's AGI framing invites scrutiny: if GPT-5 "lacks continuous learning," as he concedes, can it truly be "generally intelligent"? This ambiguity underscores a broader tension in AI ethics and competition.

For developers, GPT-5-nano's cost efficiency could democratize AI tools, but it also pressures rivals to slash prices. Enterprises eyeing the Pro tier gain potent workflow integrations, yet the seamless data access raises privacy concerns. As AI evolves from helper to expert, one truth emerges: GPT-5 isn't just a model—it's a mirror reflecting how close we are to machines that don't just assist, but truly understand. Source: Wired