OpenAI GPT-5
OpenAI positioned GPT-5 as its "best AI system yet" and a single unified model that combined the fast responses of the GPT line with the deliberate reasoning of its o-series, governed by a real-time router that decides when to answer quickly versus think longer. It claimed state-of-the-art results across coding (74.9% on SWE-bench Verified), science, and math, plus expert or "PhD-level" reasoning. OpenAI emphasized sharply reduced hallucinations and less sycophancy, and Sam Altman framed it as "the best model in the world." It was made the default for all ChatGPT users, including the free tier.
GPT-5 shipped same-day across ChatGPT (default for free users, more usage for Plus, and a higher-reasoning GPT-5 Pro for $200/month Pro subscribers) and the API in three sizes — gpt-5, gpt-5-mini, gpt-5-nano — priced at $1.25 per million input tokens and $10 per million output tokens. Under the hood it is a system of fast and "thinking" variants selected automatically by a router rather than one monolithic model.
The launch was technically credible but operationally rocky, and it became a case study in change management. On day one the automatic router malfunctioned for part of the day, making GPT-5 feel — in Altman's own words — "way dumber" and producing inconsistent quality that undercut the benchmark story. The bigger misstep was removing users' ability to choose older models: abruptly deprecating the well-liked GPT-4o triggered a backlash that forced OpenAI to restore GPT-4o, raise limits, and publicly concede the deprecation was a mistake. Benchmark leadership was real, but the gap over rivals and over OpenAI's own prior models felt more incremental than the "leap" framing implied, and the router shifted control from user to vendor in ways power users resented. The verdict is reasonably settled rather than provisional, with one caveat: OpenAI has since iterated past GPT-5, so it is no longer the live flagship.
GPT-5 marks the industry's pivot from "pick your model" to opaque, router-orchestrated systems where the vendor — not the user — decides how much reasoning you get, a control and reliability trade-off executives should weigh before standardizing on it. Its launch also showed that abruptly removing a model teams depend on is a business risk, not just a product decision.