OpenAI introduced GPT‑5: everything you need to know from the launch

OpenAI has officially introduced ChatGPT‑5 — the company’s most powerful and refined model to date.

During the launch event, CEO Sam Altman framed GPT‑5 as a breakthrough not just in performance but in purpose, describing it as a step closer to artificial general intelligence (AGI).

"GPT‑5 is a major upgrade over GPT‑4 and a significant step toward AGI," Altman said. In his analogy, GPT‑3 was like a high school student, GPT‑4 a college undergrad, and GPT‑5 "a legitimate PhD expert in any area you need on demand."

So what else did OpenAI reveal at the launch? From performance leaps to new tools, integrations, and safety upgrades, we’ve summarized the key takeaways for you below.

Proactive support

GPT‑5 isn’t just more accurate or faster, it also marks a shift from passive interaction to active support. The model can now perform complex tasks like writing full programs or generating entire websites from scratch — and accomplish them in seconds. According to Altman, "software on demand" will be one of the defining features of the GPT‑5 era.

GPT‑5 also addresses a long-standing trade-off noted during the launch: users previously had to choose between the quick responses of lightweight models and the more thoughtful, reasoning-based replies of slower ones. According to OpenAI, GPT‑5 is designed to "think just a perfect amount to give you a perfect answer."

Performance gains — and a Pro version

Compared to GPT‑4o, GPT‑5 makes 45% fewer mistakes and uses 50–80% fewer tokens per task — whether it’s visual reasoning, code generation, or scientific problem-solving. For more demanding use cases, GPT‑5 Pro delivers even deeper analysis and 22% fewer errors.

Benchmarks:

In terms of benchmarks, GPT‑5 outperforms previous OpenAI models and stands out among current offerings on the market. GPT‑5 leads across major evaluations:

• 94.6% on AIME 2025 (mathematics)
• 74.9% on SWE-bench Verified (real-world software engineering)
• 88% on Aider Polyglot (cross-language programming)
• 84.2% on MMMU (multimodal reasoning)
• 46.2% on HealthBench Hard (medical diagnostics)

The model also introduces a "Study & Learn" mode to guide users step-by-step through complex topics — especially useful in voice mode, such as when learning new languages or practicing conversational skills.

Personalization gets an upgrade

GPT‑5 brings more control over how users interact with the model:

• Users can now customize the chatbot’s tone and personality — for example, making it more concise, more supportive, or even slightly sarcastic, depending on communication preferences.
• Paid users can personalize the visual appearance of the chat interface.

• Significant improvements in memory allow the model to remember user preferences, style, and context over time.
• Integration with Gmail and Google Calendar is rolling out first to Pro users. Once enabled, GPT‑5 can access your schedule, organize your day, and even generate suggested agendas, often without requiring user input. This integration allows the model to act more like a proactive assistant, able to prioritize tasks and resurface forgotten responsibilities based on your inbox.

A smarter approach to safety

GPT‑5 introduces a more comprehensive safety approach. In addition to reducing hallucinations, OpenAI focused heavily on limiting deception — instances where the model might misrepresent its capabilities or fabricate success. According to the company, deception rates in GPT‑5 reasoning responses were reduced to 2.1%, down from 4.8% in GPT‑3.5-level models.

The model also features a significant overhaul in safety training. Whereas older models tended to either fully comply with or entirely refuse ambiguous prompts, GPT‑5 uses a new system called "safe completions." Rather than rejecting a request outright, the model attempts to provide a response that remains helpful within clearly defined safety boundaries. That could mean partially answering a question, summarizing information at a high level, or, if refusal is necessary, explaining the reason and suggesting safer alternatives.

GPT‑5 is available now to Plus, Pro, and Team subscribers. Enterprise and Edu users gain access next week. Free-tier users start with GPT‑5 but transition to GPT‑5 mini when they reach usage limits — a smaller model still more capable than GPT‑3.5.