A system that decides for itself
Previously, users often had to select the appropriate model themselves, but with GPT-5, this is no longer necessary. OpenAI combines a fast standard model, a so-called "thinking" variant for complex tasks, and an intelligent router that automatically decides which version to use.
If you type something like "think about it carefully" in your input, you'll immediately trigger Reasoning Mode. Once the quota is exhausted, a mini version kicks in.
Fewer errors, smarter security
One of the biggest problems with previous models was false facts. According to OpenAI, GPT-5 reduces the hallucination rate by around 4 percent compared to GPT-45o, and the reasoning model even reduces it by 80 percent compared to OpenAI o3. Honesty has also been improved. If a task is impossible or unclear, GPT-5 more often communicates this openly, rather than gloating with a made-up answer.
Also new is the Safe Completions security concept. Instead of rigid "hard refusal"—i.e., complete rejection of risky questions—GPT-5 now provides safe, yet helpful answers. This is especially important for so-called dual-use topics, i.e., questions that have both harmless and potentially dangerous applications. For example, in pyrotechnic inquiries, the model cannot provide specific technical values, but can refer to safety standards and alternative information sources.
Strong performance in coding, writing and health
OpenAI sees GPT-5 as being ahead in three areas in particular.
- Programming: The model achieves new best scores in benchmarks such as SWE-bench Verified (74,9%) and Aider Polyglot (88%), works more efficiently with fewer token and tool calls, and impresses with a better sense of layout, typography, and white space in frontend design.
- Writing: GPT-5 understands context and style better, can consistently implement complex text forms such as free verse, and helps turn rough ideas into fluid, coherent texts.
- Health: It reacts more sensitively to potential problems, adapts its responses to the user's knowledge level and location, and provides more precise, contextual information. It remains clear that ChatGPT is not a substitute for medical advice.
Multimodal and usable as an agent
GPT-5 can handle not only text, but also images, diagrams, and other visual content, and does so significantly more precisely than before. In agent mode, complex tasks involving multiple consecutive or parallel tool calls can be processed without losing the thread. According to OpenAI, the model responds better to errors, reliably completes long workflows, and can reliably retrieve content from very long contexts.
Large companies such as Moderna, Uber, BBVA, and Lowe's are already testing GPT-5. Initial feedback praises the increased accuracy, faster execution, and the ability to complete complex projects in hours instead of weeks. According to pilot customers, GPT-5 demonstrates clear advantages, especially for multi-step automation and code-heavy workflows.
Available now
GPT-5 is now being gradually rolled out to ChatGPT. Free users have limited access to GPT-5 & GPT-5 mini. Plus and Pro subscribers benefit from higher usage limits. GPT-5 Pro with extended thinking time is available to Pro, Enterprise, and EDU customers.
Developers can choose from three model sizes, gpt-5, gpt-5-mini, and gpt-5-nano, via the API, with prices starting at $1,25 per million input tokens (90% cache discount) and $10 per million output tokens.
Source: Press release
