OpenAI has officially introduced GPT‑5, the newest version of its AI system, now available to all ChatGPT users. This release brings measurable improvements across key domains like coding, writing, health, math, and visual reasoning—without the marketing fluff. GPT‑5 has been built to feel more like an expert thought partner, with careful attention to accuracy, usefulness, and adaptability.

GPT‑5 operates as a unified system. Depending on your prompt, it can respond quickly or apply deeper reasoning through “GPT‑5 thinking.” A built-in router selects the right approach based on the task’s complexity, user preferences, or specific instructions like “think hard about this.” When usage limits are reached, a lighter version of the model GPT‑5 mini continues handling queries.

Practical Improvements That Matter

OpenAI focused on three major areas where people use ChatGPT the most: writing, coding, and health.

1. Writing: GPT‑5 offers better support for shaping rough ideas into clear, structurally sound writing. Whether it’s an email, report, or a more creative piece, the model balances form and expression with greater fluency.
2. Coding: It handles large codebases more effectively and generates cleaner front-end designs in a single prompt. Early testers noted improvements in layout design and responsiveness.
3. Health: GPT‑5 scored the highest yet on HealthBench, OpenAI’s medical reasoning benchmark. It’s more aware of user context and better at flagging concerns while helping users ask informed questions—without replacing a medical professional.

Reliable, Responsible, and Straightforward

GPT‑5 is more accurate than previous versions. Evaluations show it makes fewer factual errors especially when using its reasoning capability by as much as 80% compared to OpenAI o3. It also better understands when not to answer, and explains why when it refuses a request.

In contrast to previous models, GPT‑5 avoids overly agreeable or flattering responses. OpenAI worked to reduce unnecessary sycophancy by more than half, and early tests show it's more honest about uncertainty and limitations.

Safer Completions, Not Just Refusals

A major change in GPT‑5 is the shift from refusal-based safety to “safe completions.” Instead of bluntly refusing certain prompts, the model now aims to provide helpful but responsible responses, especially in complex or dual-use areas like virology. It explains safety boundaries and offers alternatives when needed.

Better Instruction Following & Personalization

GPT‑5 responds more accurately to custom instructions. Users can also choose from four preset personalities Cynic, Robot, Listener, and Nerd to adjust tone and interaction style. These new options are opt-in and meet OpenAI’s internal safety and integrity standards.

Smarter Across the Board

From math to science to spatial reasoning, GPT‑5 has set new records on key benchmarks:

1. 94.6% on AIME 2025 (math)
2. 74.9% on SWE-bench Verified (real-world coding)
3. 84.2% on MMMU (multimodal understanding)
4. 46.2% on HealthBench Hard (health reasoning)
5. 88.4% on GPQA with GPT‑5 pro

GPT‑5 also uses fewer tokens to reach better results, making it more efficient than earlier models like o3.

GPT‑5 Pro for Heavy Lifting

For the most demanding tasks, GPT‑5 pro (an enhanced version of GPT‑5) is available to Pro subscribers. It offers longer, deeper reasoning and was preferred by experts in nearly 68% of high-stakes testing scenarios, especially in science, health, and engineering tasks.

Access and Availability

GPT‑5 is now the default model in ChatGPT, replacing GPT‑4o and others for signed-in users. Plus and Pro subscribers get higher usage volumes and full access to GPT‑5 pro. Free-tier users are also gaining access gradually, with GPT‑5 mini stepping in when limits are reached.

Enterprise and educational access is set to roll out within the week.