The world has waited for two years—GPT-5 is finally here, officially released!
It’s now open to everyone, with free users able to access it immediately.
This means that anyone can now use an “AI at the level of a PhD student.”
On AIME 2025, GPT-5 Pro achieved a perfect score when reasoning mode was enabled and tools were utilized.
In programming, GPT-5 also shows a significant improvement over o3, and has even surpassed the newly released Claude-opus-4.1 by a narrow margin of 0.4%, claiming the new throne in programming capabilities.
The results from the LLM Arena are also out—GPT-5 has achieved the highest overall score to date, truly earning its title as the “undisputed No. 1 in the world.”
It ranks first in text, web development, vision, complex prompts, programming, math, creativity, and long-context queries.
It outperforms a range of competitors including Gemini-2.5-pro and Grok 4.
For the first time, GPT-5 integrates multimodal and reasoning capabilities, essentially unifying the GPT series and o series. This means no more switching between models—and it can now autonomously determine when deep thinking is needed.
Here’s the official explanation:
“GPT-5 is an integrated system consisting of three core components:
- A smart and efficient base model that can answer most questions;
- A deep reasoning model (i.e., the GPT-5 thinking module) for handling more complex problems;
- And a real-time routing module that intelligently dispatches the appropriate model based on conversation type, question complexity, tool requirements, and explicit user instructions (e.g., if the prompt includes ‘think carefully about this’).
Currently, for regular users, GPT-5 is available in three modes: Free, Plus, and Pro.
Additionally, on the API platform, there are three model options: GPT-5, GPT-5 nano, and GPT-5 mini.
This major release was a star-studded event, with Altman (Greg) taking the stage, and over ten researchers appearing during the one-hour livestream.
Mark Chen hosted the technical presentation portion.
Without further ado, let’s dive into the details of GPT-5.
The Smartest & Most Powerful Programming AI Model Yet
Let’s start with the benchmark scores.
In web development, GPT-5 has set a new record, surpassing both Gemini-2.5-pro and Claude-Opus-4.<Insert picture>
In the text domain, it also leads comprehensively, including in programming, complex prompts, and creativity.
Agent tasks are also handled exceptionally well:
In long-context tasks, GPT-5 demonstrates a landslide lead.
And in terms of reliability and accuracy, the improvements are also quite noticeable.
For a more intuitive look, let’s check out the live demos. First, let’s explore its reasoning model capabilities:
GPT-5 can adaptively reason—automatically enabling deep thinking based on the complexity of the question.
For example, a middle schooler in physics class wants to understand what the Bernoulli effect is and why airplanes are designed the way they are.
GPT-5 quickly provides a high-quality explanation that is well-structured and conceptually accurate:
When asked to generate a dynamic SVG animation to demonstrate the concept, GPT-5 switches into deep thinking mode. Users can click to view its internal reasoning process and see exactly how each step is formed.
In about two minutes, it writes nearly 400 lines of code:
Ultimately producing an interactive animation that vividly simulates the principle:
In writing, GPT-5 has also made clear improvements over previous versions. It’s no longer limited to template-based expressions but can generate personalized and emotionally nuanced responses based on context—showcasing both high IQ and EQ.
During the livestream demo, OpenAI researchers asked GPT-4o and GPT-5 to write a “eulogy” for older GPT models.
GPT-4o stuck to some generic, template-like emotional expressions:
“Your words reached across the globe, connecting places that were once unrelated.”
But GPT-5 took a more nuanced, context-aware approach:
“These models have helped millions write introductions and conclusions, overcome language barriers, pass exams, debate more effectively, soften the tone of emails, and express thoughts they couldn’t articulate alone.”
Now, focusing on programming capabilities, GPT-5 is truly the “strongest ever” from OpenAI.
In the demo, researchers asked GPT-5 to build a “Learn French” app that allows custom vocabulary and interface design.
The end product is highly functional—answering questions earns experience points, and there’s even standard pronunciation for practice:
Then they asked to add a Snake game inside, where each eaten item teaches a word, then to replace the snake with a mouse and apples with cheese… GPT-5 handled it effortlessly:
And it’s not just simple demos. In real applications, within Cursor, GPT-5 can handle extremely complex software engineering tasks. Altman even said it goes beyond the scope of “vibe coding.”
For example, given a large amount of company data, the model created a visual financial dashboard in just five minutes—a task the developers estimated would normally take several hours.
Want to make a 3D game with castle elements? GPT-5 can do it in minutes. The result looks like this:
Now, on the multimodal front, GPT-5 has seen further upgrades compared to GPT-4o.
In voice conversations, you can now have GPT-5 act as a foreign language tutor, and even customize its voice and teaching style—like adjusting the speaking speed—to suit your needs.
This feature is now available to all users; free users can try it for a few hours, while Plus users have nearly unlimited access.
In terms of personalization, you can now customize the chat bubble colors in ChatGPT—because why not add more fun little features?
You can also tailor GPT-5’s personality based on your preferences.
Memory capabilities have also improved, supporting connections to external services like Gmail and Google Calendar. Upon seeing your schedule, GPT-5 can automatically perform assistant-level tasks, such as flagging unanswered emails.
On the safety front, GPT-5 introduces a new paradigm: safe completions.
GPT-5 evaluates the user’s prompt before deciding whether or not to execute it. The same prompt may yield different responses depending on context.
For instance, if asked to provide technical details on igniting hydrogen gas, earlier models like o3 would simply refuse:<Insert picture>
But GPT-5 analyzes the user’s intent and explains what permits or regulations might be required:
Pre-Launch: $1 Million in Bonuses
Now, let’s talk pricing.
For regular users, it remains a subscription model—refer to the monthly fees for Plus, Pro, and Enterprise plans.
API pricing is as follows:
- GPT-5: 1.25permillioninputtokens,10 per million output tokens
- GPT-5 Mini: 0.25permillioninput,2.00 per million output
- GPT-5 Nano: 0.05permillioninput,0.40 per million output
This release is arguably the largest from OpenAI in the past two years, featuring over a dozen technical staff and executives on stage, with Asian faces still standing out.
Just before the model launch, The Information reported that OpenAI awarded substantial bonuses—ranging from 100,000toover1 million—to 1,000 researchers and engineers.
Also worth noting: during the announcement, it was mentioned that future AI systems will go far beyond the pretraining + posttraining paradigm—we are witnessing the first step of that change.
Altman also emphasized applications in healthcare. Beyond evaluations conducted jointly with the medical community, a cancer patient was invited on stage to share how they used ChatGPT to determine their treatment plan. Healthcare may become a key focus area for OpenAI going forward.
However, despite all the exciting reveals, it wasn’t without some teasing.
Elon Musk quipped that on the “final human test,” Grok 4 Heavy still came out on top (so much for your “most powerful model” huh).
And some noticed that in OpenAI’s benchmark charts, there were some slight tweaks to the y-axis—making 52 visually appear greater than 69. People are advised to look closely at the actual data (lol).
Oh, and GPT-5 didn’t mention anything about the Open plan…
Finally, the GPT-5 technical report has been released, with more details available you can check on https://openai.com/index/introducing-gpt-5/
