On August 7, OpenAI unveiled its new flagship model, GPT-5, the most powerful and capable model to date. OpenAI states the new model demonstrates better accuracy, faster response rates, and significantly fewer errors. It combines the best features of its predecessors and offers new capabilities in text writing, programming, medicine, visual recognition, and logical thinking. From dynamic model routing and coding breakthroughs to medical accuracy and new personalities, GPT-5 brings sweeping changes.
In this blog post, we will discuss everything you need to know before testing it yourself.
One Unified Model
Unlike previous releases, GPT‑5 is not a single model, but an integrated system with dynamic routing that is capable of adapting to the complexity of the query. It means that users no longer have to choose between reasoning or fast AI, as GPT-5 will determine the complexity of the question and choose the reasoning route itself.
GPT‑5 is family of models that differ in speed and depth of analysis and includes:
gpt-5-main - standard fast model
gpt-5-main-mini - simplified and even faster, available on a free plan once the user hits usage limits
gpt-5-thinking - “thinking” model for complex tasks
gpt-5-thinking-pro - accelerated reasoning model with parallel processing. It is intended for professionals working in areas requiring deep knowledge: medicine, science, engineering, law, and analytics.
gpt-5-thinking-nano - compact reasoning model for developers that is available only in the API.
GPT-5 offers improved accuracy, faster responses, and fewer errors, with enhanced capabilities in coding, writing, and medical tasks - building apps in one go, writing with more depth, and giving smarter, context-aware medical answers. OpenAI also notes that GPT-5 is "safer," less susceptible to misinformation and manipulation. The model is also better at recognizing malicious intent from users. Besides, as announced earlier, the new policy will be applied to the upgraded ChatGPT, it will no longer provide definitive advice on difficult personal issues.
Availability and Pricing
GPT-5 is integrated into ChatGPT as a default model on the Free, Plus, Teams, and Pro plans and via API starting on the date of launch, with the Enterprise/Edu plans getting the model next week. As with GPT-4o, the difference between free and paid access is the amount of usage. Users with free accounts have access to the model with certain limitations - after exceeding the request limit, there will be an automatic switch to the lightweight version of GPT-5 mini. The Plus plan promises higher limits on model usage, as well as access to an advanced reasoning model. Pro subscribers get unlimited access to GPT-5 and GPT-5 Pro reasoning models. They will also be able to pick through legacy models that is no longer available on other plans.
GPT-5 is available in the API at $1.25/1M input tokens and $10/1M output tokens. The API also includes lightweight GPT-5-mini priced at 25 cents/1M input tokens and $2/1M output tokens and GPT-5-nano priced at 5 cents/1M input tokens and 40 cents/1M output tokens. The context window for all models is 256K tokens.
Pro, Plus, and Team subscribers can also use GPT‑5 in Codex CLI without paying for the API separately. Users can just log in via ChatGPT to be able to work with code and shell commands in the terminal.
New Personalities Choice
ChatGPT is getting a few UI updates with the release of GPT-5. OpenAI is testing four new personalities to change how the AI responds, e.g., more professional, supportive, or sarcastic with Cynic, Robot, Listener, and Nerd. These are optional, adjustable in settings, and aim to better match different communication styles. The company says this will allow ChatGPT to tailor its responses without having to specifically ask the model to respond in a certain style.
Test Results
Speaking of benchmarks, the first thing that stands out is the level of hallucinations. According to OpenAI, it has significantly decreased compared to GPT o3.
On the HealthBench Hard Hallucinations test, which measures the accuracy of AI models answering medical questions, GPT-5 "hallucinated" only 1.6% of the time. It is significantly lower than previous models GPT-4o and GPT-3, which had 12.9% and 15.8%, respectively.
On Humanity’s Final Exam, a challenging test that measures the performance of AI models in math, humanities, and science, the advanced reasoning version of GPT-5 (GPT-5 Pro) scored 42%. That’s slightly less than what xAI was able to achieve with Grok 4 Heavy, which scored 44.4% on the test.
One of the key areas of GPT‑5 application is programming and automation of engineering tasks. Thanks to its enhanced reasoning abilities and improved tool support, the model excels at generating, debugging, and analyzing code in real-world scenarios. Importantly, GPT-5 is designed with safety in mind — it refuses to create malicious scripts, avoids provocative requests, and steers clear of discussing exploits, even if they are veiled. Unlike models that simply write code, GPT-5 writes code like a person, matching it with the style of the existing project and making its output more natural and consistent. The company claims the model can autonomously create complex applications from a single prompt, accurately handle tools and files. It does not get lost in a long context. During a demonstration at a press briefing, OpenAI's head of post-training, Yann Dubois, showed how GPT-5 created a full-fledged French-learning website in seconds. The model wrote hundreds of lines of code on its own, and everything ran without errors in real time.
When speaking specifically about benchmarks, the model scored 74.9% on fixing real bugs (SWE-Bench Verified) and 88% on working with different programming languages (Aider Polyglot). Just to compare, Anthropic's Claude Opus 4.1 scored 74.5% on SWE-Bench Verified, while Google DeepMind's Gemini 2.5 Pro scored 59.6%.
GPT‑5 is fully integrated into Codex CLI, a tool that allows developers to use the model as an interactive assistant in the command line without the need to switch between windows. ChatGPT users with a Plus, Pro, or Team subscription can run Codex CLI at no additional cost for the API.
At the same time, when it comes to Tau-bench - on tasks simulating website navigation, GPT-5 shows mixed performance, with 62.6% on airline websites (slightly worse than o3) and 81.1% on shopping platforms, which is lower than, for instance, Claude Opus 4.1's 82.4%.
Bottom Line
GPT-5 represents a significant leap forward in AI technology. It adapts to query complexity and delivers accurate, context-aware responses. The model’s ability to autonomously create complex applications, as demonstrated by the French-learning website, sets a new benchmark for AI in programming and automation. It marks a big step toward AGI (artificial general intelligence). Still, as Sam Altman, OpenAI’s CEO, said, it's not there yet. While ChatGPT has 700 million weekly users, 5 million paying users, and 4 million developers utilizing the API, OpenAI hasn’t had the most powerful model on the market since the release of GPT-4. With GPT-5’s launch, the company aims to regain its leadership, especially in key industries like medicine, software development, and law.
FAQ:
What is GPT-5 and how is GPT-5 different from previous models?
GPT-5 is OpenAI’s latest flagship AI model, unveiled on August 7, 2025. Unlike its predecessors, it’s an integrated system with dynamic model routing, automatically adjusting to the complexity of queries. It offers improved accuracy, faster responses, and fewer errors, with enhanced capabilities in coding, medical tasks, text generation, and logical reasoning.
What new capabilities does GPT-5 have?
GPT-5 can build complex apps with a single prompt, write more naturally styled code, provide context-aware medical answers, and maintain accuracy even in long conversations.
How well does GPT-5 perform in tests?
GPT-5 shows significant improvements in accuracy, with much lower hallucination rates in medical tests and strong coding benchmarks. It scored 74.9% on bug fixing and 88% on multi-language programming tasks.
Is GPT-5 safer to use?
Yes, GPT-5 is designed to reduce misinformation, recognize harmful intent, refuse to create malicious code, and avoid risky or inappropriate advice.
What are the new ChatGPT personalities?
Four optional personalities are being tested — Cynic, Robot, Listener, and Nerd — allowing users to customize how ChatGPT responds in tone and style.
Can I still access older models like GPT-4o with GPT-5’s release?
Pro subscribers can access legacy models, but Free, Plus, and Team plans are limited to GPT-5 and its variants (e.g., GPT-5-mini) after the release.
How can I access GPT-5?
GPT-5 is integrated into ChatGPT across Free, Plus, Teams, and Pro plans, with varying usage limits. It’s also available via API and fully integrated into the Codex CLI for developers.
What is Codex CLI, and how does it work with GPT-5?
Codex CLI is a command-line tool that integrates GPT-5 as an interactive assistant, helping developers write code and execute shell commands without switching windows. It’s free for ChatGPT Plus, Pro, and Team subscribers, making coding workflows smoother and more efficient.
