OpenAI launched GPT-5 on August 7, 2025, and made it the new default model in ChatGPT the same day. GPT-5 is not a single model in the old sense. OpenAI describes it as a unified system that pairs a fast model for most questions with a deeper reasoning model for harder ones, and a real-time router that picks between them. This guide explains what changed, what the numbers mean, what it costs, and what to check before you change how you work.
GPT-5 at a glance: what changed
GPT-5 launched on August 7, 2025 as the new default in ChatGPT for signed-in users, replacing GPT-4o, OpenAI o3, OpenAI o4-mini, GPT-4.1 and GPT-4.5. It is a unified system with a fast model, a deeper reasoning model called GPT-5 Thinking, and a real-time router. Free users got it on day one, with usage limits.
The change that matters most for everyday users is that the model picker stopped being the main decision. Before GPT-5, ChatGPT offered a menu of named models, and users had to guess which one fit the task. GPT-5 takes that choice away in most cases. Open the app, type a question, and the system decides whether a quick answer will do or whether it should reason first. OpenAI says it plans to merge these abilities into a single model in the near future, so the router is a step on the way to that goal, not the final design.
Rollout started the same day for Free, Plus, Pro and Team users, according to OpenAI. Enterprise and Education access followed the next week. Nick Turley, OpenAI vice president of ChatGPT, told TechCrunch that this gave free users access to an AI reasoning model for the first time, since earlier reasoning models sat behind a paywall.
The unified system and the real-time router
GPT-5 is a unified system with a smart, efficient model for most questions, a deeper reasoning model for harder problems, and a real-time router that picks between them. The router looks at the conversation type, the complexity of the request, the tools needed, and any explicit instruction from the user, such as asking it to think hard about a problem.
The router is the part of GPT-5 that changes how ChatGPT feels to use. A simple question gets a fast answer, while a harder task, such as planning a project or working through a math problem, gets more thinking time. OpenAI says the router is trained on real signals, including when users switch models by hand and measured correctness, so its decisions can change over time.
Users still have some control. OpenAI says paid users can select GPT-5 Thinking from the model picker, and anyone can type a phrase such as “think hard about this” to push the system toward reasoning. Once a free user reaches the GPT-5 usage limit, the system moves them to GPT-5 mini, a smaller and faster model. If an answer looks shallow for a hard question, try the explicit instruction first, then check whether you have hit a usage limit.
GPT-5 Thinking and GPT-5 Pro
GPT-5 Thinking is the deeper reasoning part of the GPT-5 system, used for harder problems that need more steps. GPT-5 Pro is a separate version for the most demanding tasks, with extended reasoning, and it is available to Pro subscribers. It replaces OpenAI o3-pro as the top option in the GPT-5 family.
GPT-5 Pro exists for tasks where a wrong or thin answer costs real time. OpenAI says it thinks for longer to give the most complete answers it can. In tests run by OpenAI on more than 1,000 real-world reasoning prompts, outside experts preferred GPT-5 Pro over GPT-5 Thinking 67.8 percent of the time, and OpenAI says it made 22 percent fewer major errors. Those are OpenAI figures from its own tests, not independent results, so read them as the company case for the Pro tier rather than proof that every Pro answer will be better.
The practical difference is speed against depth. GPT-5 Thinking is the reasoning mode most paid users will meet. GPT-5 Pro takes longer and uses more compute, so it fits a long report, a hard research question, or a complex coding problem, not a quick question. OpenAI also says GPT-5 with thinking beat o3 while using 50 to 80 percent fewer output tokens across capabilities such as visual reasoning and agentic coding. Fewer tokens per answer can mean faster replies and lower cost, but the gain depends on the task.
Coding and agentic work
GPT-5 is OpenAI’s strongest coding model to date, according to the company, with clear gains in complex front-end generation and debugging larger repositories. It can build a website, app or game from a single prompt, and it works better as an agent that carries out multi-step tasks with tools.
Coding was the center of the GPT-5 launch. OpenAI says GPT-5 often creates a working single-page app, website or game from one prompt, with attention to spacing, typography and white space. On SWE-bench Verified, a test based on real software engineering tasks, GPT-5 scored 74.9 percent, up from 69.1 percent for o3. On Aider Polyglot, a coding test, it scored 88 percent. TechCrunch reports that the SWE-bench score sits just ahead of Anthropic Claude Opus 4.1 at 74.5 percent, and ahead of Google DeepMind Gemini 2.5 Pro at 59.6 percent.
Agentic work is the other half of the coding story. OpenAI says GPT-5 is better at instruction following and at using tools, which lets it carry out multi-step requests and adapt when the context changes. In practice, that means asking it to build a small app, check the result, and fix errors, rather than asking for one block of code at a time. Developers should still test the output, because a model that writes more code in one pass can also write more errors in one pass.
Writing, health and everyday answers
GPT-5 is OpenAI’s most capable writing collaborator yet, and its best model for health-related questions, according to the company. It handles reports, emails and memos better than GPT-4o, and it acts more like an active partner on health questions, flagging concerns and asking follow-up questions instead of giving a flat answer.
Writing is one of the three uses OpenAI named as most common in ChatGPT, along with coding and health. OpenAI says GPT-5 helps turn rough ideas into writing with literary depth and rhythm, and it handles structural ambiguity better, such as sustaining unrhymed iambic pentameter. The everyday point is simpler. The model is meant to draft and edit reports, emails and memos with less hand-holding. The best way to judge that claim is to give it a real draft and check whether the result needs less editing than GPT-4o needed.
Health answers need more care. OpenAI says GPT-5 scores far higher than any earlier model on HealthBench, an evaluation based on realistic scenarios and physician-defined criteria, and that it adapts to the user context, knowledge level and geography. It also says the model proactively flags potential concerns and asks questions to give more helpful answers. That is useful for understanding a test result or preparing questions for a doctor. It is not a substitute for a medical professional, and a health answer from any chatbot should be checked against a clinician, especially when it affects treatment or a decision about care.
Benchmarks: what OpenAI reported
OpenAI reported state-of-the-art results for GPT-5 in math, coding, multimodal understanding and health at launch. The headline figures are 94.6 percent on AIME 2025 without tools, 74.9 percent on SWE-bench Verified, 88 percent on Aider Polyglot, 84.2 percent on MMMU, and 46.2 percent on HealthBench Hard.
Benchmarks need context. OpenAI ran these tests at high reasoning effort settings, and the results for GPT-4o reflect the most recent version of that model in ChatGPT as of August 2025. A benchmark score is a result on one test, not a measure of how good the model will be on your work. The scores below come from OpenAI, and they are the company’s own results.
AIME 2025, a math competition test, gave GPT-5 94.6 percent without tools. SWE-bench Verified gave 74.9 percent. Aider Polyglot gave 88 percent. MMMU, a multimodal test, gave 84.2 percent. HealthBench Hard gave 46.2 percent. With GPT-5 Pro extended reasoning, the model scored 88.4 percent on GPQA without tools, a test of difficult science questions. TechCrunch also reports mixed results elsewhere. On Humanity’s Last Exam, GPT-5 Pro scored 42 percent with tools, just below Grok 4 Heavy at 44.4 percent. On Tau-bench, GPT-5 scored 63.5 percent on airline tasks, just below o3 at 64.8 percent, and 81.1 percent on retail tasks, just below Claude Opus 4.1 at 82.4 percent. GPT-5 leads on the tests OpenAI chose to headline, and sits close to rivals on other tests.
Accuracy, hallucinations and safety
GPT-5 is less likely to hallucinate than earlier OpenAI models, according to the company, and it is more honest about what it cannot do. With web search on, its answers contain a factual error about 45 percent less often than GPT-4o, and about 80 percent less often than o3 when it is thinking.
Hallucinations, where a model makes up information with confidence, were a known weakness of OpenAI reasoning models. OpenAI says GPT-5 with thinking shows about six times fewer hallucinations than o3 on open-ended fact-seeking prompts from the LongFact and FActScore benchmarks. TechCrunch reports the same pattern. GPT-5 with thinking hallucinated 4.8 percent of the time, against 22 percent for o3 and 20.6 percent for GPT-4o. On health hallucinations, it scored 1.6 percent, against 12.9 percent for GPT-4o and 15.8 percent for o3. Those are large drops, but 4.8 percent is not zero, and a single wrong fact can still mislead a reader who does not check it.
Safety changed in method as well as in results. OpenAI introduced safe completions, a form of safety training that teaches the model to give the most helpful answer it can within safety limits, rather than refusing outright. That can mean a partial answer, and a clear reason when it does refuse, plus a safe alternative. Deception also fell in OpenAI tests, from 4.8 percent of o3 reasoning responses to 2.1 percent for GPT-5, and sycophancy, or excessive agreement, fell from 14.5 percent to under 6 percent in targeted tests. GPT-5 Thinking was treated as High capability in the Biological and Chemical domain, with safeguards and 5,000 hours of red-teaming with partners such as CAISI and UK AISI.
ChatGPT plans, limits and price
GPT-5 is the default model for Free, Plus, Pro and Team users, with the difference between plans set by usage volume rather than by a different model. Plus costs 20 dollars a month and gives higher usage limits than free, while Pro costs 200 dollars a month, gives unlimited GPT-5 access, and adds GPT-5 Pro.
Free access is the biggest change in pricing, even though GPT-5 itself has no price for ChatGPT users. Free users get GPT-5 as their default model, and when they reach their usage limit they move to GPT-5 mini. OpenAI said full reasoning access for free users could take a few days to roll out fully at launch. Plus subscribers get more usage than free users, and Pro subscribers get unlimited access to GPT-5 plus GPT-5 Pro.
ChatGPT also gained four new preset personalities with GPT-5, in a research preview for all users. The options are Cynic, Robot, Listener and Nerd, and they are opt-in and adjustable in settings. OpenAI says they work because GPT-5 follows instructions better, so a user can set a tone once instead of asking for it in every prompt. They were available first for text chat, with voice to follow.
API pricing, sizes and developer controls
Developers get GPT-5 in three sizes through the OpenAI API: gpt-5, gpt-5-mini and gpt-5-nano. The base model costs 1.25 dollars per million input tokens and 10 dollars per million output tokens, while mini costs 0.25 dollars and 2 dollars, and nano costs 0.05 dollars and 0.40 dollars on the same basis.
The three sizes let developers trade performance against cost and speed. The full GPT-5 model is the reasoning model that powers maximum performance in ChatGPT, while the smaller sizes suit faster or cheaper work. OpenAI also gives developers a verbosity control in the API, with low, medium and high settings, so a developer can ask for short or long answers without rewriting the prompt. TechCrunch notes that one million input tokens is about 750,000 words, longer than the Lord of the Rings series.
The API version of GPT-5 is not the same as the ChatGPT system. In ChatGPT, GPT-5 is a system of reasoning, non-reasoning and router models. In the API, GPT-5 is the reasoning model, and the non-reasoning model used in ChatGPT is available as gpt-5-chat-latest. Developers can also start coding with GPT-5 in Codex CLI by signing in with ChatGPT, according to OpenAI.
What changed for users, and what to do next
The main change for most users is simple. ChatGPT now decides for itself when to answer fast and when to think, and the old model menu is no longer the first choice. That saves time on easy questions, and it puts more weight on how you phrase a hard question, since the router reads the prompt for signs of complexity.
Start with a short test on your own work. Ask GPT-5 to do a task you already did with GPT-4o, such as drafting a report or writing a piece of code, and compare the two results side by side. If the answer is thin, add a clear instruction such as “think hard about this” and see whether the result changes. Paid users can also pick GPT-5 Thinking directly from the model picker. The same test works on the ChatGPT app, whether you use an iPhone 17 or an Android 16 device, and on a desktop, including on a Mac running macOS 27.
Check the limits before you rely on GPT-5 for a deadline. Free users move to GPT-5 mini after their limit, and that smaller model may answer a hard question with less depth. If you use ChatGPT for health questions, treat the answer as a starting point for a conversation with a clinician, not as a diagnosis. If you use it for code, run the code and test it, because a confident answer is not proof that it works. The safe way to use it is to let it do the first draft, then check the facts, the code and the reasoning yourself.
Frequently asked questions
When did OpenAI launch GPT-5?
OpenAI launched GPT-5 on August 7, 2025, and rolled it out the same day to Free, Plus, Pro and Team users in ChatGPT, with Enterprise and Education access following the next week. It also launched in the OpenAI API on the same day.
Is GPT-5 free to use?
Yes, GPT-5 is the default model for free ChatGPT users, with usage limits. Once free users reach their limit, they move to GPT-5 mini. Plus costs 20 dollars a month for higher limits, and Pro costs 200 dollars a month for unlimited GPT-5 access and GPT-5 Pro.
What is the GPT-5 router?
The router is the part of the GPT-5 system that decides whether a fast model or the deeper GPT-5 Thinking model should answer. It looks at the conversation type, the complexity of the request, the tools needed, and any explicit instruction from the user, such as asking it to think hard about a problem.
What is GPT-5 Thinking?
GPT-5 Thinking is the deeper reasoning model inside the GPT-5 system, used for harder problems that need more steps. Paid users can select it from the model picker, and any user can ask for it with a phrase such as “think hard about this” in the prompt.
What is GPT-5 Pro?
GPT-5 Pro is a version of GPT-5 with extended reasoning for the most demanding tasks. It replaces o3-pro, and it is available to Pro subscribers. OpenAI says outside experts preferred it over GPT-5 Thinking 67.8 percent of the time in its own tests, with 22 percent fewer major errors.
How much does the GPT-5 API cost?
The base GPT-5 model costs 1.25 dollars per million input tokens and 10 dollars per million output tokens. GPT-5 mini costs 0.25 dollars and 2 dollars, and GPT-5 nano costs 0.05 dollars and 0.40 dollars on the same basis. Check the current OpenAI pricing page before you build, because API prices can change.
Is GPT-5 better than GPT-4o?
On the tests OpenAI published, yes. GPT-5 scored 94.6 percent on AIME 2025 without tools and 74.9 percent on SWE-bench Verified, and its answers contain a factual error about 45 percent less often than GPT-4o with web search on. On some outside tests, such as Tau-bench, it sits close to rivals, so the gain depends on the task.
Sources
- OpenAI, Introducing GPT-5: https://openai.com/index/introducing-gpt-5/
- TechCrunch, OpenAI’s GPT-5 is here: https://techcrunch.com/2025/08/07/openais-gpt-5-is-here/
3 comments