GPT-5 Review: Specs, Benchmarks & Pricing

GPT-5 is the model OpenAI launched on August 7, 2025, and at the time it became the default model in ChatGPT for every user. Instead of making people choose between a fast model and a reasoning model, GPT-5 combined both in one system that decides how hard to think about each request. It also brought big gains in math, coding, writing, and factual accuracy over GPT-4o and o3.

This GPT-5 review covers how the model works, what its benchmarks showed, how the full model compares with GPT-5 mini and nano, what it costs, and where it stands now that newer GPT models have arrived. GPT-5 is still available in Chat Smith on Android, iOS, and the web, so you can try it without a ChatGPT subscription.

GPT-5 is OpenAI's fifth-generation large language model and the first to work as a unified system. In ChatGPT, it pairs a fast main model with a deeper reasoning model, GPT-5 thinking, and a real-time router picks between them based on how complex the question is, whether it needs tools, and what the user asks for. Saying "think hard about this" nudges it toward deeper reasoning.

For developers, OpenAI released GPT-5 in three sizes: the full model, GPT-5 mini, and GPT-5 nano. All three accept text and images, support a 400,000-token context window, and let developers control reasoning effort, including a new minimal setting for fast answers. ChatGPT Pro subscribers also got GPT-5 pro, a version that reasons even longer on the hardest problems.

GPT-5 is no longer OpenAI's newest model. OpenAI retired GPT-5 Instant and Thinking from ChatGPT on February 13, 2026, after a series of GPT-5.x updates, while keeping GPT-5 available in the API. Our full list of ChatGPT models shows where it fits in the lineup.

Four areas set GPT-5 apart from the models it replaced.

GPT-5 scored 94.6% on the AIME 2025 math competition without tools, and OpenAI reported that its thinking mode matched or beat o3 while using 50% to 80% fewer output tokens. Because reasoning is built in, users no longer had to switch models for hard questions. That suits multi-step problems in math, science, and logic, the kind of work an AI homework solver walks through step by step.

OpenAI called GPT-5 its strongest coding model at launch. It scored 74.9% on SWE-bench Verified and 88% on Aider Polyglot, and OpenAI highlighted its eye for design: spacing, typography, and layout when building websites and apps from a single prompt. It could also chain dozens of tool calls in a row without losing track of the task.

With web search on, GPT-5's responses were about 45% less likely to contain a factual error than GPT-4o's, and in thinking mode about 80% less likely than o3's. OpenAI also trained it to admit when a task is impossible rather than bluff, to flatter users less, and to give partial "safe completions" instead of refusing borderline but legitimate questions outright.

GPT-5 is a stronger writer than its predecessors, better at keeping structure and rhythm in long pieces such as reports, essays, and poems, which makes it a capable engine for any AI writing tool. It scored 84.2% on MMMU, a test of reasoning over images and charts, and 46.2% on HealthBench Hard, where OpenAI said it asks better follow-up questions and flags concerns more proactively.

OpenAI launched GPT-5 in three API sizes. The table compares them using the scores OpenAI reported at launch, with all three set to high reasoning effort.

GPT-5 mini keeps most of the full model's reasoning at a fifth of the price, which makes it the practical default for many everyday tasks. If you are coming from OpenAI's earlier flagship, our GPT-5 vs GPT-4o comparison shows how much changed between the two generations.

In the OpenAI API, GPT-5 costs $1.25 per million input tokens and $10 per million output tokens, with cached input 90% cheaper. In ChatGPT, it launched as the default model across every plan:

Since February 13, 2026, ChatGPT no longer offers GPT-5, but the model remains in the API. In Chat Smith, you can still use GPT-5, GPT-5 mini, and GPT-5 nano without an OpenAI account: free accounts include limited use, and Chat Smith Pro raises usage limits for $9.99 per month or $39.99 per year.

GPT-5 is no longer OpenAI's best model. A string of GPT-5.x releases followed, and OpenAI's current lineup is led by the GPT-6 family. If you need the strongest OpenAI model today, see our GPT-6 Sol review instead.

The launch itself was rocky. Some ChatGPT users found that the automatic router sometimes picked the fast model for questions that needed more thought, and others missed GPT-4o's warmer tone, which led OpenAI to bring GPT-4o back for paid users for a while and add manual mode controls.

Its headline benchmark scores are also OpenAI's own. Thinking mode is slower and uses more tokens than quick replies, and the 400,000-token context window is smaller than the 1M-token windows many newer models offer.

You do not need an OpenAI account to use GPT-5. It is available in Chat Smith on Android, iOS, and the web: pick it from the model list and start chatting.

GPT-5 remains a solid choice for writing, analysis, and multi-step reasoning. When you want OpenAI's latest models, Chat Smith also offers GPT-6 Sol for demanding work and GPT-6 Luna for fast everyday questions.

Because Chat Smith also includes models from Anthropic, Google, xAI, and DeepSeek, you can send the same prompt to GPT-5 and its rivals and keep the answer that holds up. For research-heavy questions, the built-in Deep Research feature runs advanced search and reasoning across sources.

GPT-5 changed how people use ChatGPT: one model that decides when to think, strong scores in math and coding, fewer hallucinations, and better writing. Even after leaving ChatGPT, it remains a reliable, well-priced model for reasoning and long-form work, and GPT-5 mini is still a smart pick for everyday tasks.

For the hardest problems, newer GPT-6 models have moved ahead. The easiest way to choose is to try GPT-5 and its successors side by side in Chat Smith, an AI chatbot that puts OpenAI's models next to the latest from Anthropic, Google, xAI, and DeepSeek.