GPT-5

What it is

GPT-5 is a large language model (LLM) developed by OpenAI. It was released on 7 August 2025 at a livestream event, as the fifth GPT foundation model, succeeding GPT-4.^[1][2]

OpenAI's system card describes GPT-5 as a unified system rather than a single model. It has a smart, fast model for most questions and a deeper reasoning model for harder problems. A real-time router sits in front of the two and chooses between them based on the type of conversation, its complexity, whether tools are needed and the user's explicit intent.^[1]

Origins and lineage

OpenAI's stated succession maps GPT-4o to gpt-5-main, GPT-4o-mini to gpt-5-main-mini, OpenAI o3 to gpt-5-thinking, o4-mini to gpt-5-thinking-mini, GPT-4.1-nano to gpt-5-thinking-nano and o3 Pro to gpt-5-thinking-pro.^[1] In broader terms, GPT-5 is a successor of GPT-4, GPT-4o and o3. It was followed by GPT-5.1 in November 2025 and GPT-5.2 in December 2025.^[6]

Main features

The family has several variants. gpt-5-main and gpt-5-main-mini are the fast models. gpt-5-thinking, gpt-5-thinking-mini and gpt-5-thinking-nano are the reasoning variants offered through the API. gpt-5-thinking-pro is offered in ChatGPT and uses parallel test-time compute, meaning it spends extra computing effort on a problem at answer time.^[1]

OpenAI reported reliability gains. The hallucination rate, the share of answers in which a model confidently states false information, was 26% lower for gpt-5-main than for GPT-4o, and 65% lower for gpt-5-thinking than for OpenAI o3.^[1] All GPT-5 models use what OpenAI calls safe-completions safety training, which replaces hard refusals.^[1]

Under OpenAI's Preparedness Framework, gpt-5-thinking was treated as High capability in the biological and chemical domain. OpenAI activated the associated safeguards as a precaution.^[1]

Developer specifications

In the API, GPT-5 has a 400,000-token context window, of which 272,000 tokens can be input, and it can produce up to 128,000 output tokens. A token is a short unit of text, roughly a word or part of a word, that the model reads and writes. The knowledge cutoff is 30 September 2024. The model accepts text and image input, and developers can set reasoning effort to minimal, low, medium or high.^[3]

Prices are quoted per million tokens. GPT-5 costs $1.25 for input and $10 for output. GPT-5 mini costs $0.25 and $2, and GPT-5 nano costs $0.05 and $0.40.^[3][5] GPT-5 Pro reached the API on 6 October 2025 at $15 per million input tokens and $120 per million output tokens.^[5]

Access and reception

GPT-5 was made available to all ChatGPT users. Plus subscribers received higher limits, and Pro subscribers received unlimited GPT-5 and GPT-5 Pro access. The model was also offered in Microsoft Copilot.^[2]

Reception was mixed. Users reported inconsistent quality from the router and objected to the removal of older models. OpenAI restored GPT-4o as an option for Plus users, and on 15 August 2025 it shipped a warmer personality update.^[2]

Retirement

On 11 June 2026 OpenAI deprecated the original GPT-5, GPT-5 mini, nano and pro API snapshots, which are fixed versions of a model. Shutdown is set for 11 December 2026.^[4]