The technical foundations of ChatGPT trace back to the transformer deep-learning architecture introduced by researchers at Google Brain in the 2017 paper "Attention Is All You Need".¹⁹ Building on this self-attention mechanism, which enabled neural networks to process sequences of text in parallel rather than sequentially, OpenAI introduced its first Generative Pre-trained Transformer (GPT-1) in June 2018, trained on a corpus of over 7,000 unpublished books with 117 million parameters.²⁰ OpenAI scaled the approach with GPT-2 in 2019 and the 175-billion-parameter GPT-3 in June 2020, followed by InstructGPT in January 2022, which used human feedback to make language models follow user instructions more reliably.

On November 30, 2022, OpenAI launched ChatGPT as a free research preview powered by a fine-tuned model from the GPT-3.5 series. Designed to maintain context across multi-turn dialogue, the interface gained more than one million users within days. In March 2023, OpenAI introduced theChatGPT application programming interface (API) for third-party developers and released GPT-4, a larger multimodal model capable of accepting both image and text inputs. In May 2024, OpenAI launched GPT-4o ("omni"), its first single model trained end-to-end across text, voice, and vision, reducing response latency and expanding the context window to 128,000 tokens, followed by a smaller variant, GPT-4o mini, in July 2024.

In late 2024, OpenAI bifurcated its model development by introducing the "o-series" of reasoning models, beginning with o1-preview and o1-mini on September 12, 2024, and the full o1 model in December 2024.²¹ Unlike earlier models that generated tokens immediately, the o-series models were trained to generate an internal chain of thought before responding, improving performance in mathematics, logic, and computer programming. This track continued with o3-mini in January 2025, followed by o3 and o4-mini in April 2025, while the traditional text track produced the developer-focused GPT-4.1 and the transitional research preview GPT-4.5 (internally codenamed Orion) in early 2025.²²

On August 7, 2025, OpenAI unified its conversational and reasoning tracks with the launch of GPT-5, which replaced GPT-4o as the default model for both free and paid users. GPT-5 employed a real-time router to direct prompts either to a fast conversational model or to a deeper reasoning model depending on task complexity. The fifth generation progressed through several iterative updates: GPT-5.1 in November 2025, GPT-5.2 in December 2025, GPT-5.3-Codex and GPT-5.3 Instant in early 2026, GPT-5.4 in March 2026, GPT-5.5 in April 2026, and GPT-5.6 in June and July 2026, which introduced the Luna, Terra, and Sol capability tiers. On September 3, 2026, OpenAI debuted its sixth-generation architecture with GPT-6 Astra, designed for autonomous computer operation and complex software engineering, followed later that month by the faster GPT-6 Sol and GPT-6 Luna variants and the specialized coding model GPT-6.1 Sol.

ChatGPT is built upon large-scale neural networks that utilize the transformer architecture to predict the probability of subsequent tokens—segments of words, characters, or code—based on preceding context.²³ Text inputs are first processed using Byte Pair Encoding (BPE) tokenization, which converts natural language into numerical representations that the neural network can process. Through self-attention layers, the model weighs the relevance of every token in a prompt against every other token, allowing it to track grammatical structures, thematic consistency, and complex instructions across long documents and multi-turn conversations.

The training of ChatGPT's underlying models occurs in three primary stages, beginning with unsupervised pre-training.²⁴ During this phase, the foundation model processes vast datasets of text, images, and audio collected from the public internet, licensed third-party databases, and books.²⁵ Executed on massive clusters of graphics processing units (GPUs), pre-training teaches the model syntax, factual associations, world knowledge, and reasoning patterns by repeatedly tasking it with predicting the next token in a sequence.²⁶ However, a raw pre-trained model merely completes text patterns and may output incoherent, unhelpful, or harmful statements rather than answering a user's question directly.

To transform the base model into a conversational assistant, OpenAI applies Supervised Fine-Tuning (SFT) followed by Reinforcement Learning from Human Feedback (RLHF). In the SFT stage, human AI trainers act as both the user and the ideal assistant, creating curated datasets of prompts and high-quality responses that teach the model conversational formatting. In the RLHF stage, human evaluators rank multiple different outputs generated by the model for the same prompt from best to worst.²⁷ These rankings are used to train a separate reward model, which then guides the primary language model via reinforcement learning optimization algorithms to maximize helpfulness, factual accuracy, and adherence to safety boundaries. For reasoning models beginning with the o1 series and unified in GPT-5 and GPT-6, reinforcement learning is also applied to internal chain-of-thought sequences, training the model to break down complex problems, recognize errors in its own logic, and evaluate alternative solutions before generating a final visible response.

Accessible through its web domain at chatgpt.com and native applications for iOS, Android, macOS, and Windows, ChatGPT supports interaction in 59 languages. Beyond standard text generation, summarization, and translation²⁸, the platform integrates a suite of specialized tools.²⁹ Its web browsing capability, initially developed in partnership with Microsoft Bing in May 2023 and later expanded into SearchGPT, enables the chatbot to query the live internet and provide responses citing current websites. For more intensive inquiries, a feature named Deep Research conducts multi-step, autonomous web searches to compile comprehensive, citation-backed reports.³⁰³¹

ChatGPT's multimodal capabilities allow users to converse with the system in real time using Voice Mode, which processes vocal inflections and responds with synthesized speech, and to upload images, spreadsheets, and documents for visual recognition and data analysis. For visual creation, OpenAI integrated its DALL-E 3 model into ChatGPT in October 2023. In March 2025, OpenAI replaced DALL-E in the interface with GPT Image, which improved the rendering of legible text inside generated pictures, followed by ChatGPT Images 2.0 ahead of the formal retirement of the DALL-E 2 and DALL-E 3 models in May 2026.

For software development and writing workflows, ChatGPT includes an execution environment originally launched as Code Interpreter (later Advanced Data Analysis), which writes and runs Python code in a sandboxed environment to process files and generate charts. Users can also open Canvas, a split-screen workspace that displays the conversation on the left alongside a live, editable document or code editor on the right. With the introduction of the GPT-5.3, GPT-5.6, and GPT-6 Astra models in 2026, ChatGPT gained agentic "computer use" capabilities, allowing the software to autonomously operate computer interfaces, navigate web browsers, and execute multi-step software engineering tasks with minimal human intervention.

Personalization and third-party extensibility are managed through Memory, Custom GPTs, and plugins. When enabled by the user, the Memory function retains biographical details, stylistic preferences, and project context across separate chat sessions. In November 2023, OpenAI released GPT Builder, allowing subscribers to create custom versions of ChatGPT tailored with specific instructions and knowledge files, and opened the GPT Store marketplace in January 2024 with over three million user-created assistants. In September 2025, OpenAI added support for the Model Context Protocol (MCP) to connect ChatGPT with external servers and third-party software applications, renaming its integrated app ecosystem to plugins in July 2026.

OpenAI distributes ChatGPT through a tiered freemium business model. Unregistered visitors and free-tier account holders receive access to standard models with caps on message volume and advanced tool usage.³² Paid tiers began on February 1, 2023, with the launch of ChatGPT Plus at $20 per month, offering priority access and higher usage limits. OpenAI subsequently expanded its commercial offerings with ChatGPT Team (priced at $25 to $30 per user per month for collaborative workspaces), ChatGPT Enterprise for large corporations, ChatGPT Edu for universities, and ChatGPT Pro, a $200-per-month tier introduced for researchers and power users requiring maximum compute allocation and unrestricted access to frontier reasoning models.

User adoption of ChatGPT has outpaced all prior consumer internet applications. After reaching 100 million monthly active users in January 2023, its user base climbed to 400 million weekly active users in February 2025, 700 million in August 2025, and 900 million in February 2026. In June 2026, the ChatGPT mobile application surpassed 1 billion monthly active users, and by August 2026, total weekly active users across all platforms reached 1 billion, processing more than 2.5 billion prompts per day. A 2025 study by OpenAI and the National Bureau of Economic Research found that 73 percent of consumer messages were non-work-related, with nearly 80 percent of all interactions falling into the categories of practical guidance, information seeking, or writing. In the corporate sector, 92 percent of Fortune 500 companies used ChatGPT by 2026, accounting for more than 7 million active enterprise workplace seats and 9 million paying business users across over 1 million corporate customers.

This adoption drove rapid revenue growth and capital investment for OpenAI. Having ended 2023 with approximately $2 billion in annualized revenue and 2024 with $6 billion, OpenAI surpassed $20 billion in annualized revenue by the end of 2025 and topped $25 billion in February 2026, with Bloomberg reporting an annualized run rate exceeding $40 billion by August 2026. By early 2026, ChatGPT had more than 50 million paying consumer subscribers, while a United States advertising pilot for free users generated over $100 million in annualized revenue within six weeks of launch. In March 2026, OpenAI closed a major funding round, with financial estimates placing the raised capital between $110 billion and $122 billion and the company's post-money valuation between $730 billion and $852 billion. Despite its market leadership, intense competition from rivals—primarily Google's Gemini and Anthropic's Claude—reduced ChatGPT's share of global AI chatbot web traffic from 79.0 percent in May 2025 to 53.9 percent in May 2026, compared to 27.9 percent for Gemini and 9.2 percent for Claude.

The arrival of ChatGPT prompted widespread reassessment of white-collar labor, education, and professional standards. In primary, secondary, and higher education, the chatbot's ability to compose essays and solve homework assignments raised immediate concerns over plagiarism and cognitive offloading.³³ In a January 2023 survey of 1,000 college students, 30 percent reported having used ChatGPT for schoolwork, and several major public school districts temporarily blocked the service on campus networks. Over subsequent academic years, many educational institutions shifted from outright bans toward revising assessment methods, incorporating oral examinations, and establishing guidelines for AI-assisted tutoring and writing.

In scientific and medical publishing, academic bodies moved to regulate the tool's role in research.³⁴ The International Committee of Medical Journal Editors (ICMJE) and major scientific journals established rules prohibiting ChatGPT and other large language models from being listed as co-authors on scholarly papers, ruling that an artificial intelligence system cannot be held accountable for the accuracy, integrity, or originality of a study. Researchers noted that while ChatGPT aids in literature summarization and drafting manuscripts, its use in medicine carries risks of fabricating bibliographic citations, misinterpreting clinical data, and eroding clinical reasoning skills if relied upon without human verification.

Across the broader economy, ChatGPT accelerated the shift from individual task assistance to automated enterprise workflows. Software engineers adopted its coding models to generate boilerplate code, debug programs, and migrate legacy systems, while marketing, legal, and financial firms integrated the platform for contract review, copywriting, and data analysis. Economists and labor analysts have debated the long-term displacement effects of the technology, noting that while generative AI increases productivity for individual knowledge workers, its expanding agentic capabilities—where models execute multi-step software tasks autonomously—increasingly substitute for entry-level administrative, customer support, and programming roles.