GPT (OpenAI)

GPT (Generative Pre-trained Transformer) is a family of large language models (LLMs) developed by OpenAI. GPT models are built on the transformer architecture and implement the generative pre-training paradigm: in the first stage, the model is trained on extensive text corpora without explicit labeling, and can then be fine-tuned for specific tasks. For later generations (starting with GPT‑5), OpenAI also uses the term unified system, as the product combines a fast response mode, a deep reasoning mode, and a router.

Name

The abbreviation GPT stands for Generative Pre-trained Transformer.

• Generative: indicates that the model is capable of creating (generating) new content, such as text.
• Pre-trained: indicates that the model undergoes an extensive initial training stage on a large dataset (e.g., texts from the internet). After pre-training, the model can often be additionally "fine-tuned" for more specific tasks.
• Transformer: refers to a specific neural network architecture that is a key innovation underlying GPT and many other modern AI models.

The main characteristic of GPT is that training occurs in an autoregressive manner — the model predicts the next token based on the preceding context. That is, the model is trained to maximize the probability of the next token given a sequence of previous tokens. During training, the prediction error for the next element is minimized, which enables the generation of texts with high coherence and consistency.

Text generation process in GPT

The GPT model generates text sequentially, token by token, according to the following iterative scheme:

• Receives an initial text sequence (prompt, seed text) as input.
• Computes a probability distribution over all tokens in the vocabulary for the next text element.
• Selects the next token:
  • either by the highest probability (greedy selection),
  • or by stochastic sampling (sampling),
  • or using special filtering strategies (top-k, top-p).
• Appends the selected token to the current sequence.
• The updated sequence is again fed into the model for predicting the next token.

Transformer architecture: text processing

The data processing inside the transformer for predicting the next token involves several main stages:

• Tokenization (Tokenization). The input text is split into tokens — small text units that can be words, subwords, or punctuation marks. In GPT-3, for example, the vocabulary includes approximately 50,257 tokens.
• Token embeddings (Embeddings). Each token is converted into a fixed-length vector using an embedding matrix (W_E). The vectors encode token meanings: semantically similar tokens are located close together in the high-dimensional space. In GPT-3, the embedding dimensionality is 12,288.
• Processing in transformer layers.
  • Attention blocks: Each token interacts with other tokens in the sequence. The attention mechanism allows the model to account for context and correctly interpret word meanings.
  • Feed-forward layers: After attention, each token is processed individually through a two-layer neural network with nonlinear activation.
• Reverse transformation and Softmax. After all layers, the processed vector is transformed back into token space using a matrix (W_U), which is often a transposed version of W_E. The resulting logits vector is normalized using the Softmax function to obtain a probability distribution over all tokens.
• Next token selection (Sampling). The next token is selected based on the probability distribution. The temperature parameter controls the randomness of selection: at temperature 0, the most probable token is selected; at higher temperatures, the probability of selecting less likely options increases, which promotes greater text diversity.

GPT models

• GPT-1 (2018): the first model in the family; a 12-layer decoder-only transformer; two-stage training (pre-training + fine-tuning on NLP tasks).
• GPT-2 (2019): 1.5 billion parameters; trained on the WebText corpus; the first model capable of generating long coherent texts; improved zero-shot generation quality. Announced on February 14, 2019; the full version (1.5B) was released on November 5, 2019 due to safety concerns.
• GPT-3 (2020): 175 billion parameters; large-scale training on a combination of Common Crawl, Books, and Wikipedia; strong development of few-shot and zero-shot capabilities.
• GPT-3.5 (2022): an intermediate version between GPT-3 and GPT-4; improved instruction following through Reinforcement Learning from Human Feedback (RLHF) in the text-davinci-003 and gpt-3.5-turbo versions; context window up to 4,096 tokens in early versions and up to 16,385 tokens in later ones (gpt-3.5-turbo-16k and updated gpt-3.5-turbo).
• GPT-4 (2023): a multimodal model with text and image input (image support was deployed later, after the text-only launch); context window of 8,192 tokens in the base version and 32,768 tokens in the GPT-4-32k variant; significant improvements in accuracy, robustness, and reasoning.
• GPT-4 Turbo (2023): an optimized version of GPT-4; increased context window up to 128,000 tokens; lower latency and cost.
• GPT-4o (2024): a next-generation multimodal model (text, image, audio) with a unified neural network architecture; very high response speed and accuracy; context window of 128,000 tokens.
• GPT-4.5 (2025): a research preview; the OpenAI system card states that the model "builds on GPT-4o"; improved understanding of user queries, reduced error rate; context window of 128,000 tokens. The API model gpt-4.5-preview was declared deprecated on April 14, 2025 and shut down on July 14, 2025.
• GPT-4.1 (2025): an improved version of the GPT-4 family with a context window of up to 1 million tokens; accepts text and images as input, outputs text. Released simultaneously in three variants: GPT-4.1, GPT-4.1 mini, GPT-4.1 nano.
• GPT-5 (2025): a unified system with fast response and deep reasoning modes; context window of approximately 400,000 tokens; notable reduction in hallucinations on factual tasks.
• GPT-5.1 (2025): adaptive reasoning, improvements in coding and long-context retention.
• GPT-5.2 (2025): focus on professional work; Pro mode for frontier tasks; the agentic GPT-5.2-Codex was released based on GPT-5.2.
• GPT-5.3-Codex (2026): an agentic coding model combining coding capabilities and reasoning; 25% faster than predecessors.
• GPT-5.3 Instant (2026): an update to the most widely used conversational model in ChatGPT; released on March 3, 2026. Improved factual accuracy, web search quality, conversational flow, and reduced excessive refusals and unnecessary caveats. Available in the API as gpt-5.3-chat-latest.
• GPT-5.4 (2026): OpenAI's frontier model for professional work, introduced on March 5, 2026; the first general-purpose OpenAI model with native computer-use capabilities. In the API, gpt-5.4 is recommended as the default model for a wide range of general-purpose and coding tasks.

GPT-1

The first model, GPT-1, was introduced by OpenAI in 2018 in the paper "Improving Language Understanding by Generative Pre-Training". The model was a 12-layer decoder-only transformer built on the transformer architecture. GPT-1 training proceeded in two stages: an unsupervised generative pre-training stage (pre-training), followed by a supervised fine-tuning stage (fine-tuning).

During the pre-training stage, the model was trained on the BookCorpus, comprising over 7,000 unpublished books of various genres. A distinctive feature of this corpus was the presence of long continuous text passages, which was critically important for developing the model's ability to process complex and long-range textual dependencies.

During the fine-tuning stage, the model was adapted to solve specialized natural language processing tasks, including:

• Question Answering (QA) — generating a correct answer based on a given textual context;
• Natural Language Inference (NLI) — determining the logical relationship between two texts: entailment, contradiction, or neutrality;
• Semantic Textual Similarity — measuring the degree of semantic closeness between two text sequences.

Thanks to this approach, GPT-1 demonstrated significant superiority over previous models on a number of standard benchmarks for text comprehension tasks.

The development of GPT-1 demonstrated several key achievements and discoveries in natural language processing (NLP):

• Effectiveness of generative pre-training. It was empirically confirmed that pre-training on large corpora of unlabeled text enables the model to acquire universal language representations suitable for subsequent application in various tasks without requiring fundamental architectural changes.
• Versatility of the transformer architecture. The use of a multi-layer decoder transformer enabled the model to successfully process long-range dependencies in text, which had previously been difficult for models based on recurrent neural networks.
• Reduced dependence on labeled data. The work confirmed that large-scale pre-training on unlabeled data can significantly reduce the amount of labeled data needed to achieve high quality on target tasks.
• Foundation for further development. The results of GPT-1 laid the conceptual and technical groundwork for subsequent versions of the GPT family (GPT-2, GPT-3, and beyond).

GPT-2

The GPT-2 model was announced by OpenAI on February 14, 2019. It significantly surpassed its predecessor in size: the full version of the model contained approximately 1.5 billion parameters. For safety reasons, OpenAI initially released only smaller variants of the model; the full version (1.5B parameters) was released on November 5, 2019. Unlike GPT-1, which was trained on the BookCorpus (~5 GB), GPT-2 was trained on a specially compiled WebText corpus of approximately 40 GB, comprising textual data from high-quality internet sources. The increase in both model size and training data volume enabled GPT-2 to significantly improve text generation quality: it demonstrated the ability to create substantive articles, stories, and even coherent passages of fiction.

GPT-2 employed an autoregressive decoder-only transformer architecture similar to GPT-1, without significant changes. The model consisted of 48 self-attention layers, had a hidden state size of 1,600, and included approximately 1.5 billion parameters. The number of attention heads was 25 (maintaining a head size of 64, inherited from GPT-1: 1,600 ÷ 64 = 25). Training was performed on the next-token prediction task based on the preceding context using masked attention.

One of the main distinctions of GPT-2 was that the model was the first to demonstrate high effectiveness in zero-shot learning — the ability to solve new tasks without undergoing explicit fine-tuning on examples for those tasks. The model was trained on a large corpus of general texts and did not undergo specialized training on task-specific data. Evaluation was conducted in a zero-shot regime, in which the model performed tasks solely based on knowledge acquired during pre-training. On a number of language modeling tasks, GPT-2 achieved quality comparable to or exceeding the results of models specifically trained on specialized datasets (e.g., Wikipedia, news texts, books).

GPT-3

The GPT-3 model was introduced by OpenAI in June 2020 (the arXiv paper appeared on May 28, 2020; beta API access opened on June 11, 2020). It was the next step in the development of generative transformers after GPT-2 and was distinguished by scaling the architecture to 175 billion parameters, making it the largest language model at the time.

The architecture of GPT-3 remained fundamentally the same — a multi-layer autoregressive decoder-only transformer without radical changes. The main performance improvements were achieved through increasing the number of layers, the width of hidden layers, and the scale of training. The model was trained on a combination of several large text corpora, including Common Crawl, WebText2, Books1, Books2, and Wikipedia. The total data volume was approximately 570 GB or more (570 GB accounted for the filtered portion of Common Crawl, which dominated the training mixture).

One of the main features of GPT-3 was its capability for few-shot learning and zero-shot learning: the model could perform a wide range of natural language processing tasks, including translation, summarization, question answering, essay writing, and even programming, based on just a few examples in the text prompt or none at all.

GPT-3.5

The GPT-3.5 model was introduced by OpenAI in late 2022 as part of the evolutionary development of the GPT family. It was built on the scaled autoregressive decoder-only transformer architecture used in GPT-3, with improvements in text generation quality, context processing, and the ability to follow complex instructions. The exact number of parameters in GPT-3.5 was not officially disclosed; the davinci versions are presumably comparable in size to GPT-3 (175B), but the exact parameters of the gpt-3.5-turbo version are unknown.

The training of GPT-3.5 involved expanded use of Reinforcement Learning from Human Feedback (RLHF) methods in the text-davinci-003 and gpt-3.5-turbo versions. The earlier text-davinci-002 version was trained using supervised fine-tuning (SFT) rather than RLHF. The model was trained on expanded text corpora including Common Crawl, Books, WebText, and other high-quality sources. The context window in early popular versions (gpt-3.5-turbo) was 4,096 tokens; OpenAI subsequently released updated versions with a context of up to 16,385 tokens.

In practice, GPT-3.5 was adapted to solve a wide range of natural language processing tasks, such as:

• Generating coherent and logical text;
• Question answering (QA) and context understanding;
• Following multi-step instructions;
• Improved long-term context maintenance in dialogues.

Several key versions based on GPT-3.5 were released for different purposes:

• text-davinci-002 — the first publicly available model based on GPT-3.5, optimized for generation and instruction following (trained using SFT).
• text-davinci-003 — an improved version with even greater reasoning and complex text generation capability (trained using RLHF).
• gpt-3.5-turbo — the most performant and cost-effective version of GPT-3.5, used in the ChatGPT service since late 2022.

GPT-4

The GPT-4 model was introduced by OpenAI on March 14, 2023 in the "GPT-4 Technical Report". It represented the next stage in the development of the language model family, offering significant improvements in text comprehension, generation of meaningful and creative responses, and processing of multimodal data. The exact number of parameters and architectural details of the model were not officially disclosed — the GPT-4 technical report explicitly states that information about the architecture, model size, hardware, training compute, and dataset construction is not published. According to unofficial external estimates, GPT-4 may have used a Mixture of Experts (MoE) approach with a total scale on the order of ~1.8 trillion parameters; however, OpenAI has neither officially confirmed nor denied these figures.

GPT-4 is a multimodal model capable of accepting both text and images as input. It should be noted that at the time of the initial launch in March 2023, only the text modality was available; image input support was deployed later. The context window was 8,192 tokens in the base version and 32,768 tokens in the GPT-4-32k variant. The model used RLHF (Reinforcement Learning from Human Feedback) methods.

GPT-4 training was performed on a combination of large-scale textual and multimodal corpora. Specific details of the training data, hardware, and methodology are not disclosed in official OpenAI publications.

Training proceeded in several stages:

• large-scale unsupervised pre-training on texts and images,
• supervised fine-tuning on specialized tasks,
• a final stage of Reinforcement Learning from Human Feedback (RLHF) to improve reliability, safety, and instruction interpretation quality.

Several main versions were released based on GPT-4:

• GPT-4 (March 2023): the base version with text input support (image support added later); context window of 8,192 tokens; a GPT-4-32k variant with a 32,768-token context was also released.
• GPT-4 Turbo (November 2023): an optimized modification of GPT-4 with an increased context window of up to 128,000 tokens; reduced compute costs and accelerated generation; support for function calling and JSON output modes.
• GPT-4o (May 2024): a next-generation multimodal version; in the launch announcement it was positioned as an omni-model capable of working with text, images, and audio in real time (unlike GPT-4 Turbo, where different modalities were served by separate modules); however, the base API model gpt-4o is described as text+image input, text output; context window of 128,000 tokens.
• GPT-4.5 (February 2025): a research preview; the OpenAI system card explicitly states that the model "builds on GPT-4o"; improved generation of complex texts, increased instruction-following accuracy, and reduced hallucination rate; context window of 128,000 tokens. It was described as "the last OpenAI model without chain-of-thought" (codename — Orion). The API model gpt-4.5-preview was declared deprecated on April 14, 2025 and shut down on July 14, 2025.
• GPT-4.1 (April 2025): a stable version with a radical context expansion to 1,047,576 tokens; accepts text and images as input, outputs text; released simultaneously in three variants (GPT-4.1, GPT-4.1 mini, GPT-4.1 nano); initially available only through the API, later deployed in ChatGPT.

GPT-5

On August 7, 2025, OpenAI introduced GPT‑5 as its then "smartest, fastest, and most useful" model, with a built-in deep reasoning mode (thinking) and a focus on practical scenarios — writing, programming, health-related work, and multimodal understanding. GPT‑5 gradually became the default model for most logged-in ChatGPT users, displacing the previously used GPT‑4/4o family and o-series models.

GPT‑5 is implemented as a unified system with two main operating modes: fast, cost-effective responses for everyday queries (referred to as gpt‑5 main) and deep reasoning for complex tasks (referred to as gpt‑5 thinking). The mode is selected automatically by a router that takes into account the dialogue type, query complexity, the need for tools, and explicit user cues (e.g., "think step by step" or "analyze in depth"). In ChatGPT, users have access to Auto / Instant / Thinking / Pro modes; the mini and nano variants are primarily API models, and mini in the consumer product may be used as a fallback after rate limits are exhausted.

Several sizes and configurations of GPT‑5 are available through the API; in the OpenAI documentation, the main variants are listed as gpt‑5, gpt‑5‑mini, and gpt‑5‑nano (all supporting text and visual data). The maximum total context window for the GPT‑5 family in the API is approximately 400,000 tokens (with separate budgets for input and reasoning/output), although specific limits may vary depending on the chosen model variant and product.

On a number of web-search and factual benchmarks, GPT‑5 demonstrates a notable reduction in the frequency of hallucinations and errors compared to GPT‑4o and earlier OpenAI "thinking" models. In the official announcement, OpenAI reported error reductions of approximately 45% compared to GPT-4o and approximately 80% compared to o3 in thinking mode — these results were obtained under specific conditions: with web search enabled on anonymized prompts representative of ChatGPT production traffic.

GPT-5.1

The GPT-5.1 model was introduced by OpenAI on November 12, 2025 as the first significant iteration after the base GPT-5, aimed at improving everyday interaction, conversational quality, and adaptability. The model retains the unified system with a fast mode (GPT-5.1 Instant) and deep reasoning (GPT-5.1 Thinking), but introduces adaptive reasoning: the model dynamically determines the amount of computation depending on query complexity, making it notably faster on simple tasks without sacrificing quality on complex ones.

GPT-5.1 training was built on top of GPT-5 with an additional post-training stage that included expanded RLHF, a focus on natural tone, and reduced "coldness" of responses. The API context window is 400,000 tokens, with a maximum output of 128,000 tokens. Extended prompt caching of up to 24 hours was introduced, significantly reducing cost and latency for multi-turn dialogues.

Key features:

• GPT-5.1 Instant — the primary mode for everyday tasks; the first to use adaptive reasoning to determine when it is worth "thinking" before responding to a more complex query.
• GPT-5.1 Thinking — adaptive allocation of reasoning time; according to OpenAI, on a representative distribution of ChatGPT tasks the model is approximately twice as fast on the simplest tasks and approximately twice as slow on the hardest compared to GPT-5 Thinking.
• Improved multimodality (text + vision).
• Improved coding and agentic scenarios, as well as efficiency on simple tasks through adaptive reasoning and extended prompt caching.

GPT-5.2

The GPT-5.2 model was released on December 11, 2025 as "the most capable model in the series for professional work and learning." It is an evolution of GPT-5.1 with an emphasis on economic value: generation of tables, presentations, complex code, and end-to-end tasks. It retains the unified architecture with Instant, Thinking, and a new Pro mode (for tasks requiring maximum compute and reasoning time).

Training included an updated corpus with a knowledge cutoff of August 2025, enhanced instruction-tuning, and RLHF to reduce errors in multi-step scenarios. Context window — 400K tokens (128K max output). The model became more reliable in professional scenarios, with improved factual accuracy and tool use.

On December 18, 2025, the specialized GPT-5.2-Codex was released based on GPT-5.2 — an agentic coding model with improved context compaction, Windows support, enhanced cybersecurity, and long-horizon reasoning (tasks lasting up to several hours).

As of February 13, 2026, following the retirement of several older models, GPT-5.2 temporarily became the default model in ChatGPT. However, by early March 2026, this role was taken over by GPT‑5.3 Instant and GPT‑5.4.

GPT-5.3-Codex

GPT-5.3-Codex was introduced on February 5, 2026 as "the most powerful agentic coding model to date." It combines the frontier-coding capabilities of GPT-5.2-Codex with the professional reasoning of GPT-5.2 in a single model that is 25% faster than its predecessors.

The model is capable of performing virtually any developer task: long-running workflows, research, tool use, code execution, and interactive steering (the user can intervene in real time without losing context). Early versions of the model were used by the OpenAI team to debug their own training, deployment, and evaluations.

Key results at the time of the February 5, 2026 announcement: Terminal-Bench ~77.3%, OSWorld-Verified ~64.7%, SWE-Bench Pro ~56.8%. In the later GPT-5.4 release on March 5, 2026, OpenAI reported an updated OSWorld-Verified result of 74.0% for GPT-5.3-Codex when using a new API parameter that preserves the original image resolution.

On February 12, 2026, OpenAI also released GPT-5.3-Codex-Spark — a compact ultra-fast version in partnership with Cerebras, optimized for real-time use: over 1,000 tokens per second, text-only, 128K context. At launch, this was a rollout for ChatGPT Pro users in Codex and a small number of API design partners, rather than a broadly available API model.

GPT-5.4

On March 5, 2026, OpenAI introduced GPT‑5.4 as its new frontier model for professional work. GPT‑5.4 combines the strengths of OpenAI's latest releases in reasoning, coding, and agentic workflows and was the first in the main product line to receive built-in computer use capabilities. Simultaneously, OpenAI released GPT‑5.4 Pro — a variant for the most complex tasks, using more compute and longer reasoning.

In the API, the gpt-5.4 model is described as the recommended default for a wide range of general-purpose and coding tasks; the context window is 1,050,000 tokens, with a maximum output of 128,000 tokens. The model accepts text and images as input and outputs text.

In ChatGPT, the Auto mode as of March 7, 2026 automatically switches between GPT‑5.3 Instant and GPT‑5.4 Thinking, while GPT‑5.4 Pro is available as a separate high-capability mode. For logged-in ChatGPT users, the default model is GPT‑5.3.

GPT-5.5

On April 23, 2026, OpenAI introduced GPT-5.5 as its new frontier model for professional work and described it as a new class of intelligence for real work. GPT-5.5 was positioned as a step forward in agentic coding, computer use, knowledge work, and early scientific research, with stronger ability to plan, use tools, check its work, and carry out multi-step tasks across software and documents. OpenAI also released GPT-5.5 Pro — a higher-compute variant for harder questions and higher-accuracy work. On April 24, 2026, both GPT-5.5 and GPT-5.5 Pro became available in the API.

In the API, the gpt-5.5 model is described as OpenAI's newest frontier model for the most complex professional work. The context window is 1,050,000 tokens, with a maximum output of 128,000 tokens. The model accepts text and images as input and outputs text; reasoning.effort supports none, low, medium by default, high, and xhigh. The gpt-5.5-pro model uses more compute to think harder and provide more consistently precise answers; it is available for Responses API requests, including through the Batch API, and has the same 1,050,000-token context window and 128,000-token maximum output.

In ChatGPT, GPT-5.5 Thinking was made available to Plus, Pro, Business, and Enterprise users, while GPT-5.5 Pro was made available to Pro, Business, and Enterprise users. In Codex, GPT-5.5 was made available to Plus, Pro, Business, Enterprise, Edu, and Go plans with a 400K context window. On May 5, 2026, OpenAI introduced GPT-5.5 Instant as the updated default ChatGPT model, replacing GPT-5.3 Instant; in the API, the corresponding latest Instant model is exposed through chat-latest.

GPT-5.6

On June 26, 2026, OpenAI began a limited preview of the GPT-5.6 family, comprising GPT-5.6 Sol, GPT-5.6 Terra, and GPT-5.6 Luna. On July 9, 2026, the family entered general availability across ChatGPT, Codex, and the OpenAI API. Sol was introduced as the flagship model for complex professional work, Terra as a balanced lower-cost model for everyday workloads, and Luna as the fastest and most cost-efficient member of the family. OpenAI positioned GPT-5.6 as an advance in coding, long-horizon planning, computer use, professional knowledge work, design, cybersecurity, and scientific research, with improved performance per token and stronger ability to complete multi-step tasks using tools.

GPT-5.6 introduced max as a reasoning-effort setting above xhigh, allowing the models to spend additional time exploring alternatives, checking results, and revising their approach. OpenAI also introduced ultra, a higher-capability multi-agent mode that coordinates four agents in parallel by default. In the Responses API, Programmatic Tool Calling allows GPT-5.6 to write and execute lightweight in-memory programs that coordinate tools and process intermediate results, while a beta multi-agent capability allows the model to run concurrent subagents and synthesize their work into a single response. OpenAI stated that GPT-5.6 was released with its most extensive safety evaluation and safeguard system to date.

In the API, gpt-5.6-sol is the frontier model for complex professional work, and the gpt-5.6 alias routes requests to Sol. gpt-5.6-terra is intended to balance intelligence and cost and broadly occupies the position of the mini tier used in earlier GPT-5 families, while gpt-5.6-luna is optimized for cost-sensitive, high-volume workloads and broadly corresponds to the earlier nano tier. All three models have a 1,050,000-token context window, a maximum output of 128,000 tokens, and a knowledge cutoff of February 16, 2026. They accept text and images as input, output text, and support reasoning.effort values of none, low, medium, high, xhigh, and max.

In ChatGPT, Plus, Pro, Business, and Enterprise users received access to GPT-5.6 Sol through medium and higher reasoning-effort settings, while Pro and Enterprise users could also select GPT-5.6 Sol Pro for higher-quality results on difficult tasks. Sol Pro is a higher-compute product mode rather than a separately documented API model. In ChatGPT Work and Codex, Free and Go users received access to GPT-5.6 Terra, while Plus, Pro, Business, and Enterprise users could choose between Sol, Terra, and Luna. The max setting was made available to users with GPT-5.6 access in ChatGPT Work and Codex; ultra was made available to Pro and Enterprise users in ChatGPT Work and to Plus and higher plans in Codex.

GPT-6

On September 3, 2026, OpenAI introduced GPT-6 Astra, the first publicly announced model of the GPT-6 generation, and began a limited rollout to selected organizations. OpenAI described Astra as its most capable and most aligned model to date and positioned it as the new flagship for the hardest end-to-end work. The model was introduced with advances in computer use, browsing, software engineering, coding, cybersecurity, scientific research, and professional knowledge work, with particular emphasis on completing complex multi-step workflows across browsers, code, professional software, and document-creation tools. Astra can create and edit documents, spreadsheets, presentations, websites, and other professional artifacts while preserving templates and incorporating new instructions without losing track of the broader task.

GPT-6 Astra introduced several capabilities intended for longer-running agentic workflows. Async tool calling allows the model to continue reasoning, invoke other tools, or work on independent parts of a task while an application is executing a tool call. Mid-turn steering allows additional user instructions to be incorporated while a response is already in progress, preserving completed work rather than restarting the task. A new configuration_update mechanism allows reasoning effort to be changed during a conversation while preserving the existing prompt cache. Astra retains capabilities introduced with GPT-5.6, including Programmatic Tool Calling, multi-agent orchestration, computer use, Structured Outputs, streaming, persisted reasoning, prompt caching, compaction, and pro mode. Unlike GPT-5.6, Astra does not support the none reasoning-effort setting; supported reasoning.effort values are low, medium, high, xhigh, and max.

In the API, gpt-6-astra is the flagship GPT-6 model for complex reasoning, coding, computer use, research, and document creation. It has a 1,050,000-token context window, a maximum output of 128,000 tokens, and a knowledge cutoff of April 30, 2026. Astra accepts text and images as input and outputs text, and supports streaming, function calling, Structured Outputs, reasoning tokens, and tools including web search, file search, and computer use. Standard API pricing at launch was $10 per million input tokens, $1 per million cached input tokens, and $50 per million output tokens; prompts exceeding 272,000 input tokens use higher long-context rates. Astra is available through the Responses API and Chat Completions, although OpenAI recommends the Responses API for tool-using applications.

At launch, GPT-6 Astra began rolling out to a limited set of organizations, with broader access announced for ChatGPT Plus, Pro, Business, and Enterprise users and through the OpenAI API, Microsoft Azure, and AWS Bedrock over the following days. Users on Pro, Business, and Enterprise plans were also announced to receive a higher-compute GPT-6 Astra Pro mode; OpenAI's subsequent ChatGPT documentation refers to this product mode as GPT-6 Pro, powered by GPT-6 Astra. Plus users receive Astra in ChatGPT Work and Codex as the rollout reaches their accounts, while availability can differ between Chat, Work, and Codex during the staged deployment. Enterprise administrators control whether Astra is enabled for their workspace.

GPT-6 Astra was also the first OpenAI model classified at the Critical level for cybersecurity capability under the company's Preparedness Framework. OpenAI reported that, with appropriate tools and access, the model can identify previously unknown vulnerabilities and develop new exploitation techniques across hardened systems. The deployment therefore introduced strengthened cybersecurity safeguards and misalignment monitoring, including monitoring of agent trajectories and additional checks that can pause or stop tasks when potentially unauthorized behavior is detected. OpenAI stated that Astra showed substantially improved adherence to authorized task boundaries compared with GPT-5.6 Sol in its internal alignment evaluations.

GPT model evolution

The table covers the principal general-purpose GPT language-model lineage and its officially named size, reasoning and product variants. Voice, image-generation, embedding, moderation and search-only models are outside its scope.

GPT model evolution

Model / variant	Initial release	Parameter count	Publicly disclosed training data or knowledge cutoff	Context window	Key features and status
GPT-1	June 2018	≈117M	BooksCorpus: more than 7,000 unpublished books	512 tokens	Decoder-only Transformer; generative pre-training followed by supervised task-specific fine-tuning
GPT-2	14 February 2019	1.5B in the largest version	WebText: approximately 40 GB from about 8 million web pages	1,024 tokens	Substantially improved coherent text generation and zero-shot behavior; initially released in stages because of misuse concerns
GPT-3	28 May 2020	175B	Common Crawl, WebText2, Books1, Books2 and Wikipedia; approximately 300B tokens sampled during training	2,048 tokens	Large-scale in-context learning; strong few-shot and zero-shot capabilities without gradient-based task fine-tuning
GPT-3.5	2022	Not disclosed	Not disclosed	Approximately 4K–16K tokens, depending on the model version	Instruction tuning and reinforcement learning from human feedback; improved dialogue and instruction following; foundation of the original ChatGPT
GPT-3.5 Turbo	1 March 2023	Not disclosed	Not disclosed	Initially 4,096 tokens; later versions up to 16,385 tokens	Lower-cost chat-optimized GPT-3.5 model; later made available for fine-tuning; subsequently deprecated
GPT-4	14 March 2023	Not disclosed	Not disclosed	8,192 or 32,768 tokens	Multimodal text-and-image input; improved reasoning, reliability and instruction following
GPT-4 Turbo Preview	6 November 2023	Not disclosed	Knowledge through April 2023 at launch	128,000 tokens	Preview of the faster and less expensive GPT-4 Turbo family; substantially enlarged context window
GPT-4 Turbo	9 April 2024	Not disclosed	Knowledge cutoff: December 2023 for the final API snapshot	128,000 tokens	Production GPT-4 Turbo model with text and image input; lower cost than the original GPT-4
GPT-4o	13 May 2024	Not disclosed	Knowledge cutoff: 1 October 2023 for the principal API snapshot	128,000 tokens	End-to-end multimodal model spanning text, vision and audio; lower latency and improved real-time interaction
GPT-4o mini	18 July 2024	Not disclosed	Knowledge cutoff: 1 October 2023	128,000 tokens	Smaller and less expensive GPT-4o-family model intended for focused and high-volume tasks
GPT-4o Search Preview	March 2025	Not disclosed	Knowledge cutoff: 1 October 2023	128,000 tokens	Specialized GPT-4o variant trained to formulate and execute web searches through the Chat Completions API; subsequently deprecated
GPT-4o mini Search Preview	March 2025	Not disclosed	Knowledge cutoff: 1 October 2023	128,000 tokens	Lower-cost search-specialized GPT-4o mini variant; subsequently deprecated
GPT-4.5 Preview	27 February 2025	Not disclosed	Knowledge cutoff: 1 October 2023	128,000 tokens	Research preview emphasizing scaled unsupervised learning, broader knowledge, natural conversation and reduced hallucinations; later discontinued
GPT-4.1	14 April 2025	Not disclosed	Knowledge cutoff: 1 June 2024	1,047,576 tokens	Major improvements in coding, instruction following and tool calling; text and image input with text output
GPT-4.1 mini	14 April 2025	Not disclosed	Knowledge cutoff: 1 June 2024	1,047,576 tokens	Faster and less expensive GPT-4.1 variant retaining the approximately one-million-token context window
GPT-4.1 nano	14 April 2025	Not disclosed	Knowledge cutoff: 1 June 2024	1,047,576 tokens	Smallest and fastest GPT-4.1 variant, intended for classification, extraction, ranking and other high-volume tasks
GPT-5 Instant	7 August 2025	Not disclosed	Knowledge cutoff: 30 September 2024	128,000 tokens in the corresponding Chat API model	Fast conversational component of the original unified GPT-5 system; handled ordinary queries and could route harder requests to GPT-5 Thinking
GPT-5 Thinking	7 August 2025	Not disclosed	Knowledge cutoff: 30 September 2024	400,000 tokens	Reasoning component of the unified GPT-5 system; configurable reasoning effort, improved coding, tool use and factual reliability
GPT-5 mini	7 August 2025	Not disclosed	Knowledge cutoff: 31 May 2024	400,000 tokens	Faster and more cost-efficient GPT-5 variant for well-defined, latency-sensitive and high-volume workloads
GPT-5 nano	7 August 2025	Not disclosed	Knowledge cutoff: 31 May 2024	400,000 tokens	Fastest and least expensive original GPT-5 variant; optimized for summarization, classification and extraction
GPT-5 Pro	2025	Not disclosed	Knowledge cutoff: 30 September 2024	400,000 tokens	Higher-compute GPT-5 variant for difficult reasoning tasks; API version supports only high reasoning effort
GPT-5.1 Instant	12 November 2025	Not disclosed	Knowledge cutoff: 30 September 2024	128,000 tokens in the corresponding Chat API model	More conversational fast model with improved instruction following and adaptive reasoning
GPT-5.1 Thinking	12 November 2025	Not disclosed	Knowledge cutoff: 30 September 2024	400,000 tokens	Dynamically adjusted reasoning time according to task difficulty; clearer explanations, improved coding and reduced unnecessary reasoning
GPT-5.1 Pro	November 2025	Not disclosed	Not disclosed separately	Approximately 400,000 tokens	Higher-compute GPT-5.1 variant for difficult questions; retired from ChatGPT on 11 March 2026
GPT-5.2 Instant	11 December 2025	Not disclosed	Knowledge cutoff: 31 August 2025	128,000 tokens in the corresponding Chat API model	Fast conversational GPT-5.2 variant; improved everyday usefulness, information retrieval and instruction following
GPT-5.2 Thinking	11 December 2025	Not disclosed	Knowledge cutoff: 31 August 2025	400,000 tokens	Professional knowledge work, coding, long-context reasoning, vision and tool use; introduced the xhigh reasoning-effort setting
GPT-5.2 Pro	11 December 2025	Not disclosed	Knowledge cutoff: 31 August 2025	400,000 tokens	Higher-compute GPT-5.2 variant for complex questions where accuracy is prioritized over latency
GPT-5.3 Instant	3 March 2026	Not disclosed	Knowledge cutoff: 31 August 2025	128,000 tokens	Faster conversational model with improved factuality, web-search integration and conversational flow; later superseded by GPT-5.5 Instant
GPT-5.4	5 March 2026	Not disclosed	Knowledge cutoff: 31 August 2025	1,050,000 tokens	Frontier model for professional and agentic work; native computer use, stronger coding, multimodal reasoning and coordinated tool execution
GPT-5.4 Pro	5 March 2026	Not disclosed	Knowledge cutoff: 31 August 2025	1,050,000 tokens	Higher-compute GPT-5.4 variant producing more precise responses on difficult tasks
GPT-5.4 mini	17 March 2026	Not disclosed	Knowledge cutoff: 31 August 2025	400,000 tokens	Faster model for coding, computer use, multimodal workloads and sub-agent systems; supports text and image input and major API tools
GPT-5.4 nano	17 March 2026	Not disclosed	Knowledge cutoff: 31 August 2025	400,000 tokens	Lowest-cost GPT-5.4-class model for classification, extraction, ranking and high-volume sub-agent tasks
GPT-5.5	23 April 2026	Not disclosed	Knowledge cutoff: 1 December 2025	1,050,000 tokens	Frontier model for complex professional work, agentic coding, research, data analysis, document creation, computer use and multi-tool workflows
GPT-5.5 Pro	23 April 2026	Not disclosed	Knowledge cutoff: 1 December 2025	1,050,000 tokens	Higher-compute GPT-5.5 variant designed for increased accuracy on difficult and long-running tasks
GPT-5.5 Instant	5 May 2026	Not disclosed	Knowledge cutoff: 31 August 2025 for the corresponding dynamic API alias	400,000 tokens	Default conversational ChatGPT model after launch; improved factuality, concision, personalization, image understanding, STEM answers and web-search decisions
GPT-5.6 Sol	9 July 2026	Not disclosed	Knowledge cutoff: 16 February 2026	1,050,000 tokens	Flagship GPT-5.6 tier for complex professional work; stronger reasoning, coding, science, cybersecurity, computer use, design judgment and token efficiency
GPT-5.6 Sol Pro	9 July 2026	Not disclosed	Knowledge cutoff: 16 February 2026	1,050,000 tokens	Higher-compute product mode for the most difficult tasks; based on GPT-5.6 Sol rather than a separately disclosed base architecture
GPT-5.6 Terra	9 July 2026	Not disclosed	Knowledge cutoff: 16 February 2026	1,050,000 tokens	Balanced GPT-5.6 tier combining high capability with lower cost; broadly corresponds to the mini position in earlier GPT-5 families
GPT-5.6 Luna	9 July 2026	Not disclosed	Knowledge cutoff: 16 February 2026	1,050,000 tokens	Fastest and most cost-efficient GPT-5.6 tier for cost-sensitive, high-volume workloads; broadly corresponds to the nano position in earlier GPT-5 families
GPT-6 Astra	3 September 2026	Not disclosed	Knowledge cutoff: 30 April 2026	1,050,000 tokens	Flagship GPT-6 model for the hardest end-to-end work; state-of-the-art computer use, browsing, software engineering, cybersecurity, science and professional work; supports long-horizon multi-step workflows and up to 128,000 output tokens
GPT-6 Pro	3 September 2026	Not disclosed	Knowledge cutoff: 30 April 2026 for the underlying GPT-6 Astra model	1,050,000 tokens for the underlying GPT-6 Astra model	Higher-compute ChatGPT product mode powered by GPT-6 Astra for the most difficult tasks and longer-running workflows; rolling out to eligible Pro, Business and Enterprise plans

Specialized GPT and Codex models

Specialized GPT and Codex models

Model	Initial release	Parameter count	Context window	Distinguishing role
GPT-5-Codex	15 September 2025	Not disclosed	400,000 tokens	GPT-5 variant optimized through reinforcement learning for agentic software-engineering tasks, code review, terminal use and iterative testing
GPT-5.1-Codex	13 November 2025	Not disclosed	400,000 tokens	GPT-5.1 variant optimized for long-running agentic coding tasks in Codex and similar harnesses
GPT-5.1-Codex mini	13 November 2025	Not disclosed	400,000 tokens	Smaller and more cost-efficient GPT-5.1-Codex variant for less complex coding workflows
GPT-5.1-Codex-Max	19 November 2025	Not disclosed	400,000 tokens per context segment	Frontier agentic coding model for project-scale refactors, deep debugging and tasks lasting many hours; supports automatic context compaction
GPT-5.2-Codex	18 December 2025	Not disclosed	400,000 tokens	Long-horizon agentic coding, large refactors, migrations, Windows development, visual understanding and defensive-cybersecurity capabilities
GPT-5.3-Codex	5 February 2026	Not disclosed	400,000 tokens	Combined agentic coding with broader reasoning and professional knowledge; supported interactive steering during long-running work
GPT-5.3-Codex-Spark	12 February 2026	Not disclosed	128,000 tokens	Compact text-only real-time coding model designed for interactive steering and very high generation speed
GPT-5.5-Cyber	2026	Not disclosed	Not publicly specified	Specialized GPT-5.5 model for vetted professionals performing advanced authorized cybersecurity work
GPT-5.6 Cyber	10 August 2026	Not disclosed	400,000 tokens	Purpose-trained GPT-5.6 Sol-based cybersecurity model for approved defenders conducting advanced authorized vulnerability research, exploit validation and security testing; available through Daybreak Red with separate approval and provisioning

Architectural parameters of GPT models

GPT models with published architectural parameters

GPT models with published architectural parameters

Model	Release year	Parameter count	Number of layers	Hidden state size	Number of attention heads	Context window	Training corpus size
GPT-1	2018	≈117M	12	768	12	512 tokens	BooksCorpus; over 7,000 unpublished books
GPT-2	2019	1.5B / 1.558B	48	1,600	25	1,024 tokens	WebText; slightly over 8 million documents / 40 GB of text
GPT-3	2020	175B	96	12,288	96	2,048 tokens	≈499B-token source mixture; trained for 300B tokens

GPT models without disclosed architectural parameters (as of 2026-07-10)

Model	Release year	Public model ID / alias	API status	Context window	Max output	Input modalities	Output modalities	Knowledge cutoff	Reasoning effort	API doc
GPT-3.5 Turbo	2023	gpt-3.5-turbo	Deprecated shutdown 2026-10-23	16,385 tokens	4,096 tokens	Text	Text	2021-09-01	—	OpenAI API docs
GPT-4	2023	gpt-4	Deprecated shutdown 2026-10-23	8,192 tokens	8,192 tokens	Text	Text	2023-12-01	—	OpenAI API docs
GPT-4 Turbo	2023	gpt-4-turbo	Deprecated shutdown 2026-10-23	128,000 tokens	4,096 tokens	Text, image	Text	2023-12-01	—	OpenAI API docs
GPT-4o	2024	gpt-4o	Deprecated	128,000 tokens	16,384 tokens	Text, image	Text	2023-10-01	—	OpenAI API docs
GPT-4o mini	2024	gpt-4o-mini	Available	128,000 tokens	16,384 tokens	Text, image	Text	2023-10-01	—	OpenAI API docs
GPT-4.5 Preview	2025	gpt-4.5-preview	Removed 2025-07-14	128,000 tokens	16,384 tokens	Text, image	Text	2023-10-01	—	OpenAI API docs
GPT-4.1	2025	gpt-4.1	Available	1,047,576 tokens	32,768 tokens	Text, image	Text	2024-06-01	—	OpenAI API docs
GPT-4.1 mini	2025	gpt-4.1-mini	Available	1,047,576 tokens	32,768 tokens	Text, image	Text	2024-06-01	—	OpenAI API docs
GPT-4.1 nano	2025	gpt-4.1-nano	Deprecated shutdown 2026-10-23	1,047,576 tokens	32,768 tokens	Text, image	Text	2024-06-01	—	OpenAI API docs
GPT-5	2025	gpt-5	Available	400,000 tokens	128,000 tokens	Text, image	Text	2024-09-30	minimal, low, medium (default), high	OpenAI API docs
GPT-5 mini	2025	gpt-5-mini	Available	400,000 tokens	128,000 tokens	Text, image	Text	2024-05-31	minimal, low, medium (default), high	OpenAI API docs
GPT-5 nano	2025	gpt-5-nano	Available	400,000 tokens	128,000 tokens	Text, image	Text	2024-05-31	minimal, low, medium (default), high	OpenAI API docs
GPT-5 Pro	2025	gpt-5-pro	Available	400,000 tokens	272,000 tokens	Text, image	Text	2024-09-30	high only (default)	OpenAI API docs
GPT-5 Chat	2025	gpt-5-chat-latest	Deprecated shutdown 2026-07-23	128,000 tokens	16,384 tokens	Text, image	Text	2024-09-30	—	OpenAI API docs
GPT-5.1	2025	gpt-5.1	Available	400,000 tokens	128,000 tokens	Text, image	Text	2024-09-30	none (default), low, medium, high	OpenAI API docs
GPT-5.1 Chat	2025	gpt-5.1-chat-latest	Deprecated shutdown 2026-07-23	128,000 tokens	16,384 tokens	Text, image	Text	2024-09-30	—	OpenAI API docs
GPT-5.2	2025	gpt-5.2	Available	400,000 tokens	128,000 tokens	Text, image	Text	2025-08-31	none (default), low, medium, high, xhigh	OpenAI API docs
GPT-5.2 Chat	2025	gpt-5.2-chat-latest	Deprecated shutdown 2026-08-10	128,000 tokens	16,384 tokens	Text, image	Text	2025-08-31	—	OpenAI API docs
GPT-5.2 Pro	2025	gpt-5.2-pro	Available	400,000 tokens	128,000 tokens	Text, image	Text	2025-08-31	medium, high, xhigh	OpenAI API docs
GPT-5.3 Chat / GPT-5.3 Instant	2026	gpt-5.3-chat-latest	Deprecated shutdown 2026-08-10	128,000 tokens	16,384 tokens	Text, image	Text	2025-08-31	—	OpenAI API docs
GPT-5.3-Codex	2026	gpt-5.3-codex	Available	400,000 tokens	128,000 tokens	Text, image	Text	2025-08-31	low, medium, high, xhigh	OpenAI API docs
GPT-5.4	2026	gpt-5.4	Available	1,050,000 tokens	128,000 tokens	Text, image	Text	2025-08-31	none (default), low, medium, high, xhigh	OpenAI API docs
GPT-5.4 Pro	2026	gpt-5.4-pro	Available	1,050,000 tokens	128,000 tokens	Text, image	Text	2025-08-31	medium (default), high, xhigh	OpenAI API docs
GPT-5.4 mini	2026	gpt-5.4-mini	Available	400,000 tokens	128,000 tokens	Text, image	Text	2025-08-31	none (default), low, medium, high, xhigh	OpenAI API docs
GPT-5.4 nano	2026	gpt-5.4-nano	Available	400,000 tokens	128,000 tokens	Text, image	Text	2025-08-31	none (default), low, medium, high, xhigh	OpenAI API docs
GPT-5.5	2026	gpt-5.5	Available	1,050,000 tokens	128,000 tokens	Text, image	Text	2025-12-01	none, low, medium (default), high, xhigh	OpenAI API docs
GPT-5.5 Pro	2026	gpt-5.5-pro	Available	1,050,000 tokens	128,000 tokens	Text, image	Text	2025-12-01	medium, high (default), xhigh	OpenAI API docs
GPT-5.5 Instant / Chat Latest	2026	chat-latest	Available rolling alias	400,000 tokens	128,000 tokens	Text, image	Text	2025-08-31	—	OpenAI API docs
GPT-5.6 Sol	2026	gpt-5.6-sol alias: gpt-5.6	Available	1,050,000 tokens	128,000 tokens	Text, image	Text	2026-02-16	none, low, medium (default), high, xhigh, max; optional reasoning.mode: "pro"	OpenAI API docs
GPT-5.6 Terra	2026	gpt-5.6-terra	Available	1,050,000 tokens	128,000 tokens	Text, image	Text	2026-02-16	none, low, medium (default), high, xhigh, max; optional reasoning.mode: "pro"	OpenAI API docs
GPT-5.6 Luna	2026	gpt-5.6-luna	Available	1,050,000 tokens	128,000 tokens	Text, image	Text	2026-02-16	none, low, medium (default), high, xhigh, max; optional reasoning.mode: "pro"	OpenAI API docs
GPT-5.6 Cyber	2026	gpt-5.6-cyber alias: gpt-daybreak-red-latest	Restricted access separate Daybreak Red approval and provisioning required	400,000 tokens	128,000 tokens	Text, image	Text	2026-02-16	Reasoning supported; effort levels not publicly specified	OpenAI API docs
GPT-6 Astra	2026	gpt-6-astra	Limited rollout	1,050,000 tokens	128,000 tokens	Text, image	Text	2026-04-30	low, medium, high, xhigh, max	OpenAI API docs

Models

No.	Date	Model	Note	Source	API docs	System / model card
1	2018-06-11	GPT / GPT-1	First GPT-line model; transformer-based language model trained with unsupervised pre-training followed by supervised fine-tuning.	OpenAI	—	—
2	2019-02-14	GPT-2	Large unsupervised language model released gradually because of misuse concerns.	OpenAI	—	—
3	2020-05-28	GPT-3	175-billion-parameter language model demonstrating strong few-shot learning.	OpenAI	—	—
4	2022-01-27	InstructGPT	Instruction-following GPT-3-based models trained with reinforcement learning from human feedback.	OpenAI	—	—
5	2022-11-30	ChatGPT / GPT-3.5	Conversational model related to InstructGPT; launched as a research preview.	OpenAI	—	—
6	2023-03-01	GPT-3.5 Turbo	API model family used in ChatGPT; optimized for dialogue and lower-cost inference.	OpenAI	OpenAI API docs	—
7	2023-03-14	GPT-4	Large multimodal model accepting image and text inputs and producing text outputs.	OpenAI	OpenAI API docs	OpenAI PDF
8	2023-11-06	GPT-4 Turbo	More capable and lower-cost GPT-4 variant introduced with a 128K-token context window.	OpenAI	OpenAI API docs	—
9	2024-05-13	GPT-4o	Omni model designed for real-time multimodal interaction across text, vision, and audio.	OpenAI	OpenAI API docs	OpenAI System Card
10	2024-07-18	GPT-4o mini	Smaller and lower-cost GPT-4o-class model for cost-efficient multimodal intelligence.	OpenAI	OpenAI API docs	Evaluated in the GPT-4o System Card: OpenAI
11	2025-02-27	GPT-4.5	Research-preview GPT model focused on scaling unsupervised pre-training and improving non-reasoning intelligence.	OpenAI	OpenAI API docs	OpenAI System Card
12	2025-04-14	GPT-4.1	API model with improvements in coding, instruction following, and long-context use.	OpenAI	OpenAI API docs	—
13	2025-04-14	GPT-4.1 mini	Smaller and faster version of GPT-4.1.	OpenAI	OpenAI API docs	—
14	2025-04-14	GPT-4.1 nano	Smallest and lowest-cost version of GPT-4.1.	OpenAI	OpenAI API docs	—
15	2025-08-05	gpt-oss-120b	Large open-weight reasoning model released under the Apache 2.0 license.	OpenAI	OpenAI model docs	OpenAI Model Card
16	2025-08-05	gpt-oss-20b	Smaller open-weight reasoning model released under the Apache 2.0 license.	OpenAI	OpenAI model docs	OpenAI Model Card
17	2025-08-07	GPT-5	Unified flagship model system combining fast responses with adaptive reasoning.	OpenAI	OpenAI API docs	OpenAI System Card
18	2025-08-07	GPT-5 mini	Smaller GPT-5 API model balancing capability, cost, and latency.	OpenAI	OpenAI API docs	Covered in the GPT-5 System Card: OpenAI
19	2025-08-07	GPT-5 nano	Smallest GPT-5 API model for high-volume, low-latency, and low-cost workloads.	OpenAI	OpenAI API docs	Covered in the GPT-5 System Card: OpenAI
20	2025-08-07	GPT-5 Pro / GPT-5 Thinking Pro	Extended-reasoning GPT-5 variant intended for harder and higher-stakes tasks.	OpenAI	OpenAI API docs	Covered in the GPT-5 System Card: OpenAI
21	2025-09-15	GPT-5-Codex	GPT-5 variant optimized for agentic software-engineering tasks in Codex.	OpenAI	OpenAI API docs	OpenAI System Card Addendum
22	2025-11-12	GPT-5.1 Instant	More conversational GPT-5.1 chat model with improved instruction following.	OpenAI	OpenAI API docs	OpenAI System Card Addendum
23	2025-11-12	GPT-5.1 Thinking	GPT-5.1 reasoning model with more adaptive allocation of thinking time.	OpenAI	OpenAI API docs	OpenAI System Card Addendum
24	2025-11-13	GPT-5.1-Codex	GPT-5.1 variant optimized for long-running agentic coding and software-engineering tasks.	OpenAI	OpenAI API docs	—
25	2025-11-13	GPT-5.1-Codex mini	Smaller and lower-cost GPT-5.1-Codex model for agentic coding workloads.	OpenAI	OpenAI API docs	—
26	2025-11-19	GPT-5.1-Codex-Max	Frontier agentic coding model with context compaction for long-running software tasks.	OpenAI	OpenAI API docs	OpenAI System Card
27	2025-12-11	GPT-5.2 Instant	Instant GPT-5.2 variant for everyday ChatGPT interactions and API workloads.	OpenAI	OpenAI API docs	OpenAI System Card Update
28	2025-12-11	GPT-5.2 Thinking	Reasoning variant of GPT-5.2 with stronger performance on scientific, mathematical, and professional tasks.	OpenAI	OpenAI API docs	OpenAI System Card Update
29	2025-12-11	GPT-5.2 Pro	Pro variant of GPT-5.2 for harder reasoning and professional work.	OpenAI	OpenAI API docs	OpenAI System Card Update
30	2025-12-18	GPT-5.2-Codex	GPT-5.2 variant optimized for long-horizon agentic software engineering and cybersecurity work.	OpenAI	OpenAI API docs	OpenAI System Card
31	2026-02-05	GPT-5.3-Codex	Specialized agentic coding model for long-running and complex software-engineering tasks.	OpenAI	OpenAI API docs	OpenAI System Card
32	2026-02-12	GPT-5.3-Codex-Spark	Research-preview, text-only coding model optimized for low-latency, real-time interaction in Codex.	OpenAI	—	—
33	2026-03-03	GPT-5.3 Instant	Updated everyday ChatGPT model focused on smoother, more useful, and better-grounded responses.	OpenAI	OpenAI API docs	OpenAI System Card
34	2026-03-05	GPT-5.4	General-purpose frontier model with native computer-use capabilities.	OpenAI	OpenAI API docs	OpenAI System Card
35	2026-03-05	GPT-5.4 Pro	Pro version of GPT-5.4 for more complex professional, reasoning, and agentic tasks.	OpenAI	OpenAI API docs	Covered in the GPT-5.4 Thinking System Card: OpenAI
36	2026-03-17	GPT-5.4 mini	Smaller and faster GPT-5.4-class model for coding, tool use, and multimodal reasoning.	OpenAI	OpenAI API docs	Covered in an appendix to the GPT-5.4 Thinking System Card: OpenAI
37	2026-03-17	GPT-5.4 nano	Smallest GPT-5.4-class model for high-volume and low-cost API workloads.	OpenAI	OpenAI API docs	—
38	2026-04-23	GPT-5.5	Flagship GPT-5.5 model for professional work, coding, research, and analysis.	OpenAI	OpenAI API docs	OpenAI System Card
39	2026-04-23	GPT-5.5 Pro	Pro version of GPT-5.5 for more demanding reasoning and professional tasks.	OpenAI	OpenAI API docs	Covered in the GPT-5.5 System Card: OpenAI
40	2026-05-05	GPT-5.5 Instant	Updated default ChatGPT model with improved accuracy, clarity, and personalization.	OpenAI	OpenAI API docs	OpenAI System Card
41	2026-06-26	GPT-5.6 Sol	Flagship GPT-5.6 tier for complex professional work. Entered limited preview on 2026-06-26 and became generally available on 2026-07-09.	OpenAI preview announcement OpenAI general-release announcement	OpenAI API docs	OpenAI Deployment Safety Hub
42	2026-06-26	GPT-5.6 Terra	Balanced GPT-5.6 tier for everyday professional work. Entered limited preview on 2026-06-26 and became generally available on 2026-07-09.	OpenAI preview announcement OpenAI general-release announcement	OpenAI API docs	OpenAI Deployment Safety Hub
43	2026-06-26	GPT-5.6 Luna	Fastest and lowest-cost GPT-5.6 tier. Entered limited preview on 2026-06-26 and became generally available on 2026-07-09.	OpenAI preview announcement OpenAI general-release announcement	OpenAI API docs	OpenAI Deployment Safety Hub
44	2026-07-09	GPT-5.6 Sol Pro	Higher-compute product mode based on GPT-5.6 Sol for the most difficult reasoning and professional tasks; not a separately documented base API model.	OpenAI	—	Covered in the GPT-5.6 deployment-safety materials: OpenAI Deployment Safety Hub
45	2026-08-10	GPT-5.6 Cyber	Specialized GPT-5.6 model for approved professionals conducting advanced authorized cybersecurity research, vulnerability discovery, exploit validation, and defensive security testing.	OpenAI	OpenAI API docs	Covered in the GPT-5.6 deployment-safety materials: OpenAI Deployment Safety Hub
46	2026-09-03	GPT-6 Astra	First publicly announced GPT-6-generation model and new flagship for the hardest end-to-end work, with advances in computer use, browsing, software engineering, cybersecurity, scientific research, and long-horizon agentic workflows.	OpenAI	OpenAI API docs	OpenAI Safety Overview
47	2026-09-03	GPT-6 Pro	Higher-compute ChatGPT product mode powered by GPT-6 Astra for the most difficult and longer-running tasks; introduced as part of the staged GPT-6 rollout.	OpenAI	—	Covered in the GPT-6 Astra safety overview: OpenAI

Benchmarks

As of 10 July 2026, this table records selected benchmark results reported directly by OpenAI. It is a source-indexed record of published evaluations, not a controlled leaderboard. Only evaluation runs for which OpenAI reported at least one result in the selected columns are included. Models with no reported result in these columns are omitted rather than shown as wholly blank rows.

The date in the first column is the publication date of the evaluation source, not necessarily the model's release date. A model may appear more than once when OpenAI evaluated a different snapshot, changed the benchmark protocol, reasoning setting, tool access, task set, or harness, or published a materially different later result. Scores are percentages unless otherwise stated. An em dash (—) means that the cited source did not report that benchmark for that exact evaluation run; it does not mean that no result exists elsewhere.

Selected OpenAI-reported benchmark results

Evaluation published	Evaluated model / snapshot	Variant / evaluation setupa	MMLUb	GSM8Kc	HumanEvald	MATHe	GPQA-Df	MMMUg	MMMU-Proh	SWE-Vi	SWE-Proj
2023-03-27	GPT-3.5 GPT-4 Technical Report comparatork	MMLU and GSM8K 5-shot; HumanEval 0-shot	70.0	57.1	48.1	—	—	—	—	—	—
2023-03-27	GPT-4 base model GPT-4 Technical Report evaluationk	MMLU 5-shot; GSM8K 5-shot chain-of-thought; HumanEval 0-shot	86.4	92.0	67.0	—	—	—	—	—	—
2024-05-13	GPT-4 GPT-4o launch comparisonl	Text-evaluation comparison; source label: “GPT-4 (Initial release 23-03-14)”	86.4	—	67.0	42.5	—	—	—	—	—
2024-05-13	GPT-4 Turbol	GPT-4o launch text-evaluation comparison	86.5	—	87.1	72.6	—	—	—	—	—
2024-05-13	GPT-4o launch evaluationl	GPT-4o launch text-evaluation comparison	88.7	—	90.2	76.6	—	—	—	—	—
2024-07-18	GPT-4o mini launch evaluationm	OpenAI simple-evals; API assistant system-message prompt	82.0	—	87.2	—	—	59.4	—	—	—
2025-04-14	GPT-4o (2024-11-20)n	GPT-4.1 appendix; GPQA extraction caveat; SWE-V n=477	85.7	—	—	—	46.0	68.7	—	33.2	—
2025-04-14	GPT-4o mini GPT-4.1 comparisonn	GPT-4.1 appendix; GPQA extraction caveat; SWE-V n=477	82.0	—	—	—	40.2	56.3	—	8.7	—
2025-04-14	GPT-4.5 GPT-4.1 comparisonn	GPT-4.1 appendix; GPQA extraction caveat; SWE-V n=477	90.8	—	—	—	69.5	75.2	—	38.0	—
2025-04-14	GPT-4.1n	GPT-4.1 appendix; GPQA extraction caveat; SWE-V n=477	90.2	—	—	—	66.3	74.8	—	54.6	—
2025-04-14	GPT-4.1 minin	GPT-4.1 appendix; GPQA extraction caveat; SWE-V n=477	87.5	—	—	—	65.0	72.7	—	23.6	—
2025-04-14	GPT-4.1 nanon	GPT-4.1 appendix; GPQA extraction caveat	80.1	—	—	—	50.3	55.4	—	—	—
2025-08-05	gpt-oss-120bo	High reasoning; default system prompt; GPQA-D without tools; SWE-V n=477	90.0	—	—	—	80.1	—	—	62.4	—
2025-08-05	gpt-oss-20bo	High reasoning; default system prompt; GPQA-D without tools; SWE-V n=477	85.3	—	—	—	71.5	—	—	60.7	—
2025-08-07	GPT-5 API launch evaluationp	High reasoning; GPQA-D without tools; SWE-V n=477	—	—	—	—	85.7	84.2	78.4	74.9	—
2025-08-07	GPT-5 mini API launch evaluationp	High reasoning; GPQA-D without tools; SWE-V n=477	—	—	—	—	82.3	81.6	74.1	71.0	—
2025-08-07	GPT-5 nano API launch evaluationp	High reasoning; GPQA-D without tools; SWE-V n=477	—	—	—	—	71.2	75.6	62.6	54.7	—
2025-08-07	GPT-5 Prop	Extended reasoning; GPQA-D without tools	—	—	—	—	88.4	—	—	—	—
2025-09-15	GPT-5-Codexq	High reasoning; all 500 SWE-V tasks	—	—	—	—	—	—	—	74.5	—
2025-11-13	GPT-5.1 API evaluationr	High reasoning; GPQA-D without tools; all 500 SWE-V tasks; JSON-based apply_patch harness	—	—	—	—	88.1	85.4	—	76.3	—
2025-11-13	GPT-5 GPT-5.1 comparison re-evaluationr	High reasoning; GPQA-D without tools; all 500 SWE-V tasks; JSON-based apply_patch harness	—	—	—	—	85.7	84.2	—	72.8	—
2025-11-19	GPT-5.1-Codexs	High reasoning; all 500 SWE-V tasks	—	—	—	—	—	—	—	73.7	—
2025-11-19	GPT-5.1-Codex-Maxs	xhigh reasoning; all 500 SWE-V tasks	—	—	—	—	—	—	—	77.9	—
2025-12-11	GPT-5.1 Thinking GPT-5.2 comparisont	High reasoning; research environment; GPQA-D without tools	—	—	—	—	88.1	—	—	76.3	50.8
2025-12-11	GPT-5.2 Thinkingt	xhigh reasoning; research environment; MMMU-Pro without tools	—	—	—	—	92.4	—	79.5	80.0	55.6
2025-12-11	GPT-5.2 Prot	xhigh reasoning; research environment; GPQA-D without tools	—	—	—	—	93.2	—	—	—	—
2026-02-05	GPT-5.2-Codex GPT-5.3-Codex comparisonu	xhigh reasoning; Codex evaluation setup	—	—	—	—	—	—	—	—	56.4
2026-02-05	GPT-5.3-Codex launch evaluationu	xhigh reasoning; Codex evaluation setup	—	—	—	—	—	—	—	—	56.8
2026-03-05	GPT-5.3-Codex GPT-5.4 comparison re-evaluationv	xhigh reasoning; research environment	—	—	—	—	92.6	—	—	—	56.8
2026-03-05	GPT-5.4v	xhigh reasoning; research environment; MMMU-Pro without tools	—	—	—	—	92.8	—	81.2	—	57.7
2026-03-05	GPT-5.4 Prov	xhigh reasoning; research environment	—	—	—	—	94.4	—	—	—	—
2026-03-17	GPT-5.4 mini/nano comparison re-evaluationw	xhigh reasoning; plain MMMU-Pro result, not the Python-enabled result	—	—	—	—	93.0	—	81.2	—	57.7
2026-03-17	GPT-5.4 miniw	xhigh reasoning; plain MMMU-Pro result, not the Python-enabled result	—	—	—	—	88.0	—	76.6	—	54.4
2026-03-17	GPT-5.4 nanow	xhigh reasoning; plain MMMU-Pro result, not the Python-enabled result	—	—	—	—	82.8	—	66.1	—	52.4
2026-03-17	GPT-5 mini GPT-5.4 mini/nano comparison re-evaluationw	High reasoning; plain MMMU-Pro result, not the Python-enabled result	—	—	—	—	81.6	—	67.5	—	45.7
2026-04-23	GPT-5.5 launch evaluationx	xhigh reasoning; research environment; MMMU-Pro without tools	—	—	—	—	93.6	—	81.2	—	58.6
2026-07-09	GPT-5.5 GPT-5.6 comparison re-evaluationy	Reasoning effort not stated for these entries; MMMU-Pro without tools; source label: “SWE-Bench Pro”	—	—	—	—	93.6	—	81.2	—	59.4
2026-07-09	GPT-5.6 Soly	Reasoning effort not stated for these entries; MMMU-Pro without tools; source label: “SWE-Bench Pro”	—	—	—	—	94.6	—	83.0	—	64.6
2026-07-09	GPT-5.6 Terray	Reasoning effort not stated for these entries; MMMU-Pro without tools; source label: “SWE-Bench Pro”	—	—	—	—	92.9	—	80.7	—	63.4
2026-07-09	GPT-5.6 Lunay	Reasoning effort not stated for these entries; MMMU-Pro without tools; source label: “SWE-Bench Pro”	—	—	—	—	92.3	—	78.4	—	62.7

Table notes

• ^a “Variant / evaluation setup” records the product or API variant, reasoning setting, tool access, snapshot, harness, or evaluated task set stated by the source. Terms such as “Turbo”, “mini”, “nano”, “Instant”, “Thinking”, “Pro”, “Codex”, “high”, and xhigh are product or evaluation labels, not disclosed architectural classes. When OpenAI did not identify an exact checkpoint or reasoning setting, the table does not infer one.

• ^b MMLU entries are limited to results explicitly reported as classic MMLU. Multilingual variants such as MMMLU are not entered in this column. Prompting, answer extraction, checkpoints, and evaluation implementations differ between publications, so the values are not necessarily directly comparable.

• ^c GSM8K is not interchangeable with MGSM. In particular, the 87.0 result reported for GPT-4o mini at launch is MGSM and is therefore not entered in the GSM8K column.

• ^d HumanEval values may depend on sampling, pass@k, prompting, contamination controls, and code-execution methodology. Only values explicitly reported as HumanEval are entered.

• ^e MATH means the original MATH benchmark where the source uses that label. MATH-500, AIME, HMMT, FrontierMath, and other mathematical evaluations are not entered in this column.

• ^f GPQA-D means GPQA Diamond. Results labelled only “GPQA”, without an explicit indication that the Diamond subset was used, are omitted. Tool access, reasoning effort, answer-extraction methods, and later re-evaluations materially affect comparability.

• ^g MMMU is the original MMMU benchmark. It is kept separate from MMMU-Pro because the two are different evaluations.

• ^h MMMU-Pro values are entered separately from classic MMMU. Where a source reports both a no-tools or plain result and a Python- or tool-enabled result, this table uses the no-tools or plain result and identifies the setup in the row or source note.

• ⁱ SWE-V means SWE-bench Verified. These figures are preserved as historical OpenAI-reported results, not as a current controlled ranking. Task subsets, denominators, scaffolds, patch tools, retry policies, and the treatment of unexecutable tasks differ between releases. OpenAI later concluded that SWE-bench Verified had substantial test-design and contamination problems and no longer provided a meaningful signal for frontier software-engineering capabilities.

• ^j SWE-Pro covers results that OpenAI labels “SWE-Bench Pro”, “SWE-Bench Pro (Public)”, or “SWE-Bench Pro, Public”; the exact source label is noted where relevant. These figures should also be treated as historical reported results rather than as a clean cross-model ranking. In a July 2026 audit of the 731-task public split, OpenAI's analysis pipeline identified 200 tasks (27.4%) as broken and its human annotation campaign identified 249 tasks (34.1%); OpenAI estimated that approximately 30% of SWE-Bench Pro tasks were broken.

• ^k The GPT-4 Technical Report evaluates GPT-3.5 as a comparator and a pre-trained base GPT-4 model; the GPT-3.5 row should not be interpreted as a benchmark of the November 2022 ChatGPT product. The report gives MMLU = 70.0 / 86.4, GSM8K = 57.1 / 92.0, and HumanEval = 48.1 / 67.0 for GPT-3.5 / GPT-4. MMLU was evaluated 5-shot and HumanEval 0-shot; GPT-4 GSM8K used 5-shot chain-of-thought prompting. OpenAI mixed training-set data from MATH and GSM8K into GPT-4 pre-training and recommended interpreting the GSM8K result as between true few-shot transfer and benchmark-specific tuning. The report estimated HumanEval contamination at 25%; GPT-4 scored 65.58 on the non-contaminated subset, compared with the headline 67.0.

• ^l The GPT-4, GPT-4 Turbo, and launch GPT-4o values are transcribed from the “Text Evaluation” comparison published with GPT-4o. The source reports MMLU = 86.4 / 86.5 / 88.7, HumanEval = 67.0 / 87.1 / 90.2, and MATH = 42.5 / 72.6 / 76.6 for GPT-4 / GPT-4 Turbo / GPT-4o. Its science comparison uses the label “GPQA” rather than explicitly identifying GPQA Diamond, so those values are not entered here.

• ^m GPT-4o mini launch values were computed with OpenAI's simple-evals repository and the API assistant system-message prompt. The launch page reports MMLU = 82.0, HumanEval = 87.2, and MMMU = 59.4. Its mathematical result of 87.0 is MGSM, not GSM8K.

• ⁿ GPT-4o (2024-11-20), GPT-4o mini, GPT-4.5, and the GPT-4.1 family are taken from the GPT-4.1 comparison appendix. OpenAI's displayed table lists GPT-4o GPQA Diamond = 46.0. The accompanying footnote says that model-based answer extraction can raise GPT-4o from approximately 46 to 54, while changing GPT-4.1 by less than one percentage point; this table reproduces the displayed 46.0 rather than substituting the footnote estimate. For SWE-bench Verified, OpenAI omitted 23 of 500 tasks that could not run on its infrastructure. GPT-4.1 scored 54.6 on the remaining 477 tasks; treating the omitted tasks as zero gives 52.1.

• ^o gpt-oss values are taken from the gpt-oss-120b and gpt-oss-20b Model Card. OpenAI reports basic pass@1 results at high reasoning using the default system prompt. GPQA Diamond values are the no-tools results, and SWE-bench Verified uses a 477-task subset.

• ^p GPT-5, GPT-5 mini, and GPT-5 nano launch values use high reasoning. GPQA Diamond was run without tools. The reported MMMU-Pro result is OpenAI's average across its standard and vision sets. SWE-bench Verified omits 23 of 500 tasks and therefore uses 477 tasks. The separate GPT-5 Pro GPQA result used extended reasoning without tools.

• ^q The GPT-5-Codex launch comparison reports SWE-bench Verified = 74.5 for GPT-5-Codex at high reasoning and 72.8 for GPT-5 at high reasoning, using all 500 tasks. The later GPT-5 comparison result is represented separately because it is not the same 477-task run as the original GPT-5 launch result of 74.9.

• ^r GPT-5.1 and its GPT-5 comparison baseline were evaluated at high reasoning. SWE-bench Verified accuracy was averaged over all 500 tasks using a JSON-based apply_patch harness, producing 76.3 for GPT-5.1 and 72.8 for GPT-5. The same appendix reports GPQA Diamond = 88.1 / 85.7 and MMMU = 85.4 / 84.2 for GPT-5.1 / GPT-5.

• ^s GPT-5.1-Codex used high reasoning and scored 73.7 on all 500 SWE-bench Verified tasks. GPT-5.1-Codex-Max used xhigh reasoning and scored 77.9 on the same headline evaluation.

• ^t GPT-5.2 Thinking and GPT-5.2 Pro were evaluated with the maximum reasoning effort available in the API, xhigh; the GPT-5.1 Thinking comparison used high. OpenAI states that the benchmarks were conducted in a research environment and may differ slightly from production ChatGPT. GPQA Diamond and the displayed MMMU-Pro result are the no-tools results. The source labels the coding benchmark “SWE-Bench Pro, Public”.

• ^u The GPT-5.3-Codex launch appendix reports SWE-Bench Pro (Public) = 56.8 for GPT-5.3-Codex and 56.4 for GPT-5.2-Codex. The comparison table labels both models xhigh.

• ^v GPT-5.4 publication values were produced at xhigh reasoning in a research environment. The publication also supplies a later comparison result for GPT-5.3-Codex, including GPQA Diamond = 92.6. GPT-5.4 MMMU-Pro = 81.2 is the no-tools result.

• ^w The GPT-5.4 mini and nano comparison evaluates GPT-5.4, GPT-5.4 mini, and GPT-5.4 nano at xhigh; GPT-5 mini was evaluated at its maximum available setting, high. The table uses the source's plain MMMU-Pro row rather than its separate Python-enabled row. This later comparison reports GPT-5.4 GPQA Diamond = 93.0, compared with 92.8 in the original GPT-5.4 publication, and is therefore retained as a separate evaluation run.

• ^x GPT-5.5 launch results were produced at xhigh reasoning in a research environment. MMMU-Pro = 81.2 is the no-tools result. The launch table labels the coding benchmark “SWE-Bench Pro (Public)” and reports 58.6. OpenAI also notes evidence of benchmark memorization reported by other laboratories.

• ^y The GPT-5.6 general-release evaluation table reports GPQA Diamond = 94.6 / 92.9 / 92.3, MMMU-Pro without tools = 83.0 / 80.7 / 78.4, and SWE-Bench Pro = 64.6 / 63.4 / 62.7 for GPT-5.6 Sol / Terra / Luna. Its GPT-5.5 comparison row gives 93.6 / 81.2 / 59.4 on the same displayed benchmarks. The table does not identify the exact reasoning effort for these entries, so no effort level is inferred. The source uses the label “SWE-Bench Pro”, without “Public”; the later GPT-5.5 result of 59.4 is retained separately from its launch-time result of 58.6.

External links

• “Improving language understanding with unsupervised learning”, OpenAI, June 11, 2018
• “Better language models and their implications”, OpenAI, February 14, 2019
• “GPT-2: 6-month follow-up”, OpenAI, August 20, 2019
• “GPT-2: 1.5B release”, OpenAI, November 5, 2019
• “Language models are few-shot learners”, OpenAI, May 28, 2020
• “OpenAI API”, OpenAI, June 11, 2020
• “Introducing ChatGPT”, OpenAI, November 30, 2022
• “Function calling and other API updates”, OpenAI, June 13, 2023
• “GPT-4”, OpenAI, March 14, 2023
• “New models and developer products announced at DevDay”, OpenAI, November 6, 2023
• “Hello GPT-4o”, OpenAI, May 13, 2024
• “Introducing GPT-4.1 in the API”, OpenAI, April 14, 2025
• “Introducing GPT-4.5”, OpenAI, February 27, 2025
• “Introducing GPT-5”, OpenAI, August 7, 2025
• “GPT-5.1: A smarter, more conversational ChatGPT”, OpenAI, November 12, 2025
• “Introducing GPT-5.2”, OpenAI, December 11, 2025
• “Introducing GPT-5.2-Codex”, OpenAI, December 18, 2025
• “Introducing GPT-5.3-Codex”, OpenAI, February 5, 2026
• “Introducing GPT-5.3-Codex-Spark”, OpenAI, February 12, 2026
• “GPT-5.3 Instant: Smoother, more useful everyday conversations”, OpenAI, March 3, 2026
• “Introducing GPT-5.4”, OpenAI, March 5, 2026
• “Using GPT-5.4”, OpenAI Developers
• “Models”, OpenAI API
• “Deprecations”, OpenAI API
• “GPT-5.3 and GPT-5.4 in ChatGPT”, OpenAI Help Center
• “Retiring GPT-4o and other ChatGPT models”, OpenAI Help Center
• Generative pre-trained transformer — Wikipedia

Bibliography

• Radford, A. et al. (2018). Improving Language Understanding by Generative Pre-Training. PDF.
• Radford, A. et al. (2019). Language Models are Unsupervised Multitask Learners. PDF.
• Brown, T. B. et al. (2020). Language Models are Few-Shot Learners. arXiv:2005.14165.
• Kaplan, J. et al. (2020). Scaling Laws for Neural Language Models. arXiv:2001.08361.
• Chen, M. et al. (2021). Evaluating Large Language Models Trained on Code. arXiv:2107.03374.
• Ouyang, L. et al. (2022). Training Language Models to Follow Instructions with Human Feedback. arXiv:2203.02155.
• Hoffmann, J. et al. (2022). Training Compute-Optimal Large Language Models. arXiv:2203.15556.
• Bai, Y. et al. (2022). Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback. arXiv:2204.05862.
• OpenAI (2023). GPT-4 Technical Report. arXiv:2303.08774.
• Bubeck, S. et al. (2023). Sparks of Artificial General Intelligence: Early Experiments with GPT-4. arXiv:2303.12712.