OpenAI

Released at DevDay on September 29, seven days after GPT-6 Sol, whose docs now point to it as the newer Sol model. OpenAI pitches it as "near-Astra pe…

The two lower tiers of OpenAI's GPT-6 generation, released together on September 22 below GPT-6 Astra. GPT-6 Sol is "built to power complex coding and…

OpenAI put the Codex harness behind an API. The Agents API, in public beta from September 10, 2026, runs cloud agents on the same harness and infrastr…

Two image models released to the API and ChatGPT on September 8, 2026. Sunburst (snapshot gpt-image-2.5-sunburst-2026-09-08) is OpenAI's most capable…

OpenAI's claim on a Millennium Prize problem. The 166-page paper, authored simply "OpenAI," states as Theorem 1.1 that for every viscosity there exist…

OpenAI's GPT-6 generation flagship, announced September 3, 2026 as "the most intelligent and aligned model in the world" and pitched around computer u…

Two speech-to-text models replacing Whisper and 4o-transcribe as the OpenAI API defaults: batch GPT-Transcribe ($0.0045/min) and streaming GPT-Live-Tr…

OpenAI's full-duplex voice models (GPT-Live-1 and GPT-Live-1 mini): simultaneous listening and speaking — the model can be interrupted mid-sente…

OpenAI benchmark of 129 research-level computational-biology problems. Because tasks are built from synthetic data generated from fully-known causal s…

OpenAI's next-generation flagship series, launched June 26 in a restricted preview: Sol (strongest, with improved agentic performance in coding, biolo…

An OpenAI alignment study asking whether RL on beneficial behavior in realistic domains generalizes alignment beyond the training distribution. The au…

OpenAI benchmark of 750 life-science research tasks authored by 173 PhD scientists, graded against 19,020 rubric criteria across 7 workflow types (lit…

Pre-deployment safety methodology: replay ~1.3M de-identified real conversations against a candidate model to predict its deployed behavior rates befo…

OpenAI's most capable model at its release, when it debuted #1 globally on the AA index; it reads 38 on AA Intelligence Index v4.3. Described as "a ne…

Current flagship model. 1M token context. Knowledge cutoff August 31, 2025. Available in thinking, pro, mini, and nano variants. Architecture and para…

1,320 real-world professional tasks across 44 occupations in the 9 largest US GDP-contributing industries (real estate, government, manufacturing, hea…

Multi-component system combining a fast model, reasoning model, and router into a single interface. 400K token context. Replaced GPT-4o, o3, o4-mini,…

OpenAI's open-weight release of August 2025, its first language-model weights since GPT-2, and a backfill here: two MoE reasoning models under Apache…

Second-generation reasoning model, advancing o1's test-time compute scaling. 200K token context. Released alongside o4-mini, a smaller reasoning varia…

4,326 short-form factuality questions adversarially collected against GPT-4 to measure factual recall and calibration. Questions are designed to have…

75 ML engineering competitions from Kaggle testing whether AI agents can train models, prepare datasets, and run experiments autonomously. Each task i…

OpenAI's first reasoning model, trained with reinforcement learning to perform extended chain-of-thought reasoning before answering. Uses hidden "thin…

GPT-4o ("omni") — OpenAI's first natively multimodal model, processing text, audio, image, and video inputs and generating text and audio output…

Text-to-video generation model using a diffusion transformer operating on spacetime patches. Generates up to 1080p video, up to 20 seconds long, from…

"Improving Image Generation with Better Captions" — the third generation of DALL-E, integrated natively into ChatGPT. Key innovation is a captio…

Demonstrated that process reward models (PRMs) that evaluate each step of a reasoning chain significantly outperform outcome reward models (ORMs) that…

GPT-4 technical report. Architecture details undisclosed by OpenAI (parameter count, training data, and hardware not published). 8K token context at l…

"Robust Speech Recognition via Large-Scale Weak Supervision" — an encoder-decoder Transformer trained on 680,000 hours of multilingual audio dat…

"Training language models to follow instructions with human feedback" — introduced the three-stage RLHF pipeline that became industry standard:…

"Evaluating Large Language Models Trained on Code" — GPT-3 fine-tuned on publicly available code from GitHub. The largest variant, Codex-12B, so…

Open-source GPU programming language and compiler for writing high-performance neural network kernels. Python-like syntax that compiles to optimized G…

164 hand-written Python programming problems with function signatures, docstrings, and unit tests. Each problem tests language comprehension, algorith…

"Learning Transferable Visual Models From Natural Language Supervision" — Contrastive Language-Image Pre-training (CLIP) trained on 400M image-t…

"Zero-Shot Text-to-Image Generation" — a 12B parameter autoregressive Transformer that generates images from text descriptions. Uses a discrete…

"Language Models are Few-Shot Learners" — 175B parameter dense Transformer (96 layers, 12288 hidden, 96 heads) trained on 300B tokens from a fil…

Established power-law scaling relationships between language model performance and three key factors: model size (N), dataset size (D), and compute bu…

"Language Models are Unsupervised Multitask Learners" — 1.5B parameter decoder-only Transformer (48 layers, 1600 hidden, 25 heads) trained on We…

"Improving Language Understanding by Generative Pre-Training" — the paper that launched the GPT series. A 117M parameter decoder-only Transforme…

Introduced Proximal Policy Optimization, a family of policy gradient methods that alternate between sampling data through interaction with the environ…

Toolkit for developing and comparing reinforcement learning algorithms. Provides a standard API for RL environments (CartPole, Atari, MuJoCo, etc.) th…