New AI Models

GPT-4

GPT-4 is OpenAI's large multimodal model released on March 14, 2023, designed to accept image and text input and produce text output. Per the official announcement, it brings broader general knowledge and stronger reasoning than GPT-3.5, scoring in the 90th percentile on the Uniform Bar Exam, with the vision variant reaching the 99th percentile on the Biology Olympiad. After six months of alignment work, it is 82% less likely to respond to disallowed requests and 40% more likely to produce factual answers than GPT-3.5; the API version offers an 8,192-token context window. It suits professional, academic and complex instruction-following tasks and was available through ChatGPT Plus and the API.

Eleven English v1

Eleven English v1 is the first English-only TTS model from ElevenLabs, also known as Eleven Monolingual v1 (API id eleven_monolingual_v1). Per ElevenLabs' official announcement of Eleven Multilingual v1, this first generation model adjusts delivery based on context and conveys intent and emotion hyper-realistically, and it was described as an excellent storytelling tool that works with the platform's instant voice cloning workflow. It served as the research foundation for Eleven Multilingual v1, released in April 2023, which kept those strengths while extending speech synthesis to seven more languages. Typical use cases are English audiobooks, narration, and voiceover. The v1 models are now deprecated, and ElevenLabs' changelog lists their removal from the API on July 9, 2026, directing users to migrate to newer models such as Eleven Multilingual v2.

Point-E

Point-E is an OpenAI research system released in December 2022 that generates 3D point clouds from complex text prompts. It runs a two-stage pipeline: a text-to-image diffusion model first renders a single synthetic view, then a second diffusion model conditioned on that image produces the point cloud. According to the official paper, it generates a 3D model in only 1-2 minutes on a single GPU, one to two orders of magnitude faster than prior state-of-the-art methods that take multiple GPU-hours per sample, at the cost of somewhat lower sample quality. Pre-trained point cloud diffusion models and evaluation code are open-sourced on GitHub, suiting fast text-to-3D prototyping, creative tooling, and 3D generation research.

ChatGPT

ChatGPT is OpenAI's conversational AI assistant, released on November 30, 2022 as a free research preview. A sibling model to InstructGPT, it was fine-tuned from a GPT-3.5 series model using Reinforcement Learning from Human Feedback (RLHF). The dialogue format lets it answer follow-up questions, admit its mistakes, challenge incorrect premises, and reject inappropriate requests. It suits drafting, brainstorming, coding help, and learning, though OpenAI notes it can still produce plausible-sounding but incorrect answers.

Whisper

Whisper is OpenAI's open source general-purpose speech recognition (ASR) model, released in September 2022 under the MIT license. It was trained on 680,000 hours of multilingual and multitask supervised audio collected from the web, and uses a simple end-to-end encoder-decoder Transformer in which 30-second audio segments are converted to log-Mel spectrograms, with special tokens letting one model perform language identification, phrase-level timestamps, multilingual transcription, and speech translation into English. OpenAI reports that Whisper is far more robust to accents, background noise, and technical vocabulary, makes 50% fewer errors than LibriSpeech-specialized models in zero-shot evaluation across diverse datasets, and beats supervised SOTA on CoVoST2 zero-shot translation to English. It ships in sizes from tiny to large (39M to 1550M parameters, plus a faster turbo variant), fitting use cases such as meeting transcription, subtitle generation, podcast indexing, and multilingual voice interfaces.

GPT-3

GPT-3 is an autoregressive large language model from OpenAI with 175 billion parameters, introduced in May 2020 through the paper Language Models are Few-Shot Learners and later served to developers through the OpenAI API. Per the official paper, it performs tasks with no gradient updates or fine-tuning: tasks and examples are given purely as text in the prompt, the approach known as in-context learning or few-shot learning. It shows strong results on translation, question answering, and cloze tasks, at times matching earlier state-of-the-art systems that required fine-tuning, and handles on-the-fly tasks such as 3-digit arithmetic. Through the API it was offered as the davinci model family (for example text-davinci-002), and in March 2022 it gained edit and insert capabilities for revising existing text or code and adding content in the middle of a document. Typical uses include drafting and rewriting text, code completion, summarization, and question answering.

Codex

Codex is a GPT-3 descendant fine-tuned on publicly available GitHub code, released by OpenAI in August 2021 as a general-purpose programming model. It translates natural language into working code, is strongest in Python, and is proficient in over a dozen languages including JavaScript, Go, Ruby, Swift and TypeScript, with a 14KB Python code memory versus GPT-3's 4KB. On the HumanEval benchmark it solves 28.8% of problems at pass@1 (70.2% with 100 samples per problem), and a distinct production version powers GitHub Copilot. It supports code completion, refactoring, transpilation and code explanation, and was offered through the API in private beta.

CLIP

CLIP (Contrastive Language-Image Pre-training) is a neural network released by OpenAI on January 5, 2021 that learns visual concepts from natural language supervision. It jointly pre-trains an image encoder and a text encoder with a contrastive objective on 400 million image-text pairs collected from the internet, so it can be applied to almost any visual classification benchmark by simply providing the names of the categories as text, similar to the zero-shot capabilities of GPT-2 and GPT-3. OpenAI reports that it matches the accuracy of the original ResNet-50 on ImageNet zero-shot without using any of the 1.28M labeled examples, closes the robustness gap by up to 75%, and outperforms the best publicly available ImageNet model on 20 out of 26 transfer datasets via linear probes. Typical uses include zero-shot image classification, image-text cross-modal retrieval, OCR-style semantic understanding, and serving as a vision-language representation backbone for multimodal research.

DALL·E

DALL·E is OpenAI's first text-to-image model, released on January 5, 2021 as a 12-billion parameter version of GPT-3. Like GPT-3 it is a decoder-only transformer, but it receives text and image as a single stream of up to 1280 tokens (256 for the text and 1024 for the image) and is trained with maximum likelihood to generate all tokens autoregressively. It creates images from natural language captions across many styles, combining unrelated concepts in plausible ways (such as the avocado armchair), controlling object attributes and spatial relationships, adjusting viewpoint and 3D rendering style, and rendering text inside images. The model often fills in details implied by a caption, supports zero-shot image-to-image translation, and its best outputs are picked by reranking 512 samples with CLIP. OpenAI positions it for creative and design exploration, from fashion to interior design.

Image GPT

Image GPT (iGPT) is OpenAI's pixel-level autoregressive image generation and representation model, released on June 17, 2020. It unrolls images into pixel sequences and trains a GPT-2 style transformer with next-pixel prediction on ImageNet, using a 9-bit color palette at 32x32 to 64x64 resolutions; iGPT-S/M/L contain 76M, 455M and 1.4B parameters, plus a 6.8B iGPT-XL trained on ImageNet and web images. iGPT generates coherent image samples and completes the unseen half of a given image, and its mid-layer features work as strong classification representations: a linear probe reaches 96.3% on CIFAR-10, 95.5% on STL-10, and 72.0% top-1 on ImageNet with iGPT-XL, with fine-tuned iGPT-L matching GPipe at 99.0% on CIFAR-10. It suits research on unsupervised representation learning, generative sequence modeling, and image completion.

Jukebox

Jukebox is a neural network from OpenAI, released on April 30, 2020, that generates music with rudimentary singing as raw audio. Given a genre, artist style, and lyrics, it produces new songs from scratch: a multi-scale VQ-VAE compresses 44.1 kHz audio into discrete codes and autoregressive sparse transformers model them. Trained on 1.2 million songs with lyrics and metadata, it generates diverse, high-fidelity tracks with coherence up to several minutes, though sampling is slow at roughly 9 hours per minute of audio. Suited to music generation research, artist and genre style exploration, and creative audio experiments.

GPT-2 1.5B

GPT-2 1.5B is the final and largest version of OpenAI's GPT-2 language model, released on November 5, 2019 with full code and model weights as the last step of a staged release plan. Its text generation is noticeably convincing: in a survey run by Cornell University, human readers gave 1.5B outputs a credibility score of 6.91 out of 10, above the 774M version (6.72). A detector model released alongside identifies 1.5B GPT-2 text with about 95% accuracy, though detection becomes harder as model size grows. The model suits research on text generation, language model behavior, bias studies, and synthetic text detection.

GPT-2 774 million parameter

GPT-2 774 million parameter is an intermediate checkpoint in OpenAI's staged release of the GPT-2 language model, published on August 20, 2019 after the small 124M and medium 355M versions. As a larger autoregressive language model, it performs text generation such as prompt continuation and multi-paragraph drafting, and it served research on dialogue, code completion, and misuse detection. OpenAI released it together with a technical report and an open-source legal agreement after misuse studies with four university partners; Cornell research cited on the official page found 72% of one cohort judged GPT-2 articles credible versus 83% for real New York Times articles. The full 1558M version was planned to follow.

MuseNet

MuseNet is a deep neural network from OpenAI for music generation. It follows the same unsupervised next-token prediction approach as GPT-2, learning patterns of harmony, rhythm, and style from hundreds of thousands of MIDI files. Built on the Sparse Transformer, a 72-layer network with 24 attention heads and full attention over a 4096-token context, it can produce 4-minute pieces with up to 10 different instruments and blend styles from country to Mozart to the Beatles. Composer and instrumentation tokens let users steer generation toward a chosen composer, genre, or instrumentation, making it useful for melody sketching, style fusion experiments, and creative music production.

Sparse Transformer

Sparse Transformer is an attention architecture OpenAI released in April 2019. It reformulates self-attention from O(N^2) to O(N sqrt N) complexity, so Transformers can model sequences of tens of thousands of elements with hundreds of layers, extracting patterns from sequences 30x longer than previously possible. It set state-of-the-art density estimation results on CIFAR-10, Enwik8 and Imagenet 64, completed 64x64 ImageNet images, generated raw audio waveforms of length 65,000, and shipped with open-sourced block-sparse GPU kernels, making sparse attention a practical route to efficient long sequence modeling across text, images and audio.

Glow

Glow is an OpenAI reversible generative model for images, part of the normalizing flow family. It extends NICE and RealNVP with learned invertible 1x1 convolutions and a simpler architecture, enabling exact latent-variable inference, exact log-likelihood evaluation, and efficient parallel sampling, at about 130ms per 256x256 sample on an NVIDIA 1080 Ti. Its latent space supports interpolation between faces and attribute manipulation such as hair color, age, or beard density, without attribute labels during training. Official benchmarks show better bits per dimension than RealNVP on ImageNet 32x32, ImageNet 64x64, and LSUN. OpenAI released the code on GitHub plus an online visualization demo, so Glow suits generative modeling research, image editing experiments, and representation learning.

OpenAI Five

OpenAI Five is a team of five neural networks announced by OpenAI on June 25, 2018 to play Dota 2. It was trained entirely with reinforcement learning and self-play, using a scaled-up version of Proximal Policy Optimization on 256 GPUs and 128,000 CPU cores, accumulating roughly 180 years of game experience per day, with a separate LSTM per hero and no human data. By June 2018 it was beating amateur and semi-pro teams ranked up to the 99th percentile, and in April 2019 it became the first AI to defeat the Dota 2 world champion team OG in an esports game. The project serves as a benchmark for long-horizon planning, teamwork, and large-scale deep reinforcement learning research.

A2C

A2C (Advantage Actor-Critic) is a reinforcement learning algorithm that OpenAI released on August 18, 2017 as part of OpenAI Baselines. It is a synchronous, deterministic variant of A3C: parallel actors each collect fixed-length segments of experience, and the policy is updated after averaging over all actors, which uses GPUs more effectively since they perform best with large batch sizes. According to the official announcement, this synchronous implementation performs better than the asynchronous A3C implementation, is more cost-effective than A3C on single-GPU machines, and ships with code for training feedforward convnets and LSTMs on the Atari benchmark, making it a foundational baseline for actor-critic policy learning.

ACKTR

ACKTR, short for Actor Critic using Kronecker-factored Trust Region and pronounced like the word actor, is a reinforcement learning algorithm published on the OpenAI Baselines blog on August 18, 2017, developed by researchers at the University of Toronto and New York University in collaboration with OpenAI. It combines actor-critic methods, trust region optimization, and Kronecker-factored natural gradient updates (K-FAC), which gives it better sample efficiency than TRPO and A2C while each update costs only 10 to 25 percent more than a standard gradient step. Because its curvature estimates scale well with batch size, it fits large-scale distributed training. OpenAI benchmarked it on 49 Atari games against A2C, PPO, and ACER, and it can learn simulated robot control policies directly from low-resolution pixel inputs, making it suited to sample-hungry control and game-playing tasks.