OpenAI

OpenAI released GPT-5.4 mini and nano, their most capable small models yet optimized for coding and subagents. GPT-5.4 mini significantly improves over GPT-5 mini across coding, reasoning, multimodal understanding, and tool use while running more than 2x faster. GPT-5.4 nano is the smallest, cheapest version for tasks where speed and cost matter most. These models are designed for high-volume workloads where latency directly shapes product experience.

• GPT-5.4 mini: 2x faster than GPT-5 mini
• 54.4% on SWE-Bench Pro (mini)
• 400K context window (mini)
• GPT-5.4 nano: $0.20 per 1M input tokens
• Optimized for subagents and coding workflows

OpenAI retired GPT-4o and several related models from ChatGPT, while keeping them available in the API. The change marked a cleanup of the ChatGPT model lineup as newer GPT-5-family models became the default.

• GPT-4o removed from ChatGPT
• GPT-4.1 and GPT-4.1 mini retired too
• GPT-5 variants became the default
• Models remained available in the API

OpenAI shipped a GPT-5.2 Instant update that improved response style and quality in ChatGPT and the API. The update made answers more measured, grounded, and contextually appropriate for advice-seeking and how-to questions.

• Improved response style and quality
• More measured, grounded tone
• Better advice-seeking and how-to answers
• Rolled out in ChatGPT and the API

OpenAI launched GPT-5.3-Codex, its most capable agentic coding model yet. The release combined the Codex and GPT-5 training stacks to improve code generation, reasoning, and steerable long-running coding workflows.

• Combined Codex and GPT-5 training stacks
• Improved code generation and reasoning
• Built for steerable coding workflows
• Faster than prior Codex generations
• A step toward general-purpose coding agents

OpenAI launched AgentKit, a complete set of tools for building, deploying, and optimizing agents. The release bundled visual workflow building, connector management, embedded chat experiences, and expanded evaluation features for production agent workflows.

• Agent Builder for visual workflow design
• Connector Registry for data and tool governance
• ChatKit for embedded agent experiences
• Expanded eval and grading capabilities
• Built to speed up production agent development

OpenAI released GPT-5, its smartest and most useful model yet with built-in thinking. The launch emphasized stronger coding, reasoning, writing, health, and multimodal performance, plus a unified system that decides when to answer quickly or think longer.

• Unified system with built-in thinking
• Stronger coding and reasoning performance
• Improved writing and health use cases
• Multimodal support across common tasks
• Available to all users with tiered access

OpenAI released o3 and o4-mini, its latest reasoning models designed to think longer before responding. The launch emphasized stronger visual reasoning, agentic tool use, and major gains in coding, math, and science tasks.

• Latest o-series reasoning models
• Thinking with images and tools
• Strong coding, math, and science performance
• Agentic tool use improvements
• Codex CLI launched alongside the models

OpenAI released o3-mini, a cost-efficient reasoning model optimized for coding, math, and science. It offered adjustable reasoning effort levels and was made available to free ChatGPT users, democratizing access to reasoning capabilities while maintaining strong performance.

• Cost-efficient reasoning model
• Adjustable reasoning effort
• Free tier availability
• Optimized for STEM tasks
• Lower latency than o1

OpenAI announced o3, achieving a historic breakthrough on the ARC-AGI (Abstraction and Reasoning Corpus for Artificial General Intelligence) benchmark. In low-compute mode, o3 scored 75.7%, surpassing human-level performance (85% threshold for AGI). In high-compute mode with extended thinking time, o3 reached 87.5%, becoming the first AI system to exceed human baseline performance on this challenging abstract reasoning benchmark designed specifically to test general intelligence capabilities.

• 75.7% on ARC-AGI (low-compute mode)
• 87.5% on ARC-AGI (high-compute mode)
• First AI to surpass human baseline (85%)
• Historic AGI milestone
• Demonstrates novel reasoning capabilities

OpenAI released the full o1 model and o1 Pro Mode, featuring enhanced reasoning capabilities and extended thinking time. o1 demonstrated strong performance on math, science, and coding tasks, including continued improvements on the ARC-AGI abstract reasoning benchmark. The release also brought o1-mini to all ChatGPT users, democratizing access to reasoning models.

• Full o1 model released
• o1 Pro Mode with extended thinking
• Enhanced reasoning capabilities
• Improved ARC-AGI performance
• o1-mini available to all ChatGPT users
• Significant improvements on STEM tasks

OpenAI introduced the o1 series (o1-preview and o1-mini), featuring new large-scale reinforcement learning algorithms that train models to perform complex reasoning. o1 spent more time thinking before responding, achieving PhD-level accuracy on challenging reasoning tasks. On the ARC-AGI benchmark, o1 achieved a breakthrough score of approximately 21%, significantly advancing AI performance on abstract reasoning tasks.

• Chain-of-thought reasoning
• Trained with reinforcement learning
• PhD-level accuracy on complex tasks
• 83% on International Math Olympiad qualifying exam
• ~21% on ARC-AGI benchmark
• Ranked in 89th percentile on Codeforces

OpenAI released GPT-4o ('o' for omni), their flagship model with native multimodal capabilities across text, audio, and vision. It featured response times as fast as humans (232ms average for audio), 50+ language support, and was made available to free ChatGPT users.

• Native multimodal (text, audio, vision)
• 232ms average audio response time
• 50+ language support
• Free tier availability
• Improved vision and non-English performance

OpenAI unveiled Sora, a text-to-video AI model capable of generating high-quality videos up to 60 seconds long. Sora demonstrated understanding of physics, complex camera movements, and consistent character appearances across frames. It represented a major breakthrough in AI video generation, producing cinematic-quality outputs from text prompts alone.

• Up to 60-second video generation
• Complex camera movements and physics
• Consistent characters and scenes
• Multiple shots in single generation
• Cinematic quality outputs

OpenAI released GPT-4 Turbo with a 128K context window (4× larger than previous GPT-4), knowledge cutoff updated to April 2023, and significantly lower pricing. It also introduced JSON mode and reproducible outputs, making it more practical for production applications.

• 128K context window (4× increase)
• Knowledge updated to April 2023
• 3× cheaper input, 2× cheaper output
• JSON mode and reproducible outputs
• Better instruction following

At DevDay, OpenAI announced GPTs (custom versions of ChatGPT), the Assistants API, GPT-4 Turbo with 128K context, and multimodal capabilities. This enabled developers to build AI agents with persistent threads, built-in retrieval, and code interpreter capabilities.

• GPTs (custom ChatGPT versions)
• Assistants API for agents
• GPT-4 Turbo (128K context)
• Multimodal vision capabilities
• GPT Store announced

OpenAI announced that ChatGPT had reached 100 million weekly active users, less than a year after its launch. This milestone cemented ChatGPT as one of the fastest-growing consumer applications in history and demonstrated massive mainstream AI adoption.

• 100 million weekly active users
• Fastest-growing consumer app
• Less than 1 year since launch
• Mainstream AI adoption achieved
• 1 million+ API developers

OpenAI released GPT-4 with vision capabilities, allowing the model to understand and reason about images. It could analyze photographs, charts, diagrams, and documents, marking a major step toward truly multimodal AI systems.

• Image understanding capabilities
• Analyze charts and diagrams
• Read text from images (OCR)
• Visual reasoning and description
• Available in ChatGPT and API

OpenAI released DALL-E 3, integrated natively into ChatGPT. The model significantly improved image generation quality and understanding of complex prompts. It could render text within images and follow complex multi-paragraph prompts with high fidelity.

• Native ChatGPT integration
• Significantly improved quality
• Text rendering in images
• Follows complex multi-paragraph prompts
• Built-in safety mitigations

OpenAI released GPT-4, a large multimodal model accepting image and text inputs. It demonstrated human-level performance on professional exams (bar exam, SAT) and significantly improved reasoning capabilities. The model was trained with RLHF and constitutional AI techniques for safety.

• Multimodal (text + images)
• Human-level exam performance
• 8K and 32K context windows
• Training cost >$100M
• 82% less restricted content vs GPT-3.5

OpenAI launched ChatGPT Plus, a $20/month subscription offering faster response times, priority access during peak times, and early access to new features. This marked the beginning of monetization for consumer AI products at scale.

• $20/month subscription
• Faster response times
• Priority access during peak hours
• Early access to new features
• First major consumer AI monetization

OpenAI launched ChatGPT, a conversational AI based on GPT-3.5 fine-tuned with RLHF. The product gained 1 million users in 5 days and 100 million in 2 months, becoming the fastest-growing consumer application in history and bringing AI into mainstream consciousness.

• 1 million users in 5 days
• 100 million users in 2 months
• Fastest-growing consumer app ever
• RLHF for safety and alignment
• Brought AI to mainstream awareness

OpenAI released DALL-E 2, a significant upgrade that generated more realistic and accurate images with 4× greater resolution than the original. The model used a CLIP-based architecture and diffusion techniques. It became available to the public via API and web interface, making AI image generation accessible to millions of users.

• 4× higher resolution than DALL-E 1
• CLIP-based diffusion architecture
• Public API and web interface
• Realistic photorealistic images
• Inpainting and outpainting capabilities

OpenAI introduced InstructGPT using RLHF (Reinforcement Learning from Human Feedback). The paper showed that 1.3B parameter InstructGPT outperformed 175B GPT-3 on following instructions, demonstrating that alignment with human preferences matters more than scale. This became the foundation for ChatGPT and modern alignment techniques.

• RLHF alignment methodology
• 1.3B model outperforms 175B GPT-3 on instruction-following
• Preference-based training
• Foundation for ChatGPT
• Alignment > Scale principle

OpenAI released the Codex API, providing access to the model powering GitHub Copilot. The model was trained on 54 million GitHub repositories and could interpret natural language and execute over a dozen programming languages.

• API access to Copilot model
• Trained on 54M GitHub repos
• 12B parameters
• Support for 12+ programming languages
• Natural language to code translation

OpenAI introduced DALL-E, a 12-billion parameter version of GPT-3 trained to generate images from text descriptions. It could create novel combinations of concepts, transform existing images, and even render text. This marked a major milestone in multimodal AI.

• 12B parameter transformer model
• Text-to-image generation
• Zero-shot image generation
• Combining concepts in novel ways
• Foundation for modern image generation

OpenAI released GPT-3 with 175 billion parameters, demonstrating remarkable few-shot and zero-shot learning capabilities. The model could perform tasks it wasn't explicitly trained on, simply by providing examples in the prompt. This demonstrated the power of scale in language models.

• 175B parameters (100x GPT-2)
• Few-shot learning without fine-tuning
• Zero-shot capabilities across tasks
• Trained on 300B tokens
• API waitlist opened immediately

OpenAI released the full GPT-2 model with 1.5 billion parameters after a staged release process and partnership with researchers. The model demonstrated impressive text generation capabilities but also raised ongoing discussions about AI safety and dual-use concerns.

• Full 1.5B parameter model released
• Staged release with safety research
• Impressive zero-shot capabilities
• No fine-tuning needed for many tasks
• Established safety release practices

OpenAI released a small preview of GPT-2 but withheld the full model citing concerns about potential misuse for generating fake news. The decision sparked debate about AI safety and responsible release practices in the research community.

• 1.5B parameters (full model withheld)
• Concerns about fake news generation
• Staged release approach
• Sparked AI safety debate
• Full model released later in 2019

OpenAI introduced GPT-1 (Generative Pre-trained Transformer), demonstrating that unsupervised pre-training on large text corpora followed by supervised fine-tuning could achieve state-of-the-art results on various NLP benchmarks. This established the pre-training paradigm.

• First GPT model - 117M parameters
• Unsupervised pre-training + supervised fine-tuning
• Transformer decoder architecture
• Trained on BookCorpus (7,000 books)
• Established pre-training paradigm for NLP

OpenAI released Universe, a platform for measuring and training AI agents across diverse environments including games, web interfaces, and applications. It provided a single interface to thousands of environments using Virtual Network Computing (VNC).

• Unified interface for AI training environments
• Support for games, web, and apps
• Used VNC for environment control
• Aimed to accelerate reinforcement learning research
• Later evolved into more focused projects