AI Timeline

OpenAI begins rolling out GPT-6 Astra, advancing autonomous computer use and scientific reasoning, with reported near-perfect ARC-AGI-3 and FrontierMath Tier 4 scores.

Anthropic publishes the first complete Lean proof of Fermat's Last Theorem, formalized largely autonomously by Claude in 11 days.

OpenAI publishes a proposed Navier-Stokes Millennium Problem solution from an internal model, including a Lean proof of finite-time singularities in forced fluid flow.

Google releases AlphaGenome Atlas, mapping predicted molecular effects of all nine billion possible single-letter human DNA variants for researchers.

Google publishes WeatherNext cyclone research in Nature, achieving three-day forecasts as accurate as previous two-day forecasts, and releases the model.

OpenAI releases GPT-5.6 Sol, Terra, and Luna broadly, improving performance per token and introducing parallel agents for demanding tasks.

Moonshot AI launches Kimi K3, a 2.8-trillion-parameter multimodal model with million-token context, expanding the frontier of open models.

Claude Fable 5 helps mathematician Levent Alpöge disprove the Jacobian conjecture in dimension three, overturning a problem posed in 1939.

Google introduces Gemini Robotics 2 and ER 2, enabling whole-body control, adaptation to new robot bodies, and collaboration between robots.

Anthropic launches Fable 5 for general use and restricted Mythos 5; US export controls suspend both three days later, until July 1.

Google launches Gemini 3.5 Flash, surpassing Gemini 3.1 Pro on coding and agent benchmarks while retaining Flash speeds.

Google introduces Gemini Omni Flash, combining multimodal reasoning with video generation and conversational editing.

An internal OpenAI model disproves Erdős's unit-distance conjecture, open since 1946; external mathematicians check the proof.

Anthropic unveils Claude Mythos Preview through Project Glasswing, restricting access to selected defenders after demonstrating major advances in software vulnerability discovery.

OpenAI launches ChatGPT Images 2.0, adding image generation with reasoning, web search, and improved text rendering.

DeepSeek releases V4 Pro and Flash, open-weight models that sharply reduce the memory and compute costs of million-token contexts.

OpenAI releases GPT-5.4, unifying reasoning, coding, and native computer use with 1M-token context; smaller mini and nano variants follow.

Anthropic releases Claude Opus 4.6 and Sonnet 4.6, advancing autonomous coding and computer use with 1M-token context windows in beta.

Google upgrades Gemini 3 Deep Think for scientific reasoning, reaching an ARC Prize-verified 84.6% on ARC-AGI-2.

Axiom publishes AxiomProver's verified Lean proofs for all 12 Putnam 2025 problems, eight solved within the exam window.

Anthropic previews Cowork, extending Claude's agents to everyday work with local files.

Google DeepMind publishes AlphaGenome in Nature, advancing prediction of how DNA variants affect gene regulation.

Google introduces Gemini 3 Flash, opens the upgraded Gemini Deep Research agent to developers through the Interactions API, and updates Gemini 2.5 Flash Native Audio.

Poetiq reports results up to 75% on the ARC-AGI-2 public evaluation set using GPT-5.2 X-High with its reasoning system.

Google released Nano Banana Pro, a superior image generation and editing model based on Gemini 3.

Anthropic announced Claude Opus 4.5, delivering elite coding and agentic performance at a significantly reduced price.

DeepSeek released DeepSeekMath-V2 as open source, achieving gold-medal performance in math olympiads.

1X announced Neo, a humanoid robot marketed as the first consumer-ready model for home use.

An advanced Gemini variant, reported as Gemini 2.5 - Deep Think, achieved gold-medal-level performance at the ICPC World Finals programming contest.

OpenAI reported a reasoning and code model achieved a perfect score (12/12) in ICPC testing.

OpenAI and NVIDIA announce a letter of intent for a strategic partnership to deploy at least 10 gigawatts of NVIDIA systems for OpenAI's AI infrastructure.

Google introduced Gemini 2.5 Deep Think, a special "extended thinking" mode for solving complex problems and exploring alternatives.

Google DeepMind announced Genie 3.0, a "world model" for creating interactive 3D environments from text, maintaining consistency for several minutes.

Google launched a preview of Gemini 2.5 Flash Image (showcased as *nano-banana*), an advanced model for precise image editing, merging, and maintaining character consistency.

xAI releases Grok 4, achieving a new SOTA of 15.9% on ARC-AGI v2 and 25.4% on Humanity’s Last Exam.

OpenAI unveils the ChatGPT Agent, embedding autonomous coding, web research and tool use directly within the chat interface.

OpenAI reports that an experimental reasoning model reaches gold-medal-level performance on IMO 2025, scoring 35/42 without external tools or internet access.

An advanced version of Gemini Deep Think achieves gold-medal-level performance on IMO 2025, solving five of six problems and scoring 35/42 in an evaluation graded by the IMO.

Anthropic releases Claude Opus 4 and Claude Sonnet 4, with improved coding and optional extended thinking with tool use. Rakuten reports that Opus 4 sustained autonomous work on a coding task for seven hours.

Google releases Veo 3, a video-generation model with synchronized dialogue, sound effects, and ambient audio, alongside Imagen 4, an image model with improved detail and text rendering.

Google DeepMind introduces AlphaEvolve, a coding agent combining Gemini models with evolutionary search. On a collection of over 50 mathematical problems, it matches the best-known solutions in about 75% of cases and improves them in about 20%.

Google releases Gemini Diffusion, an experimental text diffusion model achieving high-speed text generation with enhanced control and creativity via noise refinement.

Alibaba releases Qwen 3 as open source, in various sizes, with very impressive capabilities for their size.

Google introduced Gemini 2.5 Pro, an experimental "Thinking model" with advanced reasoning and planning capabilities, a 1 million token context window, achieving top rankings across several key benchmarks.

OpenAI integrated GPT-4o Image Generation, enabling high-fidelity text-to-image creation, text rendering within images, and more.

Google expanded experimental image generation and editing within Gemini 2.0 Flash Experimental, enabling image generation and editing, including enhanced text creation capabilities.

xAI launches Grok 3, Grok 3 Reasoning and Grok 3 mini, next-generation AI models trained with 10 times the computing power of Grok 2, significantly improving SOTA performance. They include "Think" and "Big Brain" modes for advanced reasoning, as well as DeepSearch for autonomous web searches.

Anthropic introduces Claude 3.7 Sonnet, a hybrid reasoning model with improved coding performance and an optional Extended Thinking mode that exposes its reasoning process.

OpenAI unveils Deep Research, a tool for autonomous research, enabling real-time web searches and comprehensive report generation.

OpenAI releases GPT-4.5, featuring advanced pattern recognition and significantly reduced hallucinations, improving accuracy and reliability.

Sesame AI unveiled its Conversational Speech Model (CSM), enabling remarkably human-like, real-time voice interaction, incorporating emotional nuances, natural pauses, laughter, and contextual memory.

OpenAI releases Operator for Pro subscribers in the U.S., an experimental AI agent capable of browsing websites and performing actions.

DeepSeek open-sourced the reasoning models R1 and R1-Zero, which demonstrated capabilities similar to o1 across various domains at a fraction of the cost. Additionally, smaller distilled models were released, achieving high performance relative to their size.

Google published a research paper on a new language model architecture called Titans, designed to enable models to retain both short- and long-term memory. This architecture significantly improves processing for extended context windows.

OpenAI made the o3 mini reasoning model available to all users, including the free tier, featuring three reasoning levels. The model matches or comes close to o1 in several benchmarks, significantly surpasses it in coding, and remains significantly faster and more cost-efficient.

OpenAI released SORA, a video generation model, along with the full version of o1 and o1 Pro for advanced subscribers. Additionally, the company launched a live video mode for GPT 4o.

Google launched Gemini 2.0 Flash in beta. This model leads benchmarks and outperforms the previous version, Gemini Pro 1.5. Additionally, Google introduced live speech and video mode and announced built-in image generation capabilities within the model.

Google releases Gemini-2.0-Flash-Thinking, an experimental reasoning model based on Gemini 2.0 Flash.

Google introduces Veo 2 through a VideoFX waitlist, reporting improved video generation and model capabilities up to 4K and several minutes. It also updates Imagen 3, improving image quality and prompt adherence.

Deepseek open-sourced Deepseek V3, a model with 671B parameters that surpasses closed-source SOTA models across several benchmarks.

Alibaba unveiled QVQ-72B-Preview, a cutting-edge thinking model capable of analyzing images, featuring SOTA-level performance.

OpenAI previews o3 and o3-mini ahead of release. ARC Prize reports that o3 scores 75.7% on ARC-AGI's semi-private evaluation at lower compute and 87.5% at higher compute.

OPENAI has released two nextgeneration AI models to its subscribers: o1 preview and o1 mini. These models show a significant improvement in performance, particularly in tasks requiring reasoning, including coding, mathematics, GPQA, and more.

Meta releases as open source the llama 3.1 model in sizes 8B, 70B and 405B. The large model features the same capabilities as the best closed source models

Google DeepMind unveils AlphaProof and AlphaGeometry 2, which together solve four of six IMO 2024 problems at silver-medal level in a research evaluation.

Mistral AI has released a large language model Mistral Large 2 in size 123B, which presents capabilities close to the closed SOTA models.

Anthropic releases the Claude Sonnet 3.5 model, which presents better capabilities than other models with low resource usage.

OpenAI announces the GPT-4o model that presents full multimodal capabilities, including receiving and creating text, images, and audio. The model presents an impressive ability to speak with a high response speed and in natural language. The model is 2 times more efficient than the GPT-4 Turbo model, and has better capabilities for languages other than English.

Google announces AI Overviews intended to give a summary of the relevant information in Google search.

Microsoft releases the Phi-3-mini model in open source. The model comes in a reduced version of 3.8B parameters, which allows it to run on mobile devices as well, and it presents capabilities similar to GPT-3.5.

Google announces the Gemini Pro 1.5 multimodal language model capable of parsing up to a million tokens, as well as parsing video and images. The model is gradually released to developers on a waiting list.

OpenAI announces the Sora model that produces videos up to a minute long. The model is not released to the public at this time.

OpenAI's GPT-4 model is partially released, featuring multimodal image analysis and improved multi-language support.

Meta releases the LLaMA language model to approved researchers under a noncommercial research license.

ChatGPT, a chatbot by OpenAI using GPT-3.5, is released to the public and quickly becomes a viral sensation.

Stable Diffusion 1.5 becomes available.

DALL-E 2 is announced for gradual release.