Google launches AlphaGo
DeepMind's AlphaGo beat European champion Fan Hui 5-0, the first program to beat a professional Go player.
2016
Google launches WaveNet
DeepMind's generative model of raw audio produces more natural speech than the best existing text-to-speech systems.
Google launches Neural Machine Translation
Google Translate now uses GNMT for all Chinese-to-English translations, cutting translation errors by 55%-85%.
2017
Google launches AlphaGo Zero
DeepMind's AlphaGo Zero learns Go by self-play alone and beats the champion-defeating AlphaGo 100-0.
2018
OpenAI launches GPT-1
Transformers plus unsupervised pre-training set state-of-the-art results across language tasks; system released.
Google releases BERT
Google open-sources BERT, a bidirectional pre-trained language model, state of the art on 11 NLP tasks.
Google launches AlphaZero
DeepMind's AlphaZero teaches itself chess, shogi and Go, beating a world-champion program in each.
2019
OpenAI launches GPT-2
A 1.5B-parameter language model trained to predict the next word; only a much smaller model was released.
2020
Google releases T5, the text-to-text Transformer
Google Research open-sources T5, an 11B-parameter text-to-text model, with the C4 dataset.
OpenAI launches GPT-3
A 175-billion-parameter language model that performs new tasks from a few examples, without fine-tuning.
Google's AlphaFold 2 solves protein folding
DeepMind's latest AlphaFold scores a median 92.4 GDT at CASP14, recognised as a solution to protein folding.
Google releases Vision Transformer (ViT)
Google Research's ViT applies the Transformer to image patches, beating comparable CNNs with four times less compute.
2021
OpenAI launches CLIP
A neural network that learns visual concepts from natural language and classifies images zero-shot.
OpenAI launches DALL·E
A 12-billion parameter version of GPT-3 that generates images from text captions.
Google launches LaMDA
Google's conversational language model, trained on dialogue, can talk freely about a seemingly endless number of topics.
OpenAI launches Codex
Codex, the model behind GitHub Copilot, translates natural language to code; private beta in the API.
Google launches GLaM
Google Brain's trillion-weight sparse mixture-of-experts (MoE) model; competitive with GPT-3, more efficiently.
2022
OpenAI launches InstructGPT
Models trained with humans in the loop follow instructions better than GPT-3; now the API default.
Google launches PaLM
Google's 540B-parameter model, trained with Pathways on 6,144 TPU chips, sets few-shot records.
Google introduces Gato, a generalist agent
DeepMind's Gato uses one network to play Atari, caption images, chat and stack blocks with a robot arm.
OpenAI releases Whisper
An open-sourced speech recognition model approaching human-level robustness and accuracy on English.
Google launches AlphaCode
DeepMind's AlphaCode, now published in Science, writes programs at a competitive level in coding contests.
Google launches RT-1 Robotics Transformer
Google Research's multi-task robot model, trained on 130k episodes over 700+ tasks; its code is open source.
2023
OpenAI launches GPT-4
Multimodal model accepting image and text inputs and emitting text, with human-level performance on various benchmarks.
Anthropic launches Claude and Claude Instant
Anthropic offers Claude and the lighter, faster Claude Instant through a chat interface and API.
Google launches PaLM 2
Language model with stronger multilingual, reasoning and coding skills, in four sizes, powering 25+ products.
Anthropic launches Claude 2
New model with a 100K token context window, available via API and the claude.ai beta in the US and UK.
Google launches RT-2
A vision-language-action model that learns from web and robot data and turns instructions into robot actions.
Anthropic launches Claude Instant 1.2
A faster, lower-priced model with gains in math, coding, reasoning and safety, available through the API.
Google launches AlphaMissense
Google DeepMind classifies the effects of 71 million missense mutations to help pinpoint causes of disease.
OpenAI launches GPT-4 Turbo
GPT-4 Turbo offers a 128K context window and lower prices, with a Vision version.
Google launches GraphCast
Google DeepMind's weather model gives 10-day forecasts in under a minute, beating HRES on 90%+ of variables.
Google launches Lyria
Google DeepMind's music generation model, tested in YouTube Shorts via Dream Track and in Music AI tools.
Anthropic launches Claude 2.1
Adds a 200K token context window, 2x fewer false statements than Claude 2.0, system prompts and beta tool use.
Google launches GNoME for materials discovery
A deep learning tool predicted 2.2 million new crystals; 380,000 stable ones are released to researchers.
Google launches Gemini 1.0 (Ultra, Pro and Nano)
Google's first natively multimodal models, in three sizes; Ultra leads on 30 of 32 benchmarks at launch.
Mistral releases Mixtral 8x7B
Sparse mixture-of-experts (MoE) model with open weights under Apache 2.0; beats Llama 2 70B on most benchmarks.
2024
Google launches AlphaGeometry
DeepMind's neuro-symbolic system solved 25 of 30 Olympiad geometry problems; code and model open-sourced.
Google launches Imagen 2
Google's upgraded text-to-image model rolls out in Bard, ImageFX, Search and Vertex AI, with SynthID.
Google launches Gemini 1.5
Gemini 1.5 Pro, a mixture-of-experts (MoE) model, matches 1.0 Ultra with less compute and offers up to 1M tokens.
OpenAI announces Sora
A text-conditional diffusion model trained jointly on videos and images to generate video.
Google releases Gemma
Open-weights Gemma 2B and 7B models, built from Gemini research, with a Responsible Generative AI Toolkit.
Anthropic launches the Claude 3 family
Haiku, Sonnet and Opus models, in ascending order of capability, set new benchmarks at launch.
Anthropic releases Claude 3 Haiku
The fastest, most affordable model in its intelligence class, in the API and for Claude Pro users.
xAI releases Grok-1
Open release of the base model weights and architecture of a 314B-parameter mixture-of-experts (MoE) model, under the Apache 2.0 license.
Snowflake releases Snowflake Arctic
Open mixture-of-experts (MoE) model with 480B total and 17B active parameters, built for enterprise tasks such as SQL and coding, under Apache 2.0.
Google launches AlphaFold 3
Predicts structures and interactions of proteins, DNA, RNA and ligands; free via the new AlphaFold Server.
OpenAI launches GPT-4o
New flagship model that reasons across audio, vision and text in real time.
Google launches Veo
Google's most capable video model makes 1080p videos over a minute long; private preview in VideoFX.
Google launches Gemini 1.5 Flash
A lighter, faster model with a 1 million token context window; 1.5 Pro's window grows to 2 million.
Google launches Imagen 3
Google's highest-quality text-to-image model, with better detail and text rendering; preview in ImageFX.
Anthropic launches Claude 3.5 Sonnet
Outperforms Claude 3 Opus on key evaluations at twice the speed.
Google releases Gemma 2
Open models in 9B and 27B sizes, with the 27B competitive with models more than twice its size.
OpenAI launches GPT-4o mini
A smaller, cost-efficient model in the GPT-4o family.
Google launches AlphaProteo
DeepMind's first AI system for designing novel, high-strength protein binders for biology and health research.
OpenAI launches o1
A reasoning model that thinks through a long chain of thought before answering, released in preview.
OpenAI launches o1-mini
A cost-efficient reasoning model alongside o1.
Google releases DataGemma
Open models built on Gemma 2 that ground answers in Google's Data Commons statistics to reduce hallucination.
Google releases AlphaChip
DeepMind names its chip-layout method AlphaChip and releases a pre-trained checkpoint with model weights.
Anthropic launches upgraded Claude 3.5 Sonnet and 3.5 Haiku
A refreshed, more capable Claude 3.5 Sonnet and a new Claude 3.5 Haiku model.
Google launches Genie 2
A foundation world model that generates playable 3D environments from one prompt image, to train agents.
Google releases GenCast
AI ensemble weather model beats ECMWF's ENS up to 15 days ahead; code and weights are released openly.
Google launches Gemini 2.0 Flash
First Gemini 2.0 model, with native image and audio output and tool use; beats 1.5 Pro at twice the speed.
Google launches Veo 2
New video model with up to 4K resolution and improved physics and cinematography, available in VideoFX.
OpenAI launches o1 for developers
o1 comes to developers alongside Realtime API improvements and a new fine-tuning method.
DeepSeek releases DeepSeek-V3
Open-weights mixture-of-experts (MoE) model with 671B parameters, 37B active, trained on 14.8T tokens.
2025
Mistral launches Codestral 25.01
Coding model with a more efficient architecture and tokenizer, completing code about 2x faster.
MiniMax releases the MiniMax-01 series
Open models MiniMax-Text-01 and MiniMax-VL-01: 456B parameters, 45.9B active, with up to 4M tokens of context.
DeepSeek releases DeepSeek-R1
Open reasoning model on par with OpenAI o1, with code and weights under the MIT License.
Alibaba releases Qwen2.5-VL
Vision-language models in 3B, 7B and 72B sizes that read documents and charts and act as visual agents.
Alibaba releases Qwen2.5-1M
Open 7B and 14B models with 1M-token context windows, plus an open-sourced inference framework.
Alibaba launches Qwen2.5-Max
Large mixture-of-experts (MoE) model pretrained on over 20 trillion tokens, in the API and Qwen Chat.
Mistral releases Mistral Small 3
Open 24B model under Apache 2.0, built for low latency and local deployment, with base and instruct checkpoints.
OpenAI launches o3-mini
A new small model in the reasoning line.
Google launches Gemini 2.0 Flash, Flash-Lite and Pro
2.0 Flash is generally available, with 2.0 Pro Experimental (2 million tokens) and Flash-Lite in preview.
Mistral launches Mistral Saba
24B model for Arabic and South Asian languages, via API or deployed on premises.
Anthropic launches Claude 3.7 Sonnet
First hybrid reasoning model: near-instant answers or visible extended thinking.
Alibaba releases QwQ-Max-Preview
Preview reasoning model built on Qwen2.5-Max for math, coding and agent tasks, in Qwen Chat.
OpenAI launches GPT-4.5
Research preview of GPT-4.5, OpenAI's largest and best model for chat at release.
Alibaba releases QwQ-32B
32B reasoning model trained with reinforcement learning, under Apache 2.0; comparable to DeepSeek-R1.
Google launches Gemini Robotics and Gemini Robotics-ER
Two Gemini 2.0-based models: a vision-language-action model that controls robots and a spatial-reasoning model.
Google releases Gemma 3
Open models in 1B, 4B, 12B and 27B sizes with vision, a 128k-token context window and function calling.
Mistral releases Mistral Small 3.1
Open model under Apache 2.0 that adds image understanding and a context window of up to 128k tokens.
NVIDIA releases Llama Nemotron
Open reasoning model family built on Llama models; Nano and Super available now, with Ultra to follow.
OpenAI launches next-generation audio models in the API
New speech-to-text and text-to-speech models in the API; developers can instruct how the voice speaks.
Google launches Gemini 2.5 Pro
The first Gemini 2.5 model, a thinking model that debuts at #1 on LMArena, with a 1 million token context.
Google releases TxGemma
Open models in 2B, 9B and 27B sizes, fine-tuned from Gemma 2 to predict properties of therapeutics.
DeepSeek releases DeepSeek-V3-0324
Update to DeepSeek-V3 with stronger reasoning, front-end development and tool use, now under the MIT License.
Alibaba releases Qwen2.5-Omni
7B end-to-end model that takes text, images, audio and video and answers in text and streaming speech.
Alibaba launches QVQ-Max
Visual reasoning model that analyzes and reasons over images and video, in Qwen Chat.
Quasar Alpha launches in stealth on OpenRouter
Cloaked 1M-context model on OpenRouter, free to test; revealed as an early test of GPT-4.1.
Optimus Alpha launches in stealth on OpenRouter
Second cloaked 1M-context model on OpenRouter; revealed as another test of GPT-4.1.
OpenAI launches GPT-4.1
New API model family with gains in coding, instruction following and long context, plus a first nano model.
OpenAI launches o3 and o4-mini
OpenAI's smartest and most capable models at release, with full tool access.
Google launches Gemini 2.5 Flash
Preview of a hybrid reasoning model with thinking that can be switched on or off, in the Gemini API and the Gemini app.
Alibaba releases Qwen3
Six dense and two mixture-of-experts (MoE) models from 0.6B to 235B parameters, under Apache 2.0.
Mistral launches Mistral Medium 3
Model for enterprise use priced at $0.4/$2 per million tokens, via API or self-deployed on four GPUs.
Google launches AlphaEvolve
A Gemini-powered evolutionary coding agent that discovers and optimizes algorithms for math and computing.
Google launches Veo 3 and Imagen 4
Veo 3 generates video with audio for the first time; Imagen 4 adds 2k resolution and better typography.
Google launches Gemini Diffusion
Experimental text diffusion model, faster than Google's fastest model while matching its coding performance.
Mistral releases Devstral
Open agentic coding model under Apache 2.0, built with All Hands AI, scoring 46.8% on SWE-Bench Verified.
Anthropic launches Claude 4 (Opus 4 and Sonnet 4)
Opus 4 and Sonnet 4 can use tools during extended thinking; Opus 4 is aimed at long agent tasks.
DeepSeek releases DeepSeek-R1-0528
Update to DeepSeek-R1 with better benchmark scores, fewer hallucinations, and JSON output and function calling.
Anthropic launches Claude Gov models
Custom models built for U.S. national security customers in classified environments.
Mistral releases Magistral
Mistral's first reasoning model: open 24B Magistral Small under Apache 2.0, plus Magistral Medium for enterprise.
MiniMax releases MiniMax-M1
Open hybrid-attention reasoning model with a 1M-token context window, trained with the CISPO RL algorithm.
Google launches Gemini 2.5 Flash-Lite and makes 2.5 GA
Gemini 2.5 Flash and Pro become stable and generally available, and 2.5 Flash-Lite launches in preview.
Google launches Gemini Robotics On-Device
A VLA model that runs locally on robots, adapts to new tasks with 50 to 100 demonstrations, via trusted testers.
Google launches AlphaGenome
A DNA sequence model that predicts how variants affect gene regulation, in preview via API for research.
Google releases Gemma 3n
Mobile-first open models, E2B and E4B, taking image, audio, video and text input, now fully released.
Cypher Alpha launches in stealth on OpenRouter
Cloaked 1M-context model on OpenRouter with native tool calling; its lab is not revealed.
Google releases MedGemma 27B Multimodal and MedSigLIP
Open health models: MedGemma 27B Multimodal adds EHR interpretation; MedSigLIP encodes medical images.
Google releases T5Gemma
Open encoder-decoder models adapted from Gemma 2 2B and 9B, plus new T5-sized Small to XL models.
Mistral releases Devstral Small 1.1
Open 24B coding agent model under Apache 2.0, scoring 53.6% on SWE-Bench Verified; Devstral Medium is API-only.
Moonshot AI releases Kimi K2
Open mixture-of-experts (MoE) model with 1T total and 32B active parameters, built for agentic tasks.
Mistral releases Voxtral
Open speech understanding models in 24B and 3B sizes under Apache 2.0, also available on the API.
Alibaba releases Qwen3-Coder
480B mixture-of-experts (MoE) coding model with 35B active, released with the Qwen Code command-line tool.
Alibaba releases Qwen3-235B-A22B-Thinking-2507
Updated 235B thinking model with 256K context, after an Instruct-2507 non-thinking update earlier that week.
Z.ai releases GLM-4.5 and GLM-4.5-Air
Open hybrid reasoning models: GLM-4.5 (355B total, 32B active) and GLM-4.5-Air (106B, 12B active).
NVIDIA releases Llama Nemotron Super v1.5
Open 49B reasoning model that fits on a single H100 GPU, with gains in math, science, coding and tool calling.
Google launches AlphaEarth Foundations
Earth observation model producing embeddings of land and coastal waters, released as an Earth Engine dataset.
Mistral launches Codestral 25.08
Code-completion update with 30% more accepted completions, shipped in Mistral's enterprise coding stack.
Horizon Alpha launches in stealth on OpenRouter
Cloaked 256K-context model on OpenRouter with image input; revealed as an earlier iteration of GPT-5.
Horizon Beta launches in stealth on OpenRouter
Improved version of Horizon Alpha on OpenRouter, 256K context; revealed as an earlier GPT-5.
OpenAI releases gpt-oss-120b and gpt-oss-20b
Two open-weight language models under Apache 2.0, optimized for deployment on consumer hardware.
Anthropic launches Claude Opus 4.1
An upgrade to Opus 4 on agentic tasks, real-world coding and reasoning, at the same price.
Google launches Genie 3
A world model that generates interactive environments from text, navigable in real time at 24 fps and 720p.
OpenAI launches GPT-5
OpenAI's new flagship model, with leading performance across coding, math, writing and health.
Google releases Gemma 3 270M
Compact 270M-parameter open model for task-specific fine-tuning, released pretrained and instruction-tuned.
NVIDIA releases Nemotron Nano 2
Open 9B hybrid Mamba-Transformer reasoning model with 128K context, released with a 6.6T-token pretraining dataset.
DeepSeek releases DeepSeek-V3.1
Hybrid model with thinking and non-thinking modes in one, faster thinking and stronger agent skills.
Google launches Nano Banana
A new image editing model in the Gemini app keeps people's and pets' likeness consistent across edits.
OpenAI launches gpt-realtime
A more advanced speech-to-speech model; the Realtime API adds MCP servers, image input and SIP calls.
Meituan releases LongCat-Flash-Chat
Open-weights 560B-parameter MoE model activating 18.6B to 31.3B (27B on average) per token.
Sonoma Dusk Alpha launches in stealth on OpenRouter
Faster sibling of Sonoma Sky, 2M context; revealed as an early test version of Grok 4 Fast.
Sonoma Sky Alpha launches in stealth on OpenRouter
Cloaked 2M-context model on OpenRouter; revealed as an early test version of Grok 4 Fast.
Alibaba releases Qwen3-Next
80B model with 3B active on a new hybrid-attention architecture, in Instruct and Thinking versions.
Google releases VaultGemma
A 1B open model trained from scratch with differential privacy, with weights on Hugging Face and Kaggle.
DeepSeek releases DeepSeek-V3.1-Terminus
Update to V3.1 with fewer Chinese-English mix-ups and stronger code and search agents.
Alibaba releases Qwen3-Omni
Omni model that takes text, images, audio and video and answers in text and speech, with an audio captioner.
Meituan releases LongCat-Flash-Thinking
Open-weights reasoning model for logic, math, code and agent tasks.
Alibaba releases Qwen3-VL
Vision-language flagship with 235B parameters, 22B active, in Instruct and Thinking versions.
Alibaba launches Qwen3-Max
Model with over 1 trillion parameters, pretrained on 36 trillion tokens, in the API and Qwen Chat.
Google launches Gemini Robotics 1.5 and ER 1.5
Two robotics models that think before acting and call tools; ER 1.5 is available in the Gemini API.
Anthropic launches Claude Sonnet 4.5
Leads on coding, agent building and computer use, and is the most aligned Claude model to date.
DeepSeek releases DeepSeek-V3.2-Exp
Experimental open model with DeepSeek Sparse Attention; API prices cut by more than half.
OpenAI launches Sora 2
Video generation model with synchronized dialogue and sound effects.
Z.ai releases GLM-4.6
Open model update with a 200K-token context window, up from 128K, and stronger coding and reasoning.
Google launches Gemini 2.5 Computer Use model
A specialized model built on Gemini 2.5 Pro that powers agents interacting with user interfaces, in API preview.
Anthropic launches Claude Haiku 4.5
A small model with coding performance similar to Sonnet 4 at one-third the cost and over twice the speed.
Google releases C2S-Scale 27B
A 27B Gemma-based single-cell model, built with Yale, predicted a cancer drug combination confirmed in lab tests.
Google launches Veo 3.1
Veo 3.1 adds richer audio, more narrative control and better realism, with audio across Flow features.
Andromeda Alpha launches in stealth on OpenRouter
Cloaked small reasoning vision model on OpenRouter; revealed as NVIDIA Nemotron Nano 2 VL.
MiniMax releases MiniMax M2
Open agent-first model for coding, tool use and deep search, at 8% of Claude Sonnet's price and twice its speed.
NVIDIA releases Nemotron Nano 2 VL
Open 12B vision-language model with a hybrid Mamba-Transformer design, for document intelligence and video.
OpenAI releases gpt-oss-safeguard
Open-weight reasoning models that classify content against a developer-provided safety policy.
Meituan releases LongCat-Flash-Omni
Open-weights omni-modal model that understands text, images, video and speech, and speaks in real time.
Moonshot AI releases Kimi K2 Thinking
Open-source thinking model that reasons step by step while using tools, with native INT4 inference.
Polaris Alpha launches in stealth on OpenRouter
Cloaked 256K-context model on OpenRouter; revealed as an early GPT-5.1 snapshot.
OpenAI launches GPT-5.1
Upgraded GPT-5 models in ChatGPT with new tone and style controls, rolling out to paid users.
OpenAI brings GPT-5.1 to the API
GPT-5.1 reaches the API with adaptive reasoning, extended prompt caching, and apply_patch and shell tools.
Google launches SIMA 2
Gemini-powered agent that follows instructions, reasons about goals and self-improves in 3D game worlds.
Sherlock Dash Alpha launches in stealth on OpenRouter
Non-reasoning sibling of Sherlock Think; revealed as Grok 4.1 Fast with reasoning disabled.
Sherlock Think Alpha launches in stealth on OpenRouter
Cloaked 1.84M-context reasoning model on OpenRouter; revealed as an early Grok 4.1 Fast.
Google launches WeatherNext 2
AI weather model generating hundreds of forecast scenarios in under a minute, with data in Earth Engine.
Google launches Gemini 3 Pro
Google's most intelligent model, in preview via the Gemini API and Vertex AI, with stronger agentic coding.
OpenAI launches GPT-5.1-Codex-Max
An agentic coding model for Codex built for long-running, project-scale work.
Google launches Nano Banana Pro
Image generation and editing model built on Gemini 3 Pro, with legible multilingual text and up to 4K output.
Sber releases GigaChat 3 Ultra Preview
Open mixture-of-experts (MoE) model with 702B total and about 36B active parameters, natively trained on Russian, under the MIT license.
Anthropic launches Claude Opus 4.5
New flagship model with state-of-the-art coding and agentic performance at lower pricing.
Bert-Nebulon Alpha launches in stealth on OpenRouter
Cloaked multimodal 256K-context model on OpenRouter; revealed as an early Mistral Large 3.
DeepSeek releases DeepSeek-V3.2
Open reasoning-first models built for agents, V3.2 and V3.2-Speciale.
Mistral releases Mistral 3
Open Mistral Large 3, a 675B mixture-of-experts (MoE) model with 41B active, and Ministral 3 in 3B, 8B and 14B; Apache 2.0.
Z.ai releases GLM-4.6V
Open multimodal models, 106B GLM-4.6V and 9B GLM-4.6V-Flash, with native function calling and 128K context.
Mistral releases Devstral 2
Open-weights 123B dense coding model for code agents, with a 256K context window, released with the 24B Devstral Small 2 and the Mistral Vibe CLI.
OpenAI launches GPT-5.2
Frontier model for professional work, improving reasoning, long context, coding and vision, in ChatGPT and API.
NVIDIA releases Nemotron 3
Open hybrid mixture-of-experts (MoE) family; Nano (30B, 3B active) available now, Super and Ultra to follow.
OpenAI launches ChatGPT Images and GPT-Image-1.5
A faster ChatGPT Images experience, with GPT-Image-1.5 available in the API.
Google launches Gemini 3 Flash
Fast, lower-cost Gemini 3 model that surpasses 2.5 Pro on many benchmarks, at under a quarter the cost of 3 Pro.
OpenAI launches GPT-5.2-Codex
Coding model with long-horizon reasoning, large-scale code changes and stronger cybersecurity skills.
Google releases FunctionGemma
Gemma 3 270M fine-tuned for function calling, a base for local agents that turn language into API actions.
Google releases T5Gemma 2
Open encoder-decoder models based on Gemma 3, now multimodal with 128K context, in sizes up to 4B-4B.
Z.ai releases GLM-4.7
Open model focused on agentic coding, scoring 73.8% on SWE-bench and 41% on Terminal Bench 2.0.
MiniMax releases MiniMax M2.1
Open model update with stronger coding across languages such as Rust, Java, Go and C++, and in office tasks.
2026
Google releases MedGemma 1.5 and MedASR
Open MedGemma 1.5 4B adds CT, MRI and histopathology support; new open MedASR model transcribes medical speech.
Google releases TranslateGemma
Open translation models built on Gemma 3 in 4B, 12B and 27B sizes, covering 55 languages.
Meituan releases LongCat-Flash-Thinking-2601
Open-weights upgrade of LongCat-Flash-Thinking, stronger at agentic search and agentic tool use.
Alibaba launches Qwen3-Max-Thinking
Flagship reasoning model that calls search, memory and a code interpreter on its own; in Qwen Chat and API.
Moonshot AI releases Kimi K2.5
Open visual agentic model trained on 15T tokens, with an agent swarm for parallel tasks.
Alibaba releases Qwen3-Coder-Next
Model for coding agents and local use, built on Qwen3-Next-80B-A3B; over 70% on SWE-Bench Verified.
Anthropic launches Claude Opus 4.6
Upgrade to the smartest Claude model, leading on agentic coding, computer use, tool use, search and finance.
OpenAI launches GPT-5.3-Codex
A Codex-native agent model pairing frontier coding performance with general reasoning.
Pony Alpha launches in stealth on OpenRouter
Cloaked coding and agentic model on OpenRouter; revealed as an early testing version of GLM-5.
Aurora Alpha launches in stealth on OpenRouter
Cloaked reasoning model built for speed on OpenRouter, 128K context; its lab is not revealed.
Meituan releases LongCat-Flash-Lite
Lightweight 68.5B MoE model, 2.9B to 4.5B active, with over 30B parameters in N-gram embeddings.
MiniMax releases MiniMax M2.5
Open model for coding, agentic tool use, search and office work, scoring 80.2% on SWE-Bench Verified.
Z.ai releases GLM-5
Open model under MIT with 744B total and 40B active parameters, using DeepSeek Sparse Attention.
OpenAI launches GPT-5.3-Codex-Spark
A real-time coding model with 15x faster generation, in research preview for ChatGPT Pro.
Google upgrades Gemini 3 Deep Think
Upgraded Deep Think reasoning mode for Google AI Ultra subscribers, with early API access for select users.
Alibaba releases Qwen3.5
Native vision-language model with 397B parameters, 17B active, supporting 201 languages and dialects.
Anthropic launches Claude Sonnet 4.6
Full Sonnet upgrade across coding, computer use, long reasoning, agent planning, knowledge work and design.
Google launches Lyria 3
Music model that makes 30-second tracks with lyrics from text or photos, in beta in the Gemini app.
Google launches Gemini 3.1 Pro
Upgraded reasoning model in preview across Gemini API, Vertex AI, Gemini app and NotebookLM; ARC-AGI-2 77.1%.
Google launches Nano Banana 2
Image model with Nano Banana Pro's world knowledge and quality at Flash speed, in Gemini, Search and Flow.
OpenAI launches GPT-5.3 Instant
GPT-5.3 Instant targets smoother, more useful everyday conversations in ChatGPT.
Google launches Gemini 3.1 Flash-Lite
Fastest, most cost-efficient Gemini 3 series model, in preview in the Gemini API and Vertex AI.
OpenAI launches GPT-5.4
Frontier model for professional work with stronger coding, computer use, tool search and 1M-token context.
Google launches Gemini Embedding 2
Google's first natively multimodal embedding model: text, images, video, audio and documents in one space.
NVIDIA releases Nemotron 3 Super
Open 120B hybrid mixture-of-experts (MoE) model, 12B active, with a 1M-token context window for agents.
Healer Alpha launches in stealth on OpenRouter
Cloaked omni-modal model taking text, image, audio and video; revealed as Xiaomi MiMo-V2-Omni.
Hunter Alpha launches in stealth on OpenRouter
Cloaked 1T-parameter, 1M-context agentic model on OpenRouter; revealed as Xiaomi MiMo-V2-Pro.
Mistral releases Leanstral
Open-source code agent for the Lean 4 proof assistant, with 6B active parameters and weights under Apache 2.0.
Mistral releases Mistral Small 4
Open 119B mixture-of-experts (MoE) model, 6B active, uniting reasoning, image input and agentic coding; Apache 2.0.
OpenAI launches GPT-5.4 mini and nano
Smaller, faster GPT-5.4 versions for coding, tool use and high-volume API and sub-agent workloads.
MiniMax releases MiniMax M2.7
Model that helped run its own reinforcement learning experiments, scoring 56.22% on SWE-Pro.
Google launches Gemini 3.1 Flash Live
Google's highest-quality audio model, with lower latency, in the Live API, Gemini Live and Search Live.
Alibaba launches Qwen3.5-Omni
Omni model in Plus, Flash and Light sizes for text, image, audio and video input, with speech in 36 languages.
Google releases Gemma 4
Open models in four sizes under Apache 2.0, built for reasoning and agentic workflows, with function calling.
Alibaba launches Qwen3.6-Plus
Hosted model with a 1M-token context window by default and stronger agentic coding, via API.
Meituan releases LongCat-Next
Native multimodal model that takes images and speech as discrete tokens; model and tokenizer open-sourced.
Anthropic launches Claude Mythos Preview
A general-purpose model strikingly capable at security work, used in Project Glasswing to secure key software.
Z.ai releases GLM-5.1
Open model under MIT for agentic engineering, built to stay effective on long-horizon coding tasks.
Elephant Alpha launches in stealth on OpenRouter
Cloaked 100B-parameter text model, 256K context; revealed as inclusionAI's Ling-2.6-flash.
OpenAI launches GPT-5.4-Cyber for vetted defenders
Trusted Access for Cyber expands, giving vetted security defenders a cyber-tuned GPT-5.4.
Google launches Gemini Robotics-ER 1.6
Upgraded robotics reasoning model with better spatial understanding, in the Gemini API and AI Studio.
Alibaba releases Qwen3.6-35B-A3B
Mixture-of-experts (MoE) model with 35B parameters, 3B active, for agentic coding; 73.4% on SWE-bench Verified.
Anthropic launches Claude Opus 4.7
Generally available successor to Opus 4.6, with notable gains on the hardest software engineering tasks.
OpenAI launches GPT-Rosalind
A reasoning model for life sciences: drug discovery, genomics analysis and protein reasoning.
Alibaba launches Qwen3.6-Max-Preview
Preview of the next proprietary Max model, with stronger agentic coding than Qwen3.6-Plus, in Qwen Studio.
Moonshot AI releases Kimi K2.6
Open-source model built for long-horizon coding and agent swarms.
OpenAI releases Privacy Filter
An open-weight model that detects and redacts personally identifiable information in text.
OpenAI launches GPT-5.5
A model built for complex tasks like coding, research and data analysis across tools.
DeepSeek releases DeepSeek V4 Preview
Open V4-Pro (1.6T total, 49B active parameters) and V4-Flash models, with 1M-token context by default.
NVIDIA releases Nemotron 3 Nano Omni
Open 30B mixture-of-experts (MoE) model, 3B active, taking text, image, audio and video input for agents.
Owl Alpha launches in stealth on OpenRouter
Cloaked agentic model on OpenRouter with 1M context; revealed as Meituan's LongCat-2.0.
OpenAI launches GPT-5.5 Instant
ChatGPT's default model updates, with fewer hallucinations and improved personalization controls.
OpenAI launches GPT-5.5-Cyber
Trusted Access for Cyber expands to GPT-5.5 and GPT-5.5-Cyber for verified defenders.
OpenAI launches realtime voice models in the API
New realtime API voice models that reason, translate and transcribe speech.
Google launches Gemini 3.5 Flash
First Gemini 3.5 model, Google's strongest for agents and coding, in the Gemini app, AI Mode and the API.
Google launches Gemini Omni Flash
First Gemini Omni model: makes video from any mix of image, audio, video and text, and edits it by chat.
Alibaba launches Qwen3.7-Max
Proprietary model for agent work: coding, office workflows and autonomous runs of over 1,000 tool calls.
Mistral releases Mistral Medium 3.5
Open-weights 128B dense flagship model merging instruction-following, reasoning and coding, with a 256k context window, under a modified MIT license.
Anthropic launches Claude Opus 4.8
Stronger coding, agentic and professional work, with more consistency on long-running tasks.
MiniMax releases MiniMax M3
Open-weight model for coding and agents, with sparse attention, 1M-token context and native multimodality.
Alibaba launches Qwen3.7-Plus
Multimodal agent model that reads screens, operates GUIs and writes code from images, via API.
Google releases Gemma 4 12B
Encoder-free multimodal open model with native audio input that runs locally on laptops with 16GB of memory.
NVIDIA releases Nemotron 3 Ultra
Open 550B mixture-of-experts (MoE) model, 55B active, for long-running agents; Nemotron moves to OpenMDW-1.1.
Anthropic launches Claude Fable 5 and Mythos 5
Fable 5, a Mythos-class model made safe for general use; Mythos 5 goes to cyberdefenders.
Google releases DiffusionGemma
Experimental open 26B MoE text diffusion model under Apache 2.0, with up to 4x faster generation on GPUs.
Z.ai releases GLM-5.2
Open model under MIT for long-horizon tasks, with a 1M-token context window and stronger coding.
Alibaba releases Qwen-AgentWorld
Language world model that simulates agent environments across seven domains, released with AgentWorldBench.
OpenAI previews GPT-5.6 Sol
Preview of a next-generation model stronger in coding, science and cybersecurity.
Meituan releases LongCat-2.0
1.6T-parameter MoE model, ~48B active, with 1M context, trained and served on AI ASIC superpods.
Anthropic launches Claude Sonnet 5
The most agentic Sonnet yet, aimed at coding and everyday professional work.
Poolside releases Laguna XS 2.1
Open mixture-of-experts (MoE) coding model with 33B total and 3B active parameters, built for agentic coding on a local machine, under the OpenMDW-1.1 license.
Mistral releases Leanstral 1.5
Open Lean 4 proof model, 119B total and 6B active, under Apache 2.0; solves 587 of 672 PutnamBench problems.
OpenAI launches GPT-Live
A new generation of voice models for natural conversation, now powering ChatGPT Voice.
OpenAI launches GPT-5.6
GPT-5.6 family reaches general availability: flagship Sol, balanced Terra and cost-efficient Luna.
Thinking Machines releases Inkling
The lab's first open-weights model: a multimodal mixture-of-experts (MoE) model with 975B total and 41B active parameters and up to 1M tokens of context.
Moonshot AI releases Kimi K3
Open model with 2.8T parameters, the first open model at that scale, with native vision and a 1M-token context window.
Google launches Gemini 3.5 Flash Cyber
A lightweight model for finding and patching vulnerabilities, piloted with governments and trusted partners.
Google launches Gemini 3.6 Flash and 3.5 Flash-Lite
3.6 Flash improves coding and knowledge work with 17% fewer output tokens; 3.5 Flash-Lite is the fastest 3.5.
Anthropic launches Claude Opus 5
A step up for the Opus tier on long-running agents, coding and professional work.
Google launches Lyria 3.5
Music model with better musicality, lyrics and vocals, plus tempo and duration control, in Google Flow Music.
Google launches Gemini Robotics 2
Three robotics models add whole-body humanoid control, finer dexterity and multi-robot teamwork.
Alibaba releases Qwen3.8-Max
2.4T-parameter model with 95B active, the first Qwen Max-class model to get open weights.
OpenAI gives free users GPT-5.6 Luna
Improved GPT-5.6 Sol in ChatGPT; free users get unlimited everyday chats with GPT-5.6 Luna.
OpenAI launches GPT-5.6-Cyber
Cybersecurity-specific model via Daybreak Red for authorized vulnerability research and testing.
NVIDIA releases Nemotron 3.5 Lightning
Open 30B mixture-of-experts (MoE) model for agentic workloads, released with the NeMo Switchyard routing library.
Google launches Gemini 3.7 Flash
Flash model for coding and agents, with an introductory price of $0.75/1M input and $3.75/1M output tokens.
DeepSeek makes DeepSeek-V4-Pro generally available
V4-Pro reaches general availability with agent upgrades, adjustable reasoning effort and Responses API support.
Z.ai releases GLM-5.3
Coding model post-trained from the GLM-5.2 base, state of the art on CyberGym; weights to follow in two weeks.
Ox Alpha launches in stealth on OpenRouter
Cloaked reasoning model on OpenRouter for coding and agentic work, 1M context; revealed as Z.ai's GLM-5.3-Flash.
DeepSeek releases DeepSeek-V4-Flash-Vision-Exp
Experimental multimodal V4-Flash model that adds image input for agents, matching V4-Flash on text.
Alibaba releases Qwen3.8-Flash-Next
Preview of the Qwen4 architecture: 125B mixture-of-experts (MoE) model with 6B active, plus 51B n-gram embeddings.
Z.ai releases GLM-5.3-Flash
Natively multimodal open model with 320B total and 18B active parameters, at a tenth of GLM-5.2's price.
Google launches Gemini Omni 1.1 Flash
Adds scene extension to 40 seconds, first and last frame control, 360p drafts and 4K upscaling.
Anthropic launches Claude Fable 5.1 and Mythos 5.1
One model with two safeguard levels: Fable 5.1 generally available, Mythos 5.1 only via trusted access programs.
Google launches Gemini 3.8 Flash and 3.8 Flash Cyber
3.8 Flash improves coding and reasoning at 3.7's price; 3.8 Flash Cyber goes to trusted defenders via Fairwind.
Alibaba launches Qwen3.8-Max-0902
Refresh of Qwen3.8-Max with post-training for coding and workplace tasks, available via API on QwenCloud and in QwenWork.
OpenAI launches GPT-6 Astra
OpenAI's most intelligent model yet, state of the art in computer use, coding, cybersecurity and science.
Google launches WeatherNext 3
Hourly global forecasts at up to 5 km resolution, trained on live satellite data, in Search, Gemini and Maps.
DeepSeek releases DeepSeek-V4.1-Flash
Open 552B MoE model with a new causal encoder-decoder architecture and native vision; DeepSeek phases out V4-Pro.
OpenAI brings GPT-Live-1 to the API
Full-duplex voice conversations in the API, with custom voices and telephony support.
Google launches Gemini 3.8 Live
Two live voice models for conversation; the Extended Thinking one reasons while speaking, in the Gemini API.
Union Alpha launches in stealth on OpenRouter
Cloaked multimodal model on OpenRouter for research, coding and agentic work; revealed as Unbiased's Pareto.
Alibaba launches Qwen3.8-Omni-Flash
Omni model with a 1M-token context for audio and video agent tasks, such as editing and live conversation.
Anthropic launches Claude Opus 5.5
Leads in agentic coding and knowledge work, and costs 40% less to run than Opus 5 on typical workloads.
OpenAI launches GPT-6 Sol and Luna
Two GPT-6 models with different balances of capability and cost for everyday work.
Space Bunny Alpha launches in stealth on OpenRouter
Cloaked fast coding model on OpenRouter with native multimodal input; revealed as MiniMax M3.1 Flash.
Anthropic launches Claude Sonnet 5.5
Clear upgrade over Sonnet 5: 30%+ faster and up to 30% cheaper for most work.
OpenAI launches GPT-6.1 Sol
Near-Astra coding and computer use at one-fifth of Astra's standard API token prices.
Google launches Gemini 4 Argon
Frontier model for coding, knowledge work and cyber defense, rolling out first to trusted cyber defenders.
Mistral releases Mistral Large 4 (public preview)
1T-parameter multimodal model with 52B active; the API preview is out and weights are promised by month end.
Anthropic launches Claude Haiku 5.5
Fastest small Claude model, built for high-volume work like summarization, subagents and browser use.
Google makes Gemini 4 Argon generally available
Gemini 4 Argon reaches developers, enterprises and AI Pro and Ultra users in the API, Antigravity and app.
NVIDIA releases Nemotron 3.5 Super VL and Bolt
Open 120B multimodal model for always-on agents, with the Bolt 74B-A9B edge model; a 4GB Bolt E4B follows.
DeepSeek releases DeepSeek-V4.1-Flash-1017
Open update of V4.1-Flash, trained with self-feedback RL for end-to-end project development, with shorter outputs.
MiniMax releases MiniMax M3.1 Flash
Flash model for agents, coding and computer use, nearly three times faster at 70% lower usage cost; weights followed.
NVIDIA releases Nemotron 3.5 Bolt E4B
Open 4GB small model built on Gemma 4 E4B for conversational agents on RTX, Jetson and DGX Spark devices.
Moonshot AI releases Kimi K3.1 Pro
Post-trained Kimi K3 and Moonshot's most capable 3T-class open-weight model, with a 1M-token context window and a faster Highspeed variant.
OpenAI launches GPT-6.1 Terra
A fast, efficient model for computer use and writing, with API prices 50% below GPT-5.6 Terra and GPT-6.1 Sol.
No events match these filters.