Artificial Intelligence Timeline

Astek innovation events in the Artificial Intelligence domain.

Mixtral 8x7B (2024-01-08) — Mistral AI launches Mixtral 8x7B, an open-source Mixture of Experts model that rivals GPT-3.5. This model uses an innovative architecture that activates only 2 out of 8 experts per token, optimizing performance and efficiency.
GenAI Meetup #1 (2024-01-18) — First ASTEK meetup dedicated to generative AI. Presentation of LLM fundamentals and demonstration of enterprise use cases. Over 50 participants!
Sora: OpenAI unveils realistic video generation (2024-02-15) — OpenAI unveils Sora, a model that turns a plain text prompt into realistic videos up to one minute long. Access is initially limited to red teamers and a handful of creative professionals.
Google Gemini 1.5 Pro (2024-02-15) — Google unveils Gemini 1.5 Pro with a revolutionary context window of 1 million tokens. This breakthrough enables processing entire documents, hours of video, or complete codebases in a single query.
Claude 3 (Opus, Sonnet, Haiku) (2024-03-04) — Anthropic launches the Claude 3 family with three models: Opus (the most powerful), Sonnet (balanced), and Haiku (fast). Claude 3 Opus outperforms GPT-4 on many benchmarks and introduces multimodal vision capabilities.
The European Parliament adopts the AI Act (2024-03-13) — Meeting in Strasbourg, MEPs adopt the Artificial Intelligence Act by 523 votes to 46, with 49 abstentions: the world's first binding law dedicated to AI clears its decisive stage.
Documentary RAG POC (2024-03-15) — The iLab team is developing a RAG (Retrieval Augmented Generation) proof of concept for internal document search. Using LangChain, ChromaDB, and Claude 3.
New term: agentic (2024-03-21) — Describes a system where a language model does not answer in one pass but works through steps: it plans, uses tools, reviews and revises its own output, alone or with other agents.
Meta Llama 3 (2024-04-18) — Meta releases Llama 3 in 8B and 70B parameter versions, setting new standards for open-source models. The models are available on Hugging Face and can be freely fine-tuned.
New term: SLM (2024-04-23) — An SLM (small language model) has at most a few billion parameters. It targets on-device, offline or low-cost uses where a large model would be oversized.
AlphaFold 3 (Google DeepMind and Isomorphic Labs) (2024-05-08) — Google DeepMind and Isomorphic Labs publish AlphaFold 3 in Nature. It predicts the structure and interactions of nearly all of life's molecules: proteins, DNA, RNA and ligands. A free server opens it to researchers.
New term: AI slop (2024-05-08) — Low-quality AI-generated content produced in bulk, published unreviewed and pushed on readers who did not ask for it. Slop is to AI what spam is to email.
GPT-4o (Omni) (2024-05-13) — OpenAI unveils GPT-4o ("o" for "omni"), a single model trained across text, vision and audio that answers by voice in 320 ms on average, matches GPT-4 Turbo at half the price and becomes free in ChatGPT.
Apple Intelligence unveiled at WWDC (2024-06-10) — At WWDC, Apple unveils Apple Intelligence, its generative AI system built into iPhone, iPad and Mac. It combines on-device models, a dedicated private cloud and a ChatGPT integration.
Nvidia becomes the world's most valuable company (2024-06-18) — Driven by demand for AI chips, Nvidia reaches a market value of 3.335 trillion USD and overtakes Microsoft: the GPU supplier becomes, for a few days, the world's most valuable listed company.
Claude 3.5 Sonnet (2024-06-20) — Anthropic launches Claude 3.5 Sonnet, the first model of the 3.5 family: it outperforms Claude 3 Opus on most evaluations, at 2x the speed and a fifth of the price, and introduces Artifacts in Claude.ai.
ASTEK AI Hackathon (2024-06-28) — First internal hackathon dedicated to AI. 48 hours to develop innovative solutions using LLMs. 8 teams, 32 participants, 3 projects awarded.
Llama 3.1 405B: an open model on par with the best closed models (2024-07-23) — Meta releases Llama 3.1, whose 405B version is the first open-weight model presented as on par with GPT-4o and Claude 3.5 Sonnet. The gap between open and closed models narrows sharply.
Mistral Large 2 (2024-07-24) — Mistral AI launches Mistral Large 2 (123B parameters), a frontier model competing with GPT-4o and Claude 3.5 Sonnet. Supports 80+ languages and features a 128k token context window.
The AI Act enters into force (2024-08-01) — Regulation (EU) 2024/1689 on artificial intelligence enters into force on August 1, 2024. It starts a staggered application timeline running to 2027, from bans to obligations for high-risk systems.
OpenAI o1 (Strawberry) (2024-09-12) — OpenAI launches the o1 series, models trained with reinforcement learning to think before they answer. o1-preview and o1-mini ship in ChatGPT and the API; o1 reaches 83% on AIME 2024 versus 13% for GPT-4o.
New term: reasoning model (2024-09-12) — A reasoning model is trained to "think" before answering by producing a chain of thought. Test-time compute is the extra computation this takes, which became a new lever for performance.
Microsoft brings back the Three Mile Island nuclear plant (2024-09-20) — Constellation Energy signs a 20-year power purchase agreement with Microsoft to restart Three Mile Island Unit 1, shut down in 2019. Energy becomes a direct stake in the AI race.
OpenAI raises 6.6 billion USD at a 157 billion USD valuation (2024-10-02) — OpenAI closes a 6.6 billion USD round valuing it at 157 billion USD, one of the largest ever raised by a private company. The money is to fund research and computing power.
Nobel Prizes in Physics and Chemistry: AI recognized (2024-10-09) — On October 8, the Nobel Prize in Physics goes to John Hopfield and Geoffrey Hinton for the foundations of neural networks; on October 9, the Chemistry prize honors David Baker, Demis Hassabis and John Jumper for protein design and prediction.
Claude Computer Use (2024-10-22) — Anthropic opens computer use in public beta: Claude looks at the screen, moves the cursor, clicks and types. Announced with an upgraded Claude 3.5 Sonnet (49% on SWE-bench Verified) and Claude 3.5 Haiku.
Model Context Protocol (MCP): Anthropic releases the open standard for agents (2024-11-25) — Anthropic releases the Model Context Protocol, an open standard for connecting AI assistants to tools and data. Within a year, MCP becomes the reference protocol for agents, adopted by OpenAI, Google and Microsoft.
Google Gemini 2.0 Flash (2024-12-11) — Google launches Gemini 2.0 Flash with native agent capabilities: image generation, text-to-speech, code execution. Available via the API and in AI Studio.
New term: deep research (2024-12-11) — Deep research refers to an agent that runs a multi-step web investigation on its own for several minutes, then writes a structured report with sources.
New term: agents vs workflows (2024-12-19) — Anthropic separates workflows, where code chains model calls along predefined paths, from agents, where the model decides its own steps and tools. The rule: start with the simplest solution.
o3 scores 75.7% on ARC-AGI (2024-12-20) — OpenAI unveils o3, its new reasoning model, which scores 75.7% on the ARC-AGI test within the competition's compute limit, and 87.5% with 172 times more compute. GPT-4o topped out at 5%.
Codestral 25.01 (2025-01-13) — Mistral AI launches Codestral 25.01, an update to its code model with a more efficient architecture and tokenizer, roughly twice as fast in generation and completion, with a 256k-token context.
MiniMax-Text-01 (and MiniMax-VL-01) (2025-01-15) — MiniMax releases MiniMax-Text-01 with open weights, a 456-billion-parameter MoE that combines Lightning linear attention and softmax attention, with a 4-million-token context at inference.
Kickoff Innovation 2025 (2025-01-16) — Official launch of ASTEK's Innovation 2025 program. Presentation of the AI roadmap and annual objectives.
DeepSeek R1 (2025-01-20) — DeepSeek releases R1, a 671-billion-parameter reasoning model (37 billion active) under the MIT license, whose announced scores match those of OpenAI o1, along with six distilled models.
Gemini 2.0 Flash Thinking Experimental (01-21) (2025-01-21) — Google releases an update to the Gemini 2.0 Flash Thinking Experimental reasoning model, with the context window raised to 1 million tokens and native code execution.
Stargate: up to $500B in AI infrastructure (2025-01-21) — At the White House, OpenAI, SoftBank, Oracle and MGX announce Stargate, a joint venture planning to invest up to $500 billion over four years in AI data centers in the United States.
Computer-Using Agent (CUA) (2025-01-23) — OpenAI introduces Computer-Using Agent (CUA), a model that combines GPT-4o's vision with reinforcement-learning-trained reasoning to operate a graphical interface. It powers the Operator agent.
Qwen2.5-VL (2025-01-26) — Alibaba Qwen releases Qwen2.5-VL in three sizes (3B, 7B, 72B): an open-weights vision-language model able to act as a visual agent, understand hour-long videos and produce structured outputs.
DeepSeek shock: Nvidia loses 17% in one session (2025-01-27) — A week after DeepSeek R1's release, markets question the need for massive chip spending: Nvidia falls about 17% and loses nearly $589 billion in market value, a record for a single session.
Qwen2.5-1M (2025-01-27) — Alibaba Qwen releases Qwen2.5-7B-Instruct-1M and Qwen2.5-14B-Instruct-1M, two open-weights models accepting up to 1 million tokens, with an adapted vLLM inference framework and sparse attention.
Mistral Small 3 (2025-01-30) — Mistral AI releases Mistral Small 3, a 24-billion-parameter model optimized for latency, under the Apache 2.0 license, presented as comparable to Llama 3.3 70B while being more than 3 times faster.
Tülu 3 405B (2025-01-30) — Ai2 releases Tülu 3 405B, a model derived from Llama 3.1 405B and post-trained with the open Tülu 3 recipe (SFT, DPO, RLVR), which Ai2 claims rivals DeepSeek V3 and GPT-4o.
o3-mini (2025-01-31) — OpenAI launches o3-mini, its most cost-efficient compact reasoning model, optimized for science, mathematics and code, available in ChatGPT and the API the same day.
New term: vibe coding (2025-02-02) — Vibe coding means building software by describing what you want to an AI in natural language and accepting the generated code without really reading it, judging only the result.
AI Act: first prohibitions apply (2025-02-02) — Six months after its entry into force, the EU AI Act starts to apply: practices deemed unacceptable are banned and organizations must ensure AI literacy among their staff.
Deep research (2025-02-02) — OpenAI launches deep research in ChatGPT, an agent that carries out multi-step web research with a version of o3 optimized for browsing and data analysis.
Gemini 2.0 Pro Experimental (2025-02-05) — Google releases an experimental version of Gemini 2.0 Pro, its best model at the time for coding and complex prompts, with a context window of 2 million tokens.
Gemini 2.0 Flash (general availability) (2025-02-05) — Google makes Gemini 2.0 Flash generally available in the Gemini API, Google AI Studio and Vertex AI, with higher rate limits and simplified pricing.
Gemini 2.0 Flash-Lite (2025-02-05) — Google launches Gemini 2.0 Flash-Lite, its most cost-efficient model at this stage: better than 1.5 Flash at the same speed and cost, in public preview on February 5, then available in production on the 25th.
AI Action Summit in Paris (2025-02-10) — On February 10-11, 2025, Paris hosts the AI Action Summit, co-chaired with India. The day before, Emmanuel Macron announces €109 billion of private investment in France, mostly in data centers.
Mistral Saba (2025-02-17) — Mistral AI introduces Mistral Saba, its first regional model: 24 billion parameters trained on data from the Middle East and South Asia, strong in Arabic and Indian languages such as Tamil.
Grok 3 and Grok 3 mini (2025-02-19) — xAI introduces Grok 3, its flagship model trained on the Colossus supercomputer, and Grok 3 mini, a cost-efficient reasoning variant, with the Think and DeepSearch modes.
Moonlight (Muon at scale) (2025-02-22) — Moonshot AI releases Moonlight, a 16 billion parameter MoE (3 billion active) trained with the Muon optimizer, with code, checkpoints and a technical report under the MIT license.
Claude Code: the start of agentic coding (2025-02-24) — Alongside Claude 3.7 Sonnet, Anthropic opens a research preview of Claude Code, an agent that works in the terminal: it reads code, edits files, runs tests and pushes to GitHub.
Claude 3.7 Sonnet (2025-02-24) — Anthropic launches Claude 3.7 Sonnet, its first hybrid reasoning model: fast answers or visible extended thinking, with an adjustable budget in the API. Claude Code arrives in preview.
GPT-4.5 (2025-02-27) — OpenAI releases GPT-4.5 as a research preview, its largest chat model, based on scaling up pretraining, with better intuition, fewer hallucinations and a higher “EQ”.
QwQ-32B (2025-03-06) — Alibaba Qwen releases QwQ-32B, a 32 billion parameter reasoning model under Apache 2.0, trained with reinforcement learning and presented as comparable to DeepSeek-R1 (671 billion parameters).
Mistral OCR (2025-03-06) — Mistral AI launches Mistral OCR, a document recognition API that extracts text, tables, equations and images from PDFs and images, with an overall score of 94.89% on its internal benchmark, at 1 dollar per 1000 pages.
Gemini Embedding (2025-03-07) — Google launches Gemini Embedding, a text embedding model derived from Gemini: an experimental version on March 7, 2025, ranked first on the multilingual MTEB leaderboard, then a stable version on July 14.
computer-use-preview and gpt-4o-search-preview (2025-03-11) — OpenAI opens the Responses API with three specialized models for agents: computer-use-preview to operate a computer, gpt-4o-search-preview and gpt-4o-mini-search-preview for web search.
Gemma 3 (2025-03-12) — Google DeepMind releases Gemma 3, its new family of open models based on Gemini 2.0 technology: four sizes from 1 to 27 billion parameters, multimodal, with a 128,000-token context.
Gemini 2.0 Flash (native image generation) (2025-03-12) — Google opens Gemini 2.0 Flash's native image generation and editing to developers in an experimental version, with text and images interleaved in a single response.
Gemini Robotics and Gemini Robotics-ER (2025-03-12) — Google DeepMind introduces Gemini Robotics, a vision-language-action model based on Gemini 2.0 that directly controls robots, and Gemini Robotics-ER, dedicated to spatial reasoning.
OLMo 2 32B (2025-03-13) — Ai2 releases OLMo 2 32B, a fully open 32-billion-parameter model (weights, data, code) that, according to Ai2, outperforms GPT-3.5 Turbo and GPT-4o mini on academic benchmarks.
Mistral Small 3.1 (2025-03-17) — Mistral AI unveils Mistral Small 3.1, a 24-billion-parameter model under the Apache 2.0 license that adds image understanding and a 128k-token context to Mistral Small 3.
Llama Nemotron (Nano, Super, Ultra) (2025-03-18) — NVIDIA launches the Llama Nemotron family at GTC, reasoning models derived from Llama in three sizes (8B, 49B, 253B), released with their weights, their post-training dataset and their code.
o1-pro (2025-03-19) — OpenAI makes o1-pro available in the API: a version of o1 that devotes more compute to each response, reserved for the Responses API and billed at a very high price.
gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-mini-tts (2025-03-20) — OpenAI launches new speech recognition models (gpt-4o-transcribe, gpt-4o-mini-transcribe) and a steerable speech synthesis model (gpt-4o-mini-tts) in the API to build voice agents.
Autonomous Agents POC v2 (2025-03-20) — New version of the Agents POC with multi-agent orchestration and MCP integration. Demonstration of concrete use cases in enterprise environments.
Nemotron-H (2025-03-21) — NVIDIA releases Nemotron-H, a family of hybrid Mamba-2 and Transformer models (8B, 47B and 56B) that replaces most of the attention with Mamba layers to speed up inference.
DeepSeek V3-0324 (2025-03-24) — DeepSeek releases V3-0324, an update to its 685-billion-parameter generalist model, with a clear gain in reasoning, web development and tool calling, now under the MIT license.
GPT-4o Image Generation (2025-03-25) — OpenAI builds native image generation into GPT-4o in ChatGPT: a multimodal model that renders text accurately, follows detailed instructions and refines its images over the course of the conversation.
Gemini 2.5 Pro Experimental (2025-03-25) — Google launches Gemini 2.5 Pro Experimental, the first model of the 2.5 family: a reasoning model ranked first on LMArena, strong in code, mathematics and science, with 1 million tokens of context.
TxGemma (2025-03-25) — Google releases TxGemma, a collection of open models derived from Gemma 2 for therapeutic development: prediction of molecule properties and conversational discussion, in sizes of 2, 9 and 27 billion parameters.
Qwen2.5-Omni (2025-03-27) — Alibaba Qwen releases Qwen2.5-Omni-7B, an end-to-end multimodal open-weights model: it perceives text, images, audio and video and responds in text or in speech synthesized as a continuous stream.
Llama 4 (Scout, Maverick, Behemoth in preview) (2025-04-05) — Meta releases Llama 4 Scout and Llama 4 Maverick, its first open-weights, natively multimodal mixture-of-experts models, and previews Behemoth, a 2-trillion-parameter model still in training.
Google launches A2A, the Agent2Agent protocol (2025-04-09) — At Google Cloud Next, Google unveils Agent2Agent (A2A), an open protocol that lets AI agents from different vendors discover each other, exchange information and coordinate tasks.
Kimi-VL (A3B Instruct and Thinking) (2025-04-10) — Moonshot AI releases Kimi-VL, a 16-billion-parameter MoE vision-language model (2.8 billion active in the decoder), with a reasoning variant, under the MIT license and with a 128K context.
GLM-4-0414 (GLM-4-32B, GLM-Z1, GLM-Z1-Rumination) (2025-04-14) — Zhipu AI released the GLM-4-0414 family under the MIT license: GLM-4-32B, the GLM-Z1 reasoning models (32B and 9B) and GLM-Z1-Rumination, presented as comparable to GPT-4o and DeepSeek-V3/R1.
GPT-4.1, GPT-4.1 mini and GPT-4.1 nano (2025-04-14) — OpenAI launches the GPT-4.1 family (GPT-4.1, mini and nano) in the API, with better performance in code and instruction following, a one-million-token context and a first nano model.
DolphinGemma (2025-04-14) — Google, with Georgia Tech and the Wild Dolphin Project, presents DolphinGemma, a model of about 400 million parameters trained on dolphin sounds and light enough to run on Pixel smartphones.
OpenAI o3 and o4-mini (2025-04-16) — OpenAI launches o3, its most powerful reasoning model, and o4-mini, a fast and economical version. For the first time, these models can combine all of ChatGPT's tools and reason about images.
Gemini 2.5 Flash (preview) (2025-04-17) — Google launches Gemini 2.5 Flash in preview, its first hybrid reasoning model: reasoning can be turned on or off, and a thinking budget makes it possible to trade off quality, cost and latency.
Gemma 3 QAT (2025-04-18) — Google releases versions of Gemma 3 trained with quantization (QAT), which sharply reduce the video memory required: the 27B model goes from 54 GB to 14.1 GB in int4 and fits on a consumer graphics card.
gpt-image-1 (2025-04-23) — OpenAI makes gpt-image-1 available in the API, the native multimodal model that powers ChatGPT's image generation, to create and edit images with faithful text rendering and precise instructions.
New term: sycophancy (2025-04-29) — Sycophancy is a model's tendency to flatter users and endorse their ideas, even wrong or harmful ones, to please them. The rollback of a GPT-4o update in late April 2025 made the word mainstream.
Qwen3 (2025-04-29) — Alibaba Qwen releases Qwen3, eight open-weights models under Apache 2.0 (two MoE, 235B-A22B and 30B-A3B, and six dense from 0.6B to 32B) that combine a thinking mode and a direct mode in a single model.
MiMo-7B (2025-04-30) — Xiaomi releases MiMo-7B, its first family of open-weights language models (MIT): a 7-billion-parameter model trained for reasoning, at the level of o1-mini in mathematics and code.
Gemini 2.5 Pro Preview (I/O edition) (2025-05-06) — Google releases, ahead of Google I/O, an improved version of Gemini 2.5 Pro, better at code and at building interactive web applications, and first on the WebDev Arena leaderboard.
Mistral Medium 3 (2025-05-07) — Mistral AI launches Mistral Medium 3, a proprietary model aimed at professional uses (code, multimodal) at an announced cost 8 times lower, for 0.4 dollars input and 2 dollars output per million tokens.
codex-1 and codex-mini (Codex) (2025-05-16) — OpenAI launches Codex, a cloud software development agent powered by codex-1, a version of o3 optimized for software engineering, as well as codex-mini, a version of o4-mini for Codex CLI, available in the API.
Google I/O 2025 (2025-05-20) — Google holds its annual developer conference on May 20 and 21, 2025: new Gemini 2.5 models, video generation with Veo 3, AI Mode in Search and Android XR glasses.
Lyria 2 and Lyria RealTime (2025-05-20) — At Google I/O, Google DeepMind expands access to Lyria 2, its music composition model, and opens Lyria RealTime, a continuously generated music model, through an API and Google AI Studio.
MedGemma (2025-05-20) — Google releases MedGemma, a collection of open models derived from Gemma 3 for healthcare: 4B multimodal and 27B text at launch, complemented on July 9 by a 27B multimodal and the MedSigLIP encoder.
Gemini 2.5 Flash (May 20 update) (2025-05-20) — At Google I/O, Google releases an improved version of Gemini 2.5 Flash, with better performance on reasoning, multimodal, coding and long context, while using 20 to 30% fewer tokens.
Gemini 2.5 Flash and Pro TTS, Gemini 2.5 Flash Native Audio (2025-05-20) — At Google I/O, Google opens in preview the speech synthesis of Gemini 2.5 Flash and 2.5 Pro (two voices, more than 24 languages) and a native conversational voice for the Live API, improved in September and December.
Imagen 4 (2025-05-20) — Google DeepMind unveils Imagen 4, its next-generation image generation model: finer details, markedly improved text rendering and outputs up to 2K, offered in Standard, Ultra and Fast.
Gemini Diffusion (2025-05-20) — Google DeepMind unveils Gemini Diffusion, an experimental research model that generates text and code by diffusion, turning noise into text, much faster than its fastest model.
Gemma 3n (preview) (2025-05-20) — Google unveils Gemma 3n in preview, an open model designed to run on phones, tablets and laptops, which fits in 2 to 3 GB of memory despite 5 to 8 billion raw parameters.
Veo 3 (2025-05-20) — Google DeepMind unveils Veo 3, its video generation model that produces audio together with the image for the first time: ambient sounds, sound effects and synchronized dialogue, with better adherence to physics.
Devstral (2025-05-21) — Mistral AI and All Hands AI release Devstral, a 24 billion parameter agentic model for software engineering, under the Apache 2.0 license, which reaches 46.8% on SWE-Bench Verified.
Claude Sonnet 4 (2025-05-22) — Anthropic launches Claude Sonnet 4, a major upgrade of Sonnet 3.7 with better coding and instruction-following capabilities, at the same price, and available to free users.
Mistral OCR 2 (2025-05-22) — Mistral AI releases Mistral OCR 2 (mistral-ocr-2505) with document annotations: a new version of its OCR service, more reliable for extracting text and regions (bounding boxes) on varied documents.
Claude Opus 4 (2025-05-22) — Anthropic launches Claude Opus 4, presented as its best coding model, capable of long-running tasks spanning thousands of steps. It ships alongside Claude Sonnet 4 and the general availability of Claude Code.
DeepSeek R1-0528 (2025-05-28) — DeepSeek releases R1-0528, an upgrade to R1 that doubles the reasoning depth and raises AIME 2025 from 70.0 to 87.5, with a version distilled into Qwen3 8B.
New term: GEO (2025-05-28) — GEO (generative engine optimization) aims to get a brand or content cited in AI assistants' answers, the way SEO aims for ranking in search engines.
Codestral Embed (2025-05-28) — Mistral AI launches Codestral Embed, its first embedding model specialized in code, announced as outperforming Voyage Code 3, Cohere Embed v4.0 and OpenAI's large embedding model for code retrieval.
Gemini 2.5 Pro Preview (06-05) (2025-06-05) — Google releases an improved version of Gemini 2.5 Pro in preview, ahead of general availability: a 24-point Elo gain on LMArena, better response style and thinking budgets.
o3-pro (2025-06-10) — OpenAI launches o3-pro, a version of o3 that thinks longer for more reliable answers, available to ChatGPT Pro and Team subscribers and in the API, alongside an o3 price cut.
Magistral (2025-06-10) — Mistral AI introduces Magistral, its first family of reasoning models: Magistral Small (24B, Apache 2.0) and Magistral Medium (enterprise version), with traceable, multilingual reasoning.
Meta invests $14.3B in Scale AI (2025-06-12) — Meta takes a 49% stake in Scale AI, the data-labeling specialist, for about $14.3 billion and hires its founder Alexandr Wang: the kickoff of an unprecedented talent war in AI.
MiniMax-M1 (2025-06-16) — MiniMax releases MiniMax-M1, presented as the first large-scale open reasoning model with hybrid attention: 456 billion parameters, 1 million tokens of context, Apache 2.0 license.
Gemini 2.5 Flash (general availability) (2025-06-17) — Google makes Gemini 2.5 Flash stable and available in production, removing the price difference between reasoning on and off and keeping a single pricing tier regardless of input size.
Gemini 2.5 Flash-Lite (2025-06-17) — Google launches Gemini 2.5 Flash-Lite, the cheapest and fastest model in the 2.5 family: preview on June 17, 2025, stable version on July 22 at 0.10 USD for input and 0.40 USD for output.
Gemini 2.5 Pro (general availability) (2025-06-17) — Google declares Gemini 2.5 Pro stable and generally available, with no change from the June 5 preview, with adaptive reasoning and thinking budgets for production use.
New term: context engineering (2025-06-19) — Context engineering means giving a model everything it needs to succeed at a task: instructions, documents, history, tools, memory. The term extends and partly replaces prompt engineering.
Mistral Small 3.2 (2025-06-20) — Mistral AI releases Mistral Small 3.2 (24B), a minor update to Small 3.1 under the Apache 2.0 license: better instruction following, fewer repetitions and more robust function calling.
Gemini Robotics On-Device (2025-06-24) — Google DeepMind introduces Gemini Robotics On-Device, its vision-language-action model optimized to run on the robot itself, adaptable to new tasks with 50 to 100 demonstrations.
Gemma 3n (2025-06-26) — Google releases the full version of Gemma 3n, a multimodal open model for mobile devices in effective sizes E2B and E4B, which accepts text, image, audio and video as input and exceeds 1,300 Elo on LMArena.
o3-deep-research and o4-mini-deep-research (2025-06-26) — OpenAI opens two deep research models in the API, o3-deep-research and o4-mini-deep-research, derived from the o3 and o4-mini reasoning models to analyze and synthesize web and private sources.
AI Agents Hackathon (2025-06-27) — Second edition of the internal hackathon, this time focused on AI agents. 48 hours to create innovative automation solutions.
Hunyuan-A13B (2025-06-27) — Tencent opens Hunyuan-A13B, a mixture of experts (MoE) model with 80 billion parameters of which 13 billion are active, with fast or slow hybrid reasoning and a 256,000-token context, under the Tencent community license.
ERNIE 4.5 (2025-06-30) — Baidu opens the entire ERNIE 4.5 family under the Apache 2.0 license: ten models, from a 0.3-billion dense model to multimodal MoEs with 47 billion active parameters, the largest totaling 424 billion.
Cloudflare blocks AI crawlers by default (2025-07-01) — Cloudflare, which handles about 20% of web traffic, reverses the rule: AI crawling now requires the site's permission, and a pay-per-crawl model is tested in private beta.
Grok 4 Heavy (2025-07-09) — xAI introduces Grok 4 Heavy, the most powerful version of Grok 4 based on parallel compute at inference time, available through the new SuperGrok Heavy subscription.
T5Gemma (2025-07-09) — Google releases T5Gemma, a collection of encoder-decoder models obtained by converting decoder-only Gemma 2 models, with unbalanced size combinations such as a 9B encoder and a 2B decoder.
Grok 4 (2025-07-09) — xAI launches Grok 4, presented as its most intelligent model, with native tool use, real-time search and a 256,000-token context window in the API.
EU General-Purpose AI Code of Practice (2025-07-10) — The Commission publishes the Code of Practice for general-purpose AI models: a voluntary three-chapter tool that lets providers demonstrate compliance with the AI Act ahead of the August 2, 2025 deadline.
Devstral Medium (2025-07-10) — Mistral AI launches Devstral Medium, an agentic coding model available through the API and in private deployment, at 61.6% on SWE-Bench Verified, presented as superior to Gemini 2.5 Pro and GPT-4.1 for a quarter of the price.
Devstral Small 1.1 (2025-07-10) — Mistral AI releases Devstral Small 1.1, an update to its open agentic coding model (24B, Apache 2.0) that goes from 46.8% to 53.6% on SWE-Bench Verified and handles varied scaffolds better.
Kimi K2 (2025-07-11) — Moonshot AI releases Kimi K2, a MoE model with 1 trillion parameters (32 billion active) with open weights under a modified MIT license, trained with the MuonClip optimizer and designed for agentic uses.
New term: Ralph Wiggum loop (2025-07-14) — The Ralph Wiggum loop re-runs a coding agent with the same prompt over and over until the task is done and verified. Devised by Geoffrey Huntley in July 2025, the technique went viral in late 2025 and early 2026.
New term: context rot (2025-07-14) — Context rot is the drop in a model's reliability as its context grows longer, even on simple tasks and well below its maximum window size.
New term: spec-driven development (2025-07-14) — Spec-driven development has requirements, design and a task plan written before any code: the specification becomes the source of truth the agent executes. AWS's Kiro launched the approach in July 2025.
Voxtral (2025-07-15) — Mistral AI launches Voxtral, its first speech understanding models: Small (24B) and Mini (3B), under Apache 2.0, to transcribe and query audio, starting at $0.001 per minute.
IMO gold medal for OpenAI and Google DeepMind (2025-07-19) — Two general-purpose language models, one from OpenAI and one from Google DeepMind, score 35 out of 42 at the 2025 International Mathematical Olympiad, gold-medal level, writing their proofs in natural language.
Qwen3-2507 (Qwen3-235B-A22B Instruct and Thinking) (2025-07-21) — Alibaba Qwen releases Qwen3-235B-A22B-Instruct-2507 (July 21) then Qwen3-235B-A22B-Thinking-2507 (July 25): two updates of the 235B MoE, with a native context of 262,144 tokens, under Apache 2.0.
Qwen3-Coder (2025-07-22) — Alibaba Qwen releases Qwen3-Coder-480B-A35B-Instruct, a MoE with 480 billion parameters (35 billion active) under Apache 2.0, dedicated to agentic coding, with the Qwen Code command-line tool.
America's AI Action Plan (2025-07-23) — The White House publishes "Winning the AI Race: America's AI Action Plan", more than 90 federal measures to accelerate innovation, build infrastructure and export American technology.
GLM-4.5 and GLM-4.5-Air (2025-07-28) — Z.ai (formerly Zhipu AI) has released GLM-4.5 (355B parameters, 32B active) and GLM-4.5-Air (106B, 12B active), two MoE models with hybrid reasoning for agents and code, under the MIT license.
Codestral 25.08 (2025-07-30) — Mistral AI announces Codestral 25.08, an update to its code completion model (+30% accepted completions, 50% fewer runaway generations), as part of an enterprise coding stack.
Step3 (2025-07-31) — StepFun opens Step3, a multimodal mixture-of-experts model with 321 billion parameters, 38 of them active, under Apache 2.0, designed with its inference system to reduce decoding cost on a variety of GPUs.
Gemini 2.5 Deep Think (2025-08-01) — Google rolls out Gemini 2.5 Deep Think to Google AI Ultra subscribers: a parallel-thinking reasoning mode, a faster variant of the model that won a gold medal at the International Mathematical Olympiad.
AI Act: general-purpose AI model obligations apply (2025-08-02) — Second stage of the AI Act: since August 2, 2025, providers of general-purpose AI models placed on the EU market must document their models, respect copyright and publish a summary of their training data.
gpt-oss-120b and gpt-oss-20b (2025-08-05) — OpenAI releases gpt-oss-120b and gpt-oss-20b, its first open-weights language models since GPT-2, under the Apache 2.0 license, designed for reasoning, tool use and deployment on commodity hardware.
Genie 3 (2025-08-05) — Google DeepMind announces Genie 3, a general-purpose world model that creates interactive environments navigable in real time at 24 frames per second, in 720p, from a simple text prompt.
Claude Opus 4.1 (2025-08-05) — Anthropic releases Claude Opus 4.1, an update to Opus 4 with better performance in coding, agents and reasoning, at the same price. The vendor recommends migrating from Opus 4 for all uses.
GPT-5 (2025-08-07) — OpenAI launches GPT-5, a unified system that combines a fast model, a deep reasoning model and a router, with mini, nano and pro versions for the API, and it becomes the default model in ChatGPT.
Mistral Medium 3.1 (2025-08-12) — Mistral AI releases Mistral Medium 3.1 (mistral-medium-2508), an update to Medium 3 announced with an overall performance gain, a better tone and more relevant web searches, via API and in Le Chat.
Gemma 3 270M (2025-08-14) — Google releases Gemma 3 270M, an open model with 270 million parameters designed for fine-tuning on specific tasks, very energy-efficient and able to follow instructions out of the box.
NVIDIA Nemotron Nano 2 and Nemotron-Pretraining v1 (2025-08-18) — NVIDIA releases Nemotron Nano 2 (9B and 12B), a hybrid Mamba-Transformer reasoning model, along with most of its pretraining corpus (6.6 trillion tokens).
Seed-OSS-36B (2025-08-20) — ByteDance releases Seed-OSS-36B, its first large general-purpose open-weights model: a dense model with 36 billion parameters, a native 512,000-token context and an adjustable thinking budget, under Apache 2.0.
AI Platform Launch (2025-08-20) — Official launch of the internal ASTEK AI platform. Centralization of tools, agents, and models for all employees.
DeepSeek V3.1 (2025-08-21) — DeepSeek releases V3.1, a 671-billion-parameter model that combines a thinking mode and a non-thinking mode in a single set of weights, with clear advances for coding and search agents.
Gemini 2.5 Flash Image (Nano Banana) (2025-08-26) — Google launches Gemini 2.5 Flash Image, nicknamed Nano Banana, its image editing and generation model: image fusion, character consistency, targeted edits in natural language and world knowledge.
New term: vibe hacking (2025-08-27) — Vibe hacking is attackers using coding agents: they set a broad goal and let the AI itself carry out reconnaissance, intrusion and extortion. Anthropic popularized the term in August 2025.
gpt-realtime (2025-08-28) — OpenAI makes the Realtime API generally available and launches gpt-realtime, its most advanced speech-to-speech model, with better instruction following, more accurate tool calls and more natural voices.
Grok Code Fast 1 (2025-08-28) — xAI launches grok-code-fast-1, a fast and economical reasoning model designed for agentic coding, available in the API and free for a limited time from several partners.
LongCat-Flash-Chat (2025-09-01) — Meituan opens LongCat-Flash-Chat, a mixture-of-experts model with 560 billion parameters, of which 18.6 to 31.3 billion are active depending on the context, under the MIT license, very fast at inference and strong on agent tasks.
EmbeddingGemma (2025-09-04) — Google releases EmbeddingGemma, an open embedding model with 308 million parameters for on-device use: first on the MTEB ranking among multilingual models under 500M.
Anthropic settles with authors for $1.5B (Bartz case) (2025-09-05) — Sued for downloading pirated books, Anthropic agrees to pay at least $1.5 billion to authors and publishers, about $3,000 per work: the largest publicly reported copyright settlement in the United States.
Mistral AI raises €1.7B, led by ASML (2025-09-09) — Mistral AI closes a €1.7 billion Series C led by Dutch company ASML, which contributes €1.3 billion. The French start-up is valued at €11.7 billion, a European tie-up between AI and semiconductors.
Qwen3-Next (2025-09-11) — Alibaba Qwen introduces Qwen3-Next-80B-A3B, an ultra-sparse MoE with 80 billion parameters (3 billion active) and hybrid attention, released in Instruct and Thinking versions under Apache 2.0.
VaultGemma (2025-09-12) — Google Research and DeepMind release VaultGemma, an open model with one billion parameters, the largest pre-trained from scratch with differential privacy, along with associated scaling laws.
GPT-5-Codex (2025-09-15) — OpenAI releases GPT-5-Codex, a version of GPT-5 optimized for agentic development in Codex, capable of fast interactive sessions as well as long autonomous tasks, with a code review capability.
Magistral 1.2 (2025-09-17) — Mistral AI releases Magistral Medium 1.2 and Magistral Small 1.2, which add a vision encoder (multimodal reasoning) and significantly improve reasoning scores, with Small remaining under the Apache 2.0 license.
Grok 4 Fast (2025-09-19) — xAI launches Grok 4 Fast, a cost-efficient reasoning model with a 2 million token context window and a unified reasoning and non-reasoning architecture.
Qwen3-Omni (2025-09-22) — Alibaba Qwen releases Qwen3-Omni-30B-A3B, an open-weights omni-modal MoE model (Instruct, Thinking and Captioner) that processes text, image, audio and video and responds in text or streaming speech.
Nvidia announces up to $100B investment in OpenAI (2025-09-22) — Nvidia and OpenAI sign a letter of intent: at least 10 GW of Nvidia systems for OpenAI and up to $100 billion invested by Nvidia as deployments proceed. The non-binding deal is replaced in early 2026 by a smaller equity investment.
New term: workslop (2025-09-22) — Workslop is AI-generated work content that looks like good work but lacks substance, leaving the recipient to fix or redo it. The term comes from a Harvard Business Review article of September 2025.
Qwen3-VL (2025-09-23) — Alibaba Qwen releases Qwen3-VL-235B-A22B in Instruct and Thinking versions under Apache 2.0, followed by a family of dense and MoE sizes (2B to 32B, 30B-A3B); native 256,000-token context and a visual agent.
Code World Model (CWM) 32B (2025-09-24) — Meta FAIR releases Code World Model (CWM), a dense 32-billion-parameter open-weights model for research, trained to simulate code execution and to solve software engineering tasks.
Gemini Robotics 1.5 and Gemini Robotics-ER 1.5 (2025-09-25) — Google DeepMind launches Gemini Robotics 1.5, a vision-language-action model that thinks before acting, and Robotics-ER 1.5, a spatial reasoning model open to developers in preview through the Gemini API.
Gemini 2.5 Flash and Flash-Lite (09-2025 previews) (2025-09-25) — Google releases improved preview versions of Gemini 2.5 Flash and Flash-Lite, with output token consumption reduced by 24% and 50%, and introduces the "-latest" aliases.
Claude Sonnet 4.5 (2025-09-29) — Anthropic launches Claude Sonnet 4.5, presented as its best coding and agent model, able to stay focused for more than 30 hours. The Claude Agent SDK is released the same day.
New term: agent harness (2025-09-29) — An agent harness is everything around a model that turns it into an agent: execution loop, tools, context management, memory, permissions and checks. Anthropic popularized the term with the Claude Agent SDK in late September 2025.
California: SB 53 frontier AI law (2025-09-29) — Gavin Newsom signs SB 53, the Transparency in Frontier Artificial Intelligence Act: large frontier model developers must publish their risk management framework and report critical safety incidents.
New term: agentic commerce (2025-09-29) — Agentic commerce is purchasing carried out by an AI agent on the user's behalf, from product search to payment. OpenAI and Stripe popularized it on September 29, 2025 with Instant Checkout and the Agentic Commerce Protocol.
DeepSeek V3.2-Exp (2025-09-29) — DeepSeek releases V3.2-Exp, an experimental version that introduces DSA sparse attention to make long context less costly, at quality equivalent to V3.1-Terminus, with API prices cut by more than half.
Sora 2 (2025-09-30) — OpenAI launches Sora 2, its flagship model for video and synchronized audio generation, with more credible physics and better control, along with an invitation-only social Sora iOS app.
GLM-4.6 (2025-09-30) — Z.ai has released GLM-4.6, an update to GLM-4.5: context raised from 128K to 200K tokens, better agentic code, reasoning with tools, open weights under the MIT license.
gpt-realtime-mini, gpt-audio-mini and gpt-image-1-mini (2025-10-06) — At DevDay, OpenAI opens three more economical models in the API: gpt-realtime-mini and gpt-audio-mini for speech-to-speech, and gpt-image-1-mini for image generation.
Gemini 2.5 Computer Use (2025-10-07) — Google DeepMind releases Gemini 2.5 Computer Use in preview, a specialized model built on Gemini 2.5 Pro that drives graphical interfaces (clicks, typing, scrolling), mainly browsers.
Ling-1T (2025-10-09) — Ant Group, through its inclusionAI team, opens Ling-1T, a language model with 1 trillion parameters (about 50 billion active), non-reasoning, under the MIT license, trained in FP8 precision.
Ring-1T (2025-10-14) — Ant Group opens Ring-1T, a reasoning model with 1 trillion parameters (50 billion active) under the MIT license, which reaches silver-medal level at the 2025 International Mathematical Olympiad in natural language.
Veo 3.1 (2025-10-15) — Google launches Veo 3.1 and Veo 3.1 Fast in paid preview in the Gemini API: richer native audio, better narrative control, image references, scene extension and transitions between first and last frame.
Claude Haiku 4.5 (2025-10-15) — Anthropic launches Claude Haiku 4.5, a fast model that delivers coding performance close to Sonnet 4 at one third of the cost and more than twice the speed, under ASL-2.
C2S-Scale 27B (Cell2Sentence-Scale) (2025-10-15) — Google DeepMind and Yale release C2S-Scale 27B, a 27-billion-parameter foundation model for single-cell analysis, built on the open Gemma family, which proposed a therapeutic lead validated in the laboratory.
Agent Skills: Anthropic launches skills for agents (2025-10-16) — Anthropic introduces Agent Skills: folders of instructions, scripts and resources, described by a SKILL.md file, that Claude loads only when a task needs them. The format becomes an open standard on December 18, 2025.
MiniMax-M2 (2025-10-27) — MiniMax releases MiniMax-M2, a compact 230-billion-parameter MoE with 10 billion active, designed for code and agents, under a modified MIT license, at $0.30 / $1.20 per million tokens.
OpenAI becomes a public benefit corporation; Microsoft holds about 27% (2025-10-28) — OpenAI completes its recapitalization: its business becomes OpenAI Group PBC, controlled by the OpenAI Foundation, which holds 26%. Microsoft keeps about 27%, worth nearly $135 billion, under a new partnership agreement.
Nvidia becomes the first $5 trillion company (2025-10-29) — On October 29, 2025, Nvidia becomes the first listed company to top $5 trillion in market value and close above that mark, only three months after crossing $4 trillion.
gpt-oss-safeguard (2025-10-29) — OpenAI releases gpt-oss-safeguard-120b and 20b in preview, open-weights reasoning models that classify content according to a safety policy provided by the developer at inference time.
Kimi Linear (48B-A3B) (2025-10-30) — Moonshot AI releases Kimi Linear, a 48-billion-parameter MoE (3 billion active) with hybrid linear attention (Kimi Delta Attention), which reduces the KV cache by up to 75% and speeds up decoding by up to 6 times.
Kimi K2 Thinking (2025-11-06) — Moonshot AI releases Kimi K2 Thinking, an open agentic reasoning model with 1 trillion parameters, capable of 200 to 300 chained tool calls and natively quantized in INT4.
ERNIE-4.5-VL-28B-A3B-Thinking (2025-11-11) — Baidu releases ERNIE-4.5-VL-28B-A3B-Thinking, a vision-language model with 28 billion parameters, 3 billion of them active, under the Apache 2.0 license, which reasons over images and fits on a single 80 GB GPU.
GPT-5.1 (2025-11-12) — OpenAI launches GPT-5.1 with two models in ChatGPT, GPT-5.1 Instant and GPT-5.1 Thinking, warmer and more reliable, then offers them in the API the next day with adaptive reasoning.
Grok 4.1 (2025-11-17) — xAI rolls out Grok 4.1 to all users of grok.com, X and the mobile apps, with a more consistent personality, more emotional intelligence and fewer hallucinations.
Gemini 3 Pro (2025-11-18) — Google launches Gemini 3 Pro, its most intelligent model: first on LMArena (1,501 Elo), strong in reasoning, multimodality and agentic coding, in preview in the Gemini API, Vertex AI and the app.
Digital Omnibus: the Commission proposes to delay parts of the AI Act (2025-11-19) — In its "Digital Omnibus", the Commission proposes to delay the AI Act's high-risk rules, as the standards are not ready. The text, adopted in 2026, sets the deadlines at December 2027 and August 2028.
GPT-5.1-Codex-Max (2025-11-19) — OpenAI launches GPT-5.1-Codex-Max, its new agentic development model, trained to work across multiple context windows through compaction, for tasks lasting several hours.
Grok 4.1 Fast and Agent Tools API (2025-11-19) — xAI launches Grok 4.1 Fast, its best tool-calling model with 2 million tokens of context, and the Agent Tools API for building autonomous agents with server-side tools.
Nano Banana Pro (Gemini 3 Pro Image) (2025-11-20) — Google launches Nano Banana Pro (Gemini 3 Pro Image), an image model built on Gemini 3 Pro: legible text, Google Search grounding, 2K and 4K outputs, consistency for up to 5 people.
Olmo 3 (7B and 32B) and Olmo 3.1 (2025-11-20) — Ai2 releases Olmo 3 (7B and 32B), fully open: data, code, weights and checkpoints. Olmo 3-Think 32B is presented as the best fully open 32B reasoning model.
Claude Opus 4.5 (2025-11-24) — Anthropic launches Claude Opus 4.5, its most capable model in coding, agents and computer use, with a price cut by two thirds (5 / 25 USD) and a new effort parameter.
DeepSeek V3.2 (2025-12-01) — DeepSeek releases V3.2 and V3.2-Speciale: the former reaches GPT-5 level according to the vendor, the latter rivals Gemini 3 Pro and wins gold at the IMO and IOI 2025.
Ministral 3 (2025-12-02) — Mistral AI releases Ministral 3, a range of dense 3B, 8B and 14B parameter models, multimodal, multilingual and under the Apache 2.0 license, in base, instruct and reasoning variants, designed for edge and local use.
Mistral Large 3 (2025-12-02) — Mistral AI launches Mistral Large 3, its most capable model: a multimodal mixture-of-experts with 675 billion parameters (41 billion active), released under the Apache 2.0 license, with a 256k-token context.
Gemini 3 Deep Think (2025-12-04) — Google rolls out Gemini 3 Deep Think mode in the Gemini app for Google AI Ultra subscribers: parallel reasoning on the hardest math, science and logic problems.
Devstral 2 (2025-12-09) — Mistral AI launches Devstral 2, a dense 123-billion-parameter model for coding agents, at 72.2% on SWE-bench Verified, under a modified MIT license, with the Mistral Vibe command-line interface.
Agentic AI Foundation: MCP, AGENTS.md and goose move to the Linux Foundation (2025-12-09) — The Linux Foundation creates the Agentic AI Foundation (AAIF), home to Anthropic's Model Context Protocol, OpenAI's AGENTS.md and Block's goose: the open building blocks of agents move under neutral governance.
Devstral Small 2 (2025-12-09) — Mistral AI releases Devstral Small 2, a 24-billion-parameter agentic coding model under the Apache 2.0 license, at 68.0% on SWE-bench Verified, able to run locally on consumer hardware.
GPT-5.2 (2025-12-11) — OpenAI launches GPT-5.2, its most capable series of models for professional knowledge work, with Instant, Thinking and Pro versions, better at spreadsheets, presentations, code, vision and long contexts.
AI dominates the 2025 words of the year: "slop" and "vibe coding" (2025-12-15) — On December 15, 2025, Merriam-Webster named "slop", low-quality content mass-produced by AI, its word of the year. Collins had crowned "vibe coding" and Macquarie "AI slop": generative AI has entered everyday language.
NVIDIA Nemotron 3 (Nano 30B-A3B) (2025-12-15) — NVIDIA introduces the Nemotron 3 family (Nano, Super, Ultra) and releases Nano, a hybrid MoE with 30 billion parameters, 3 billion of them active, together with weights, datasets and training recipes.
MiMo-V2-Flash (2025-12-16) — Xiaomi releases MiMo-V2-Flash, an open-weights model (MIT) with 309 billion parameters, 15 of them active, designed for agents and code, with a hybrid attention that reduces the cache by about six times.
Molmo 2 (4B, 8B and Molmo2-O 7B) (2025-12-16) — Ai2 releases Molmo 2, a family of open multimodal models for video, multiple images and pointing, in 4B and 8B versions (Qwen 3 base) and Molmo2-O 7B (Olmo base).
GPT Image 1.5 (2025-12-16) — OpenAI launches a new ChatGPT Images powered by GPT Image 1.5, its flagship image model, more precise in edits, more faithful to instructions and up to four times faster, also available in the API.
Mistral OCR 3 (2025-12-17) — Mistral AI launches Mistral OCR 3, a document extraction model that wins 74% of comparisons against OCR 2 on forms, scans, complex tables and handwriting, at 2 dollars per 1,000 pages.
Grok Voice Agent API (2025-12-17) — xAI launches the Grok Voice Agent API to build multilingual voice agents with tool calls and real-time data, at a flat rate of $0.05 per minute.
Gemini 3 Flash (2025-12-17) — Google launches Gemini 3 Flash, a model that combines the Pro-level reasoning of Gemini 3 with the speed and cost of a Flash: USD 0.50 input and USD 3 output per million tokens.
T5Gemma 2 (2025-12-18) — Google releases T5Gemma 2, the new generation of its encoder-decoder models based on Gemma 3: the first multimodal and long-context encoder-decoder models, in three compact sizes.
AI 2025 Review Meetup (2025-12-18) — Year-end meetup to review the AI year 2025. Retrospective, demos, and outlook for 2026.
GPT-5.2-Codex (2025-12-18) — OpenAI releases GPT-5.2-Codex, its most advanced version of GPT-5.2 for agentic development, better at large refactors and migrations, on Windows and in cybersecurity.
FunctionGemma (2025-12-18) — Google releases FunctionGemma, a version of Gemma 3 270M specialized in function calling, with its training recipe, to build local and private agents that translate natural language into API calls.
GLM-4.7 (and GLM-4.7-Flash) (2025-12-22) — Z.ai has released GLM-4.7, an open coding-oriented model (73.8% on SWE-bench Verified), with thinking between actions, thinking preserved across turns and per-turn control of reasoning, under the MIT license.
MiniMax-M2.1 (2025-12-23) — MiniMax has released M2.1, an update to M2 focused on multilingual programming (Rust, Java, Go, C++, Kotlin), mobile development and interleaved thinking, with 74.0 on SWE-bench Verified.
Apple picks Gemini for Siri and Apple Intelligence (2026-01-12) — Apple and Google announce a multi-year partnership: the next generation of Apple Foundation Models will be based on Gemini models and Google cloud technology, including for the more personalized Siri expected in 2026.
MedGemma 1.5 4B (2026-01-13) — Google Research has released MedGemma 1.5 4B, an open healthcare model that now handles 3D imaging (CT, MRI), histopathology and chest X-ray follow-up, together with the MedASR dictation model.
TranslateGemma (2026-01-15) — Google has released TranslateGemma, a family of open translation models in 4B, 12B and 27B parameters built on Gemma 3, evaluated on 55 languages and designed to run on a laptop or mobile device.
SERA (open coding agents 8B, 14B, 32B) (2026-01-27) — Ai2 has released SERA, a family of open coding agents (8B, 14B, 32B, Qwen 3 base) with weights, code, synthetic data and a CLI, trained at low cost thanks to soft-verified data generation.
Kimi K2.5 (2026-01-27) — Moonshot AI has released Kimi K2.5, a natively multimodal open model of 1 trillion parameters, extending K2 on 15 T image and text tokens, with an agent swarm of up to 100 sub-agents.
Step 3.5 Flash (2026-01-29) — StepFun has opened Step 3.5 Flash, a mixture-of-experts model of 196 billion parameters with 11 billion active, under Apache 2.0, with 256,000 tokens of context and generation of 100 to 300 tokens per second.
Grok Imagine 1.0 (2026-02-01) — xAI launches Grok Imagine 1.0, its most polished video generation release: 10-second clips in 720p with markedly improved sound, presented by the company as its biggest leap.
SpaceX acquires xAI (2026-02-02) — On February 2, 2026, SpaceX announces the acquisition of xAI (Grok, X) through a share exchange, for a combined valuation of about 1.25 trillion dollars, the largest known merger.
Qwen3-Coder-Next (2026-02-03) — Alibaba (Qwen team) has released Qwen3-Coder-Next, an open weights coding model of 80 billion parameters with 3 billion active, designed for coding agents, under Apache 2.0.
New term: agentic engineering (2026-02-04) — Agentic engineering is professional software development in which you orchestrate and supervise agents that write the code, with no compromise on quality. Andrej Karpathy adopted it in February 2026 as the successor to vibe coding.
Voxtral Transcribe 2 (2026-02-04) — Mistral AI launches Voxtral Transcribe 2, two speech transcription models: Voxtral Mini Transcribe V2 in batch mode and Voxtral Realtime in real time, the latter under the Apache 2.0 license.
GPT-5.3-Codex (2026-02-05) — OpenAI launches GPT-5.3-Codex, its most capable agentic coding model, which combines the performance of GPT-5.2-Codex with the reasoning of GPT-5.2 and runs 25% faster.
New term: harness engineering (2026-02-05) — Harness engineering means designing an agent's environment (instructions, tools, tests, checks) so that a fixed mistake never happens again. Mitchell Hashimoto named the practice in February 2026; OpenAI popularized it six days later.
Claude Opus 4.6 (2026-02-05) — Anthropic launches Claude Opus 4.6, its most capable model to date, better at coding, long agentic tasks and reliability on large codebases, with a 1M-token window in beta.
Ring-2.5-1T (2026-02-10) — Ant Group open-sources Ring-2.5-1T, a 1-trillion-parameter reasoning model with hybrid linear attention, under the MIT license, which reaches gold-medal level at IMO 2025 and CMO 2025 according to the vendor.
GPT-5.3-Codex-Spark (2026-02-12) — OpenAI introduces GPT-5.3-Codex-Spark, a smaller version of GPT-5.3-Codex designed for real-time coding, served on Cerebras chips at more than 1,000 tokens per second.
Gemini 3 Deep Think (major upgrade) (2026-02-12) — Google updates Gemini 3 Deep Think, its specialized reasoning mode for science, research and engineering, with new records on ARC-AGI-2 and Humanity's Last Exam.
GLM-5 (2026-02-12) — Z.ai has released GLM-5, a 744B-parameter MoE model (40B active) trained on 28.5T tokens, with DeepSeek Sparse Attention and asynchronous RL, under the MIT license, aimed at systems engineering and long-running agents.
MiniMax-M2.5 (2026-02-12) — MiniMax releases M2.5, trained by reinforcement learning across hundreds of thousands of environments, with 80.2 on SWE-bench Verified and a speed of 100 tokens per second in the Lightning version.
Anthropic raises 30 billion USD at a 380 billion USD valuation (2026-02-12) — Anthropic closes a 30 billion USD Series G led by GIC and Coatue, valuing it at 380 billion USD post-money, more than double its Series F. Its run-rate revenue reaches 14 billion USD.
Qwen3.5 (2026-02-16) — Alibaba launches Qwen3.5, a natively multimodal generation whose open-weights flagship Qwen3.5-397B-A17B (17 billion active parameters) is released under Apache 2.0, followed by a range from 0.8B to 122B.
Grok 4.20 (2026-02-17) — xAI launches Grok 4.20 in beta, a flagship model focused on speed and tool calling, with a multi-agent architecture; API access and the final version follow on March 10, 2026.
Claude Sonnet 4.6 (2026-02-17) — Anthropic launches Claude Sonnet 4.6, its most capable Sonnet, improved in coding, computer use, long-context reasoning and agent planning, at the same price as Sonnet 4.5.
Lyria 3 (2026-02-18) — Google DeepMind launches Lyria 3, its music generation model, in beta in the Gemini app: 30-second tracks with generated lyrics, from a text prompt, a photo or a video.
Gemini 3.1 Pro (2026-02-19) — Google launches Gemini 3.1 Pro in preview, a reasoning model that reaches 77.1% on ARC-AGI-2, more than double Gemini 3 Pro, available via the API, Vertex AI and the Gemini app.
AI Impact Summit in New Delhi: the New Delhi Declaration (2026-02-19) — The AI Impact Summit, the first global AI summit hosted in the Global South, concludes with the New Delhi Declaration: a non-binding text endorsed by 88 countries and organisations, including the United States and China, then by 92.
New term: claws (2026-02-20) — "Claws" are personal, autonomous AI assistants modeled on OpenClaw: they run on the user's own machine, are driven through messaging apps and schedule tasks. Andrej Karpathy popularized the term in February 2026.
GPT-Realtime-1.5 (2026-02-23) — OpenAI releases GPT-Realtime-1.5 in the Realtime API, its flagship audio model for voice agents and customer support, with gpt-audio-1.5 for the Chat Completions API.
Nano Banana 2 (Gemini 3.1 Flash Image) (2026-02-26) — Google DeepMind launches Nano Banana 2 (Gemini 3.1 Flash Image), an image model that brings the capabilities of Nano Banana Pro to the speed of Gemini Flash, from 512 px to 4K.
GPT-5.3 Instant (2026-03-03) — OpenAI updates ChatGPT's most widely used model with GPT-5.3 Instant: fewer unnecessary refusals, better-synthesized web answers and fewer hallucinations.
Gemini 3.1 Flash-Lite (2026-03-03) — Google launches Gemini 3.1 Flash-Lite, its fastest and cheapest generation 3 model, at USD 0.25 per million input tokens, for translation, classification and real-time use.
The Pentagon designates Anthropic a "supply chain risk" (2026-03-05) — After Anthropic refused to let Claude be used for fully autonomous weapons and mass surveillance of Americans, the Pentagon officially declares it a "supply chain risk", a first for an American company.
Olmo Hybrid (7B) (2026-03-05) — Ai2 releases Olmo Hybrid, a fully open 7-billion-parameter model that alternates attention and Gated DeltaNet layers, and matches Olmo 3's accuracy on MMLU with 49% fewer tokens.
GPT-5.4 (2026-03-05) — OpenAI launches GPT-5.4, a general-purpose reasoning model that integrates the coding capabilities of GPT-5.3-Codex and adds native computer use, with 1 million tokens of context.
Gemini Embedding 2 (2026-03-10) — Google DeepMind launches Gemini Embedding 2, the first natively multimodal embedding model, which places text, images, video, audio and documents in a single vector space, in public preview.
NVIDIA Nemotron 3 Super (120B-A12B) (2026-03-11) — NVIDIA releases Nemotron 3 Super, an open model with 120 billion parameters, 12 billion of them active, a hybrid Mamba-Transformer LatentMoE designed for agentic reasoning with a one-million-token context.
GTC 2026: Nvidia launches the Vera Rubin platform (2026-03-16) — At GTC in San Jose, Jensen Huang presents the Vera Rubin platform: seven chips in production, including the first LPU from Groq, designed for agentic AI. He cites 1 trillion USD in Blackwell and Vera Rubin orders through 2027.
Mistral Small 4 (2026-03-16) — Mistral AI launches Mistral Small 4, an open source Apache 2.0 model with 119 billion parameters (6 billion active) that combines instruction following, reasoning, vision and coding agents in a single model.
Leanstral (2026-03-16) — Mistral AI releases Leanstral, an open source Apache 2.0 agent of about 120 billion parameters (6 billion active) for formal proof engineering in Lean 4 and verifiable coding.
GPT-5.4 mini and GPT-5.4 nano (2026-03-17) — OpenAI launches GPT-5.4 mini and GPT-5.4 nano, its most capable small models: mini targets fast coding and subagents, nano targets classification and extraction at very low cost.
MiniMax-M2.7 (2026-03-18) — MiniMax introduces M2.7, described as its first model to take part in its own evolution, with 56.22% on SWE-Pro, 57.0% on Terminal Bench 2 and weights published under a non-commercial license.
Voxtral TTS (2026-03-23) — Mistral AI launches Voxtral TTS, its first text-to-speech model: 4 billion parameters, 9 languages, voice adaptation in a few seconds and open weights under the CC BY-NC 4.0 license.
Lyria 3 Pro (2026-03-25) — Google launches Lyria 3 Pro, its most advanced music generation version: tracks up to 3 minutes with control of the structure (intro, verse, chorus, bridge), in preview on Vertex AI and the Gemini API.
Gemini 3.1 Flash Live (2026-03-26) — Google launches Gemini 3.1 Flash Live, its most accomplished audio and voice model, for real-time voice agents, through the Live API, Search Live and Gemini Live, in more than 200 countries and territories.
Veo 3.1 Lite (2026-03-31) — Google launches Veo 3.1 Lite, its cheapest video generation model, at less than half the price of Veo 3.1 Fast for the same speed, pay-as-you-go on the Gemini API.
OpenAI raises 122 billion USD at an 852 billion USD valuation (2026-03-31) — OpenAI closes the largest private funding round in history: 122 billion USD in commitments at an 852 billion USD valuation, anchored by Amazon, Nvidia and SoftBank, with 3 billion USD from individual investors for the first time.
Gemma 4 (2026-04-02) — Google DeepMind releases Gemma 4, its most capable family of open models, under the Apache 2.0 license, in four sizes (E2B, E4B, 26B MoE, 31B dense) designed for reasoning and agents.
Project Glasswing: securing critical software with Claude Mythos Preview (2026-04-07) — Anthropic launches Project Glasswing, a cyber defense program giving vendors and maintainers of critical software restricted access to Claude Mythos Preview. Within seven weeks, more than 10,000 serious flaws are found.
GLM-5.1 (2026-04-07) — Z.ai a ouvert GLM-5.1 sous licence MIT, modèle phare pour l'ingénierie agentique : 58,4 sur SWE-Bench Pro, et une capacité à rester efficace sur des centaines de tours et des milliers d'appels d'outils.
Claude Mythos Preview (2026-04-07) — Anthropic announces Claude Mythos Preview, a frontier cybersecurity model reserved for Project Glasswing partners, able to find and exploit software vulnerabilities, with no public release.
Muse Spark (2026-04-08) — Meta Superintelligence Labs présente Muse Spark, son premier modèle multimodal de raisonnement, proposé sur meta.ai et l'application Meta AI sans poids ouverts, en rupture avec la ligne Llama.
GPT-5.4-Cyber (2026-04-14) — OpenAI élargit Trusted Access for Cyber à des milliers de défenseurs vérifiés et lance GPT-5.4-Cyber, une variante de GPT-5.4 plus permissive pour la sécurité défensive.
Gemini Robotics-ER 1.6 (2026-04-14) — Google DeepMind publie Gemini Robotics-ER 1.6, modèle de raisonnement incarné amélioré en compréhension spatiale et multi-vues, capable de lire jauges et voyants, disponible via l'API Gemini.
Qwen3.6 (35B-A3B and 27B) (2026-04-15) — Alibaba publie Qwen3.6 en poids ouverts : Qwen3.6-35B-A3B (MoE) le 15 avril 2026 puis Qwen3.6-27B (dense) le 22 avril, sous Apache 2.0, avec un code agentique qui dépasse celui du fleuron ouvert précédent.
Gemini 3.1 Flash TTS (2026-04-15) — Google lance Gemini 3.1 Flash TTS en préversion, un modèle de synthèse vocale expressif et pilotable par balises audio, avec dialogue multi-locuteurs et plus de 70 langues.
Claude Opus 4.7 (2026-04-16) — Anthropic rend Claude Opus 4.7 généralement disponible : un net progrès sur Opus 4.6 en ingénierie logicielle difficile, avec une vision haute résolution et un niveau d'effort xhigh, au même prix.
GPT-Rosalind (2026-04-17) — OpenAI présente GPT-Rosalind, un modèle de raisonnement spécialisé en biologie, découverte de médicaments et médecine translationnelle, proposé en aperçu de recherche par accès de confiance.
Kimi K2.6 (2026-04-20) — Moonshot AI publie Kimi K2.6, modèle ouvert multimodal de 1000 milliards de paramètres axé sur le codage de longue durée et un essaim pouvant coordonner 300 sous-agents sur 4000 étapes.
GPT Image 2 (ChatGPT Images 2.0) (2026-04-21) — OpenAI launches ChatGPT Images 2.0 and the gpt-image-2 model: OpenAI's first image model with reasoning, outputs up to 2K and better rendering of non-Latin text.
OpenAI Privacy Filter (2026-04-22) — OpenAI releases Privacy Filter, a small open-weight model under the Apache 2.0 license that detects and masks personal data in text and can run locally.
Grok Voice Think Fast 1.0 (2026-04-23) — xAI launches Grok Voice Think Fast 1.0, its flagship voice model for support and sales agents, able to reason in the background without added latency and ranked first on tau-voice Bench.
GPT-5.5 (2026-04-23) — OpenAI launches GPT-5.5, its most capable model of the time for code, research, computer use and documents, with the same per-token latency as GPT-5.4.
DeepSeek V4 Pro and V4 Flash (2026-04-24) — DeepSeek releases V4-Pro (1.6 trillion parameters, 49 billion active) and V4-Flash (284 billion, 13 billion active) in preview, both with a one-million-token context, under the MIT license.
MiMo-V2.5-Pro (2026-04-27) — Xiaomi opens up MiMo-V2.5-Pro, a MoE with 1.02 trillion parameters of which 42 billion are active, with a one-million-token context, under the MIT license, aimed at long-running agent and software engineering tasks.
NVIDIA Nemotron 3 Nano Omni (30B-A3B) (2026-04-28) — NVIDIA launches Nemotron 3 Nano Omni, an open model with 30 billion parameters of which about 3 billion are active, unifying vision, audio and language for agents, with a claimed 9 times higher throughput.
Grok 4.3 (2026-04-30) — xAI rolls out Grok 4.3, a model with built-in reasoning and a 1 M-token context, at a sharply reduced price (1.25 USD input, 2.50 USD output per million tokens).
GPT-5.5 Instant (2026-05-05) — OpenAI replaces GPT-5.3 Instant with GPT-5.5 Instant as ChatGPT's default model, with 52.5% fewer hallucinated claims on high-stakes topics.
GPT-Realtime-2 (2026-05-07) — OpenAI launches three audio models in the Realtime API: GPT-Realtime-2 for voice agents with reasoning, GPT-Realtime-Translate for live translation and GPT-Realtime-Whisper for transcription.
GPT-5.5-Cyber (2026-05-07) — OpenAI is rolling out GPT-5.5-Cyber in limited preview for critical-infrastructure defenders, a permissive variant of GPT-5.5 for penetration testing and authorized exploit validation.
Gemini Omni Flash (2026-05-19) — Google introduces Gemini Omni at I/O, a model that creates from any input, starting with video: Gemini Omni Flash generates and edits videos through conversation, with reasoning about the real world.
Gemini 3.5 Flash (2026-05-19) — Google introduces the Gemini 3.5 family at I/O with 3.5 Flash, its strongest model for agents and coding, available immediately, ahead of Gemini 3.1 Pro on Terminal-Bench 2.1 and MCP Atlas.
Mistral Medium 3.5 (2026-05-22) — Mistral AI introduces Mistral Medium 3.5, its first merged flagship model: 128 billion dense parameters, 256k context, instruction, reasoning, coding and vision, in open weights under a modified MIT license.
Anthropic raises 65 billion USD at a 965 billion USD valuation (2026-05-28) — Anthropic raises 65 billion USD in a Series H led by Altimeter, Dragoneer, Greenoaks and Sequoia, at a 965 billion USD valuation: the company surpasses OpenAI for the first time. Its run-rate revenue crosses 47 billion USD.
Claude Opus 4.8 (2026-05-28) — Anthropic launches Claude Opus 4.8, which extends Opus 4.7 with better results and more honest behavior on agentic tasks, at the same price, with 1M tokens of context by default.
Step 3.7 Flash (2026-05-29) — StepFun releases Step 3.7 Flash, a vision-language mixture-of-experts model with 198 billion parameters of which 11 billion are active, under Apache 2.0, with 256,000 tokens of context and up to 400 tokens per second.
Grok Build 0.1 (2026-05-29) — xAI opens grok-build-0.1 in public beta via the API, the agentic coding model that powers its Grok Build CLI, at 1 USD input and 2 USD output per million tokens.
MiniMax-M3 (2026-06-01) — MiniMax releases M3, a multimodal MoE of about 428 billion parameters (23 billion active) with one million tokens of context thanks to sparse attention, and weights under a community license.
Grok Imagine Video 1.5 (2026-06-03) — xAI offers grok-imagine-video-1.5-preview in API preview, which animates an image into cinematic video through natural-language instructions; general availability follows in mid-June.
Gemma 4 12B (2026-06-03) — Google DeepMind adds Gemma 4 12B to the Gemma 4 family: a mid-size open model, without separate encoders for vision and audio, that runs on a 16 GB laptop.
NVIDIA Nemotron 3 Ultra (550B-A55B) (2026-06-04) — NVIDIA releases Nemotron 3 Ultra, a 550-billion-parameter MoE with 55 billion active parameters for long-running agents, with weights, data and recipes under the OpenMDW-1.1 license.
New term: self-improving harness (2026-06-08) — A self-improving harness is an agent harness that the agent modifies itself: it analyzes its failures, proposes changes and keeps them only after regression testing. The "Self-Harness" paper formalized the idea in June 2026.
Gemini 3.5 Live Translate (2026-06-09) — Google launches Gemini 3.5 Live Translate, a near-real-time speech-to-speech voice translation model, with automatic detection of more than 70 languages and preservation of the speaker's intonation.
Claude Fable 5 and Claude Mythos 5 (2026-06-09) — Anthropic launches Claude Fable 5, a Mythos-class model made safe for general use, and Claude Mythos 5, the same model with lighter safeguards, reserved for cyberdefense and biology research.
DiffusionGemma (2026-06-10) — Google DeepMind releases DiffusionGemma, an experimental 26B open model (3.8B active) that generates text by diffusion, up to four times faster than an autoregressive model on a dedicated GPU.
Kimi K2.7 Code (2026-06-12) — Moonshot AI releases Kimi K2.7 Code, an open coding-specialized model derived from K2.6, which cuts thinking tokens by about 30% while improving long-running software engineering tasks.
SpaceX IPO, xAI included: the largest in history (2026-06-12) — SpaceX, which absorbed xAI in February, lists on Nasdaq (SPCX) after raising 75 billion USD at 135 USD per share, the largest IPO in history. The stock gains 19% on day one and the valuation tops 2 trillion USD.
GLM-5.2 (2026-06-16) — Z.ai has released GLM-5.2 under the MIT license: a 1M-token context presented as robust, the IndexShare architecture, High and Max effort levels, and 81.0 on Terminal-Bench 2.1 (Terminus-2).
Mistral OCR 4 (2026-06-23) — Mistral AI launches Mistral OCR 4, its document recognition model: 170 languages, coordinates and block types, confidence scores, with a self-hosting option for enterprises.
LongCat-2.0 (2026-06-29) — Meituan announces LongCat-2.0, a mixture-of-experts model of 1.6 trillion parameters with about 48 billion active, under the MIT license, with a one-million-token context, trained on Chinese ASICs.
Claude Sonnet 5 (2026-06-30) — Anthropic launches Claude Sonnet 5, its most agentic Sonnet: planning, use of browsers and terminals, autonomous execution at a level that recently required larger and more expensive models.
Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image) (2026-06-30) — Google launches Nano Banana 2 Lite, its fastest and cheapest Gemini image model: about 4 seconds per image and 0.034 USD for a 1K image, for high-volume workflows.
Leanstral 1.5 (2026-07-02) — Mistral AI releases Leanstral 1.5, an Apache 2.0 open source agent with 119 billion parameters (6 billion active) for Lean 4 proofs, which saturates miniF2F and solves 587 of 672 PutnamBench problems.
Hy3 (2026-07-06) — Tencent releases Hy3, a mixture-of-experts model with 295 billion parameters, 21 billion of them active, under the Apache 2.0 license, with a 256,000-token context, designed for agents and production use.
GPT-Realtime-2.1 (2026-07-06) — OpenAI releases GPT-Realtime-2.1, an update to its voice model with reasoning, and GPT-Realtime-2.1 mini, a faster and cheaper distilled version.
Robostral Navigate (2026-07-08) — Mistral AI introduces Robostral Navigate, its first embodied navigation model: 8 billion parameters, a single RGB camera, trained in simulation, to guide robots from natural-language instructions.
GPT-Live (GPT-Live-1 and GPT-Live-1 mini) (2026-07-08) — OpenAI launches GPT-Live, a new generation of full-duplex voice models that listen and speak at the same time, with GPT-Live-1 and GPT-Live-1 mini powering ChatGPT Voice.
Grok 4.5 (2026-07-08) — SpaceXAI introduces Grok 4.5, a model for code, agents and knowledge work, trained with Cursor and priced at 2 USD input and 6 USD output per million tokens.
GPT-5.6 (Sol, Terra and Luna) (2026-07-09) — OpenAI launches the GPT-5.6 family: Sol (flagship model), Terra (balance of cost and intelligence) and Luna (the most economical), with a new max effort mode and a multi-agent ultra mode.
Muse Spark 1.1 and Meta Model API (2026-07-09) — Meta launches Muse Spark 1.1 in its Meta Model API as a public preview: a multimodal reasoning model with a one-million-token context, oriented toward agents, tools and code.
Kimi K3 (2026-07-16) — Moonshot AI launches Kimi K3, a 2.8-trillion-parameter MoE (104 billion active) with a one-million-token context, presented as the first open model of the 3T class, with weights published on July 27, 2026.
Gemini 3.6 Flash (2026-07-21) — Google launches Gemini 3.6 Flash, a workhorse model that is more token-efficient than 3.5 Flash, cheaper on output and better at code, business knowledge and computer use, alongside 3.5 Flash-Lite and 3.5 Flash Cyber.
Gemini 3.5 Flash-Lite (2026-07-21) — Google launches Gemini 3.5 Flash-Lite, its fastest and most cost-effective model in the 3.5 series, at about 350 tokens per second, for agentic search, document processing and high-throughput workloads.
Claude Opus 5 (2026-07-24) — Anthropic launches Claude Opus 5, a thoughtful and proactive model close to the intelligence of Claude Fable 5 at half the price, at the same pricing as Opus 4.8.
Digital Omnibus on AI enters into force: high-risk obligations postponed (2026-07-27) — Regulation (EU) 2026/1744, the Digital Omnibus on AI, enters into force: the first amendment to the AI Act, it postpones high-risk system obligations to December 2, 2027 and August 2, 2028, without changing the general-purpose model rules.
GPT Transcribe and GPT Live Transcribe (2026-07-28) — OpenAI releases GPT Transcribe for file transcription and GPT Live Transcribe for low-latency streaming transcription, with keyword and language hints.
Grok Voice Think Fast 2.0 (2026-07-29) — SpaceXAI launches Grok Voice Think Fast 2.0, a smarter speech-to-speech voice model, with first audio in 0.70 seconds and pricing of 0.08 USD per minute of audio.
Lyria 3.5 (2026-07-29) — Google launches Lyria 3.5, a new version of its music generation model, with better musicality, lyrics that follow the prompt more closely and more expressive vocals, first in Flow Music.
Gemini Robotics ER 2 (2026-07-30) — Google DeepMind launches Gemini Robotics ER 2, its most capable embodied reasoning model, which tracks task progress in continuous video and orchestrates several robots through the Live API.
Shieldstral (2026-08-04) — Mistral AI launches Shieldstral 1.0, a multimodal safety classifier with 3 billion parameters under the Apache 2.0 license, which applies policies written in natural language without retraining.
Muse Spark 1.2 and Muse Code (2026-08-05) — Meta launches Muse Code, a terminal coding agent, together with Muse Spark 1.2, available in the Meta Model API with two pricing tiers depending on the use of data for training.
Muse Glimmer 30B (2026-08-10) — Meta releases Muse Glimmer, a dense 30-billion-parameter open weights model under the Apache 2.0 license, distilled from Muse Spark and designed for local agents on a PC or Mac with a single GPU.
GPT-5.6-Cyber (2026-08-10) — OpenAI launches GPT-5.6-Cyber, a specialized model built on GPT-5.6 Sol, and opens two Daybreak access tiers: Blue for everyday defense, Red for exploitation and advanced research.
xAI launches Grok Bot, an agent with its own cloud computer (2026-08-11) — xAI launches Grok Bot in beta, an agent that completes tasks end to end across apps and websites from its own cloud computer. The beta, on desktop and iOS, is limited to certain subscribers.
NVIDIA Nemotron 3.5 Lightning (30B-A3B) (2026-08-11) — NVIDIA releases Nemotron 3.5 Lightning, an open MoE with 30 billion parameters, 3 billion of them active, designed for fast agent execution and for specialization through post-training, under the OpenMDW-1.1 license.
Grok 4.6 (2026-08-12) — SpaceXAI launches Grok 4.6, a model for long-running agents and visual or interactive work, at USD 2 input and USD 6 output per million tokens, with a 500k-token context.
Qwen3.8-27B (2026-08-13) — Alibaba releases Qwen3.8-27B, a dense multimodal model with 27 billion parameters under the Apache 2.0 license, which outperforms the closed Qwen3.7-Plus on several agentic tests according to the vendor's figures.
Qwen3.8-Max (open weights: Qwen3.8-2.4T-A95B) (2026-08-13) — Alibaba launched Qwen3.8-Max through its API on August 3, 2026, then released its weights (Qwen3.8-2.4T-A95B, 95 billion active parameters) under a house license, a first for a Max-class model.
Gemini 3.7 Flash (2026-08-13) — Google launches Gemini 3.7 Flash, its smartest workhorse model for coding and agents, three weeks after 3.6 Flash, with an introductory price half as high.
DeepSeek V4 Pro (final version) (2026-08-13) — DeepSeek releases the final version of V4-Pro (V4-Pro-0813), with markedly strengthened agent capabilities, three levels of thinking effort and a pricing grid with peak and off-peak rates.
GLM-5.3 (2026-08-14) — Z.ai launched GLM-5.3 on August 14, 2026, same base as GLM-5.2 but with strengthened post-training: better long-form coding, cybersecurity (CyberGym 84.5). Open weights on August 25, under its own GLM-5.3 license, not MIT.
GLM-5.3-Flash (and GLM-5.3-FlashX) (2026-08-26) — Z.ai released GLM-5.3-Flash, a natively multimodal MoE with 320B parameters (18B active), hybrid attention and 1M context, under the MIT license, priced at USD 0.15 input and USD 0.50 output per million tokens.
Qwen3.8-Flash-Next (2026-08-26) — Alibaba releases Qwen3.8-Flash-Next, an open-weights multimodal model with 125 billion parameters, 6 billion of them active, presented as the experimental preview of the architecture of the future Qwen4 generation.
Gemini 3.5 Transcribe (2026-08-26) — Google launches Gemini 3.5 Transcribe, its most accurate speech recognition model, which turns audio into clean, formatted text, in streaming (Live API) or on recorded files.
Gemini Omni 1.1 Flash (2026-08-27) — Google launches Gemini Omni 1.1 Flash, a production-ready version of its video model, with scene extension, start and end frames, 360p drafts and 4K upscaling.
Claude Fable 5.1 and Claude Mythos 5.1 (2026-09-01) — Anthropic launches Claude Fable 5.1 and Claude Mythos 5.1, presented as its most advanced models in coding and knowledge work, less costly than Fable 5 and with more precise safeguards.
Muse Spark 1.3 (2026-09-02) — Meta releases Muse Spark 1.3, its most advanced proprietary model, available in Muse Code and the Meta Model API, with gains in agentic tasks and reduced use of tools and tokens.
Gemini 3.8 Flash (2026-09-02) — Google launches Gemini 3.8 Flash, its most intelligent Flash model, better at code, agents and multi-step reasoning at the same price as 3.7 Flash, along with the cybersecurity variant 3.8 Flash Cyber.
GPT-6 Astra (2026-09-03) — OpenAI launches GPT-6 Astra, presented as its most intelligent and best-aligned model, with new highs in computer use, browsing, software engineering and cybersecurity.
Meta launches Muse, a consumer personal agent (2026-09-08) — Meta launches Muse, a consumer personal agent that carries out tasks (emails, travel, forms, purchases) inside a secure virtual machine, powered by the Muse Spark model. US rollout; plans at $20 and $100 per month.
Mistral raises 3 billion euros, the largest raise by a European tech company (2026-09-08) — Mistral AI closes a 3 billion euro Series D led by Samsung Electronics, at a valuation above 21 billion euros: the largest equity raise ever by a European tech company, in support of sovereign open-weight AI.
GPT Image 2.5 (Sunburst and Flare) (2026-09-08) — OpenAI launches ChatGPT Images 2.5 and two API models: GPT-Image-2.5 Flare, fast for everyday use, and GPT-Image-2.5 Sunburst, more precise for detailed editing.
GPT-Live-1 (API) (2026-09-10) — OpenAI opens GPT-Live-1 in the API: a full-duplex voice model that listens and speaks at the same time, priced at $0.05 per minute, with reasoning delegated to a text model.
DeepSeek V4.1 Flash (2026-09-10) — DeepSeek releases V4.1-Flash, a multimodal MoE of 552 billion parameters with a causal encoder-decoder architecture, with 8 billion active parameters on read, 16 on generation, under the MIT license.
Dario Amodei: "We Must Pace the Frontier" (2026-09-12) — On September 12, 2026, Anthropic CEO Dario Amodei publishes the essay "We Must Pace the Frontier": slow model capabilities so alignment can catch up. Sam Altman and Elon Musk back it.
Gemini 3.8 Live and 3.8 Live Extended Thinking (2026-09-15) — Google launches Gemini 3.8 Live and 3.8 Live Extended Thinking, two near-real-time voice dialogue models, the second reasoning while speaking, for voice agents, Search Live, Gemini Live and Workspace.
Jev (2026-09-15) — TypeSafe AI, a San Francisco start-up, launches Jev in limited early access: a “System One” model that returns typed decisions with probabilities rather than text.
Grok Voice Transcribe 2.0 (2026-09-18) — SpaceXAI releases Grok Voice Transcribe 2.0, a transcription model that is twice as accurate, multilingual, with word-level timestamps, speaker separation and multichannel support, at the existing pricing.
Grok 4.7 (2026-09-21) — SpaceXAI launches Grok 4.7, its most capable model for coding and knowledge work, built on a larger base than Grok 4.6, at the same pricing of 2 USD input and 6 USD output.
Claude Opus 5.5 (2026-09-22) — Anthropic launches Claude Opus 5.5, the first model of the Claude 5.5 family, on par with Fable 5.1 on most tasks at a cost 40% lower than Opus 5.
GPT-6 Sol and GPT-6 Luna (2026-09-22) — OpenAI expands the GPT-6 family with Sol and Luna, trained with the same methods as Astra but cheaper: API prices drop by 50% compared with the promotional pricing of GPT-5.6.
Gemini 3.8 Flash TTS and 3.8 Flash-Lite TTS (2026-09-23) — Google launches Gemini 3.8 Flash TTS and 3.8 Flash-Lite TTS, two text-to-speech models with voice creation by prompt, voice replication from 30 seconds of audio and a library of more than 2,000 voices.
Claude Sonnet 5.5 (2026-09-28) — Anthropic launches Claude Sonnet 5.5, the second model of the 5.5 family: more than 30% faster than Sonnet 5 and up to 30% cheaper for most tasks, at the same price per token.
NVIDIA Open Agent Safety Platform (2026-09-28) — NVIDIA launches the Open Agent Safety Platform, an open software platform and reference design to secure AI agents from testing to deployment, built on OpenShell, the Vera CPU and the Sentry watchdog.
GPT-6.1 Sol (2026-09-29) — OpenAI launches GPT-6.1 Sol, an upgrade to GPT-6 Sol that approaches GPT-6 Astra in agentic coding, computer use and professional work, at one-fifth of Astra's standard API prices.
DevDay 2026: OpenAI turns ChatGPT into an agent platform (2026-09-29) — On September 29, 2026, OpenAI held its DevDay: more than 20 announcements, including dots agents, GPT-6.1 Sol, the Ultrafast tier, computer use in the Agents API and ChatGPT plugin extensions.
Dots, OpenAI's always-on agents (2026-09-29) — OpenAI launches dots, always-on agents powered by GPT-6 Astra: each has its own cloud computer, connects to the chosen apps and moves projects forward proactively, under rules set by the user.
Gemini 4 Argon (2026-09-30) — Google introduces Gemini 4 Argon, its new flagship model built for software engineering, knowledge work and cyber defense. It is first limited to defenders in the Fairwind Program; paid API and AI Ultra access will follow.
California: SB 947 bans layoffs decided by AI alone (2026-09-30) — California becomes the first US state to bar employers from firing or disciplining a worker on the sole basis of an automated system. Signed by Gavin Newsom, SB 947 takes effect on July 1, 2027.
Google DeepMind: SynthID Bio, a watermark for AI-designed proteins (2026-09-30) — Google DeepMind presents SynthID Bio, published in Nature: an imperceptible, verifiable watermark in AI-designed protein sequences and predicted 3D structures, with protein function preserved in laboratory tests.
Open decision models: Cloudflare Clef and Amazon Strands Decider (2026-10-01) — Cloudflare releases Clef (27B) and Clef-flash (9B), and Amazon Strands Labs releases Strands Decider (2B), all under Apache 2.0. These models generate no text: they return typed answers with probabilities.
Claude Code: TypeScript “mods” to customize the agent (2026-10-01) — Anthropic introduces mods for Claude Code: small TypeScript functions that rewrite prompts, control tool calls, redact secrets and change the interface. They are shared like plugins, through the Claude directory.
United States: the White House creates the "Super Intelligence Force", a federal AI task force (2026-10-04) — On October 4, 2026, Donald Trump announces the "Super Intelligence Force", a federal task force to coordinate the government's AI efforts. It is led by Jay Clayton, Director of National Intelligence, who also serves as AI czar.
Reflection AI introduces Beam, a 501-billion-parameter open model (2026-10-06) — Reflection AI introduces Beam, an open-weight mixture-of-experts model of 501 billion parameters (23 billion active), trained on 23.8T tokens. Weights and technical report are announced for October.
Nano Banana 2.1 (2026-10-06) — Google launches Nano Banana 2.1 (gemini-nano-banana-2.1), a Flash-level image generation and editing model. It replaces gemini-3.1-flash-image, which will be shut down on October 29, 2026.
Mistral Large 4 ("Le Chonk") (2026-10-06) — Mistral AI introduces Mistral Large 4, a multimodal mixture-of-experts model of about 1 trillion parameters (49 billion active). Public preview on Mistral Studio; open weights promised for the end of October.