All 66 verified AI news entries from 2025, including 40 model releases, in date order with a link to each official source.
Showing 66 of 66 entries
GoogleModelsMajor
Gemini 3 Flash pairs Gemini 3-class reasoning with Flash speed at $0.50/$3 per million tokens, and replaced 2.5 Flash as the default in the Gemini app and AI Mode in Search.
Image: Google
xAIVoiceMajor
Developers can build real-time voice agents on the stack behind Grok in Tesla vehicles, with tool calling, web and X search, dozens of languages, and compatibility with the OpenAI Realtime API format.
Image: xAI
NVIDIAModels, Important
NVIDIA announced Nemotron 3 in Nano, Super, and Ultra sizes, shipping the 30B Nano with a 1M-token context and open training data on day one.
Image: NVIDIA
The White HousePolicyMajor
Executive Order 14365 created a Justice Department task force to contest state AI laws and asked for a single federal standard.
OpenAIModels, Important
GPT-5.2 became the GPT-5 family flagship, adding gains in instruction following, vision, coding and tool calling.
Image: OpenAI
Linux FoundationPlatforms, Important
The Linux Foundation launched the Agentic AI Foundation (AAIF) as a neutral home for Anthropic's Model Context Protocol, Block's goose, and OpenAI's AGENTS.md.
Image: Linux Foundation
n8nTools, Important
n8n released version 2.0 in beta on December 8, 2025, with isolated Code node execution, stricter defaults, and a Save/Publish split; the stable 2.0.x line followed on December 15.
Image: n8n
AmazonAgentsMajor
At re:Invent 2025, AWS previewed three autonomous agents: the Kiro autonomous agent for development, AWS Security Agent for design reviews and penetration tests, and AWS DevOps Agent for incident response. The security and DevOps agents were generally available by April 2026.
Image: Amazon
AmazonModels, Important
At re:Invent, AWS made Nova 2 Lite generally available with a 1M-token context, previewed Nova 2 Pro and Omni, and added services for custom models and browser agents.
AmazonAgents, Important
Policy checks every agent tool call against rules written in natural language, Evaluations scores live agent behavior with 13 built-in evaluators, and AgentCore Memory gained episodic learning. Policy and Evaluations started in preview and later became generally available.
Image: Amazon
Mistral AIModels, Important
Mistral 3 includes a 675B-parameter mixture-of-experts flagship and dense 3B, 8B, and 14B models, all open-weight under Apache 2.0 with image understanding.
Image: Mistral AI
DeepSeekModels, Important
DeepSeek-V3.2 became the default model behind its app and API, with thinking built into tool use, while V3.2-Speciale was offered briefly for maximum reasoning.
RunwayCreative, Important
Runway's Gen-4.5 text-to-video model ranked first on the Artificial Analysis Text-to-Video leaderboard at launch and is included in all paid Runway plans.
Image: Runway
Mario ZechnerAgents, Important
pi is a terminal coding agent built from small MIT-licensed packages: a multi-provider LLM API, an agent loop and a terminal UI. It uses a system prompt and tools under 1,000 tokens, four default tools and no permission prompts. It later powered OpenClaw's agent.
Model Context ProtocolPlatforms, Important
The Model Context Protocol published a backward-compatible revision that adds experimental tasks, URL-mode elicitation, tool calling inside sampling, and simpler OAuth client registration.
SunoCreative, Important
Suno partnered with WMG to build new models on licensed music, let opted-in WMG artists allow use of their voices and likenesses, and require a paid plan to download songs.
Image: Suno
AnthropicModelsMajor
Anthropic launched Claude Opus 4.5, pitched as its strongest model for coding, agents and computer use, and cut Opus pricing to $5 input and $25 output per million tokens.
Google DeepMindCreative, Important
Nano Banana Pro (Gemini 3 Pro Image) improves in-image text, fact-grounded infographics and multi-image editing, with 2K and 4K output and SynthID watermarking.
Image: Google DeepMind
xAIAgents, Important
xAI positioned Grok 4.1 Fast as its strongest tool-calling model, with a 2M-token context window and low per-token pricing, alongside server-side agent tools.
Image: xAI
GoogleModelsMajor
Google released Gemini 3 Pro in preview across the Gemini app, Search, the API and Vertex AI, and introduced Google Antigravity, a development environment where agents work in the editor, terminal and browser.
Image: Google
MicrosoftAgents, Important
At Ignite 2025, Microsoft introduced Agent 365 to track, govern and secure AI agents from any platform, plus IQ layers that give agents business context. Agent 365 opened as early access through the Frontier program.
Image: Microsoft
xAIModels, Important
Grok 4.1 became the default in Auto mode on grok.com, X, and the mobile apps, with xAI focusing on style, personality, and emotional intelligence and reporting a top LMArena text ranking at launch.
Image: xAI
OpenAIModels, Important
GPT-5.1 arrived in the API as the new GPT-5 flagship, alongside Codex variants and 24-hour prompt cache retention.
Image: OpenAI
Terminal-BenchBenchmarks, Important
The command-line agent benchmark released a harder, more carefully verified version and a new package for running agent evaluations in cloud containers.
OpenCodeAgents, Important
OpenCode, the MIT-licensed coding agent from the team behind the SST framework, reached 1.0 on October 31, 2025 with a new terminal interface. It works with models from many providers, including local ones, and its repository passed 220,000 GitHub stars by November 2026.
MiniMaxModels, Important
MiniMax-M2 launched with open weights and an API priced at $0.30 input and $1.20 output per million tokens, which MiniMax put at 8% of Claude Sonnet's price at about twice the speed.
AnthropicModels, Important
Claude Haiku 4.5 matched Claude Sonnet 4's coding performance at one-third the cost and more than twice the speed, priced at $1/$5 per million tokens.
OpenAICreativeMajor
Sora 2 generates video with dialogue and sound effects and more realistic physics, and launched in an invite-based iOS app in the US and Canada with a consent-based characters feature.
Image: OpenAI
AnthropicModelsMajor
Claude Sonnet 4.5 reported 77.2% on SWE-bench Verified and 61.4% on OSWorld, stayed focused for over 30 hours on complex tasks, kept $3/$15 pricing and arrived with the Claude Agent SDK.
DeepSeekModels, Notable
DeepSeek-V3.2-Exp introduced DeepSeek Sparse Attention for cheaper long-context work, matched V3.1-Terminus on benchmarks, and came with an API price cut of 50% or more.
SunoCreative, Notable
Suno v5 promised clearer, more immersive audio, more authentic vocals, finer creative control and better genre handling, and let users remaster older tracks.
Image: Suno
OpenAIModels, Important
GPT-5-Codex adjusts its thinking time to the task, worked independently for more than 7 hours in OpenAI's tests, and was trained for code review across the Codex CLI, IDE extension, cloud and GitHub.
Image: OpenAI
OpenAIVoice, Important
The gpt-realtime speech-to-speech model improved instruction following, tool calling and natural speech, while the Realtime API gained SIP phone calling, remote MCP servers and image input at 20% lower prices.
Image: OpenAI
Google DeepMindCreative, Important
Gemini 2.5 Flash Image added character consistency, multi-image fusion, prompt-based edits and Gemini world knowledge, priced at about $0.039 per image, with SynthID watermarks.
Image: Google DeepMind
DeepSeekModels, Important
DeepSeek-V3.1 merged chat and reasoning into one hybrid model with 128K context, stronger tool use and agent skills, Anthropic API compatibility and open weights.
OpenAIModelsMajor
GPT-5 combined a fast model, a deeper thinking model and a real-time router in ChatGPT, reached free users, and launched in the API in three sizes with new verbosity and minimal-reasoning controls.
Image: OpenAI
OpenAIModelsMajor
The Apache 2.0 gpt-oss models are mixture-of-experts reasoning models: the 120b nears o4-mini on core reasoning on a single 80 GB GPU, and the 20b matches o3-mini and runs in 16 GB of memory.
Image: OpenAI
AnthropicModels, Notable
Claude Opus 4.1 upgraded Opus 4 on agentic tasks, real-world coding and research at the same price, and Anthropic recommended it as a drop-in replacement.
Image: Anthropic
Z.aiModels, Notable
GLM-4.5 (355B total, 32B active) and GLM-4.5-Air (106B total, 12B active) offer thinking and instant modes, 128K context and native function calling, with open weights.
Alibaba QwenModels, Notable
Qwen3-Coder-480B-A35B-Instruct supports 256K context natively (1M with extrapolation), was trained with large-scale code and agent reinforcement learning, and shipped with a terminal agent.
Moonshot AIModels, Important
Kimi K2 is a mixture-of-experts model with 1T total and 32B active parameters, trained with the MuonClip optimizer and tuned for tool use, released as Base and Instruct weights.
xAIModels, Important
Grok 4 was trained with reinforcement learning at pretraining scale to use tools such as code and search, launched in the API with a 256K context window, and came with a new SuperGrok Heavy tier.
Image: xAI
Google DeepMindModels, Important
Stable Gemini 2.5 Pro and Flash arrived for production use, and the new 2.5 Flash-Lite preview targeted high-volume, low-latency tasks with thinking budgets and a 1M-token context.
Image: Google DeepMind
Mistral AIModels, Notable
Magistral Small (24B, Apache 2.0) and the enterprise Magistral Medium reason step by step across many languages, with a fast Think mode in Le Chat.
Image: Mistral AI
ElevenLabsVoice, Important
Eleven v3 added inline audio tags such as [whispers] and [excited], multi-speaker dialogue and support for 70+ languages, aimed at media rather than real-time agents.
Image: ElevenLabs
Black Forest LabsCreative, Notable
FLUX.1 Kontext models accept text plus reference images to make targeted edits while keeping characters consistent, with [pro] and [max] in the API and a 12B open-weight [dev] variant.
Image: Black Forest Labs
AnthropicModelsMajor
Claude Opus 4 led SWE-bench and Terminal-bench at launch and can work on long tasks for hours; both models can use tools during extended thinking, and Claude Code reached general availability.
Image: Anthropic
Google DeepMindCreativeMajor
At I/O 2025, Veo 3 became the first Veo model to produce synchronized audio such as dialogue and ambient sound, while Imagen 4 improved fine detail and typography at up to 2K.
Image: Google DeepMind
Mistral AIModels, Notable
Mistral Medium 3 targeted near-frontier performance at $0.40 input and $2 output per million tokens, with self-hosting possible on four or more GPUs.
SunoCreative, Notable
Suno v4.5 broadened genre coverage and mashups, improved vocals and prompt interpretation, sped up generation and extended songs to 8 minutes.
Image: Suno
Alibaba QwenModels, Important
Qwen3 spans six dense models and two mixture-of-experts models up to 235B parameters, each able to switch between step-by-step thinking and fast answers, with support for 119 languages.
OpenAIModelsMajor
o3 and o4-mini were trained to decide when to search, run code or inspect images during reasoning, replaced o1 and o3-mini in ChatGPT, and arrived with the open-source Codex CLI.
Image: OpenAI
OpenAIModels, Important
The API-only GPT-4.1 family improved coding and instruction following over GPT-4o, raised context to 1M tokens, cut prices, and started the retirement of GPT-4.5 Preview.
Image: OpenAI
MetaModels, Important
Llama 4 Scout (109B total, 10M-token context) and Maverick (400B total) are natively multimodal MoE models with 17B active parameters, distilled from the still-training Behemoth.
MidjourneyCreative, Notable
Midjourney's V7 alpha improved prompt understanding, textures and the coherence of hands and bodies, turned on personalization by default and added a cheaper, faster Draft Mode.
RunwayCreative, Notable
Gen-4 uses reference images to keep characters, objects and locations consistent across shots, without fine-tuning, starting with 5- and 10-second clips.
Image: Runway
Google DeepMindModelsMajor
Gemini 2.5 Pro Experimental reasons before answering, shipped with a 1M-token context window and led several math, science and coding benchmarks at launch.
Image: Google DeepMind
OpenAICreativeMajor
Native GPT-4o image generation brought accurate text rendering, detailed prompt following and chat-based editing to ChatGPT users, including the free tier, with C2PA metadata on every image.
Image: OpenAI
OpenAIVoice, Notable
gpt-4o-transcribe and gpt-4o-mini-transcribe improved on Whisper's word error rate, and gpt-4o-mini-tts let developers describe how a voice should speak.
Image: OpenAI
Google DeepMindModels, Important
Gemma 3 came in 1B, 4B, 12B and 27B sizes with a 128K context window, function calling, support for over 140 languages and vision in the larger models.
Image: Google DeepMind
OpenAIModels, Important
GPT-4.5 scaled up unsupervised pre-training for broader knowledge and fewer hallucinations, launching first for ChatGPT Pro and as an expensive API preview.
Image: OpenAI
AnthropicModelsMajor
Claude 3.7 Sonnet can answer instantly or use extended thinking with a configurable budget, at unchanged pricing, alongside a research preview of the Claude Code terminal agent.
xAIModels, Important
Grok 3 and Grok 3 mini arrived in beta for X Premium subscribers, trained on the Colossus cluster, with a visible reasoning mode and a DeepSearch research agent.
Image: xAI
Google DeepMindModels, Important
Gemini 2.0 Flash reached general availability with a 1M-token context window, while 2.0 Pro Experimental (2M tokens) and the low-cost 2.0 Flash-Lite entered testing.
Image: Google DeepMind
OpenAIModels, Important
o3-mini reached ChatGPT and the API with three reasoning effort levels, function calling and Structured Outputs, and became the first reasoning model offered to free ChatGPT users.
Image: OpenAI
DeepSeekModelsMajor
DeepSeek-R1 claimed reasoning performance on par with OpenAI o1, shipped with MIT-licensed weights, six distilled smaller models and a low-cost API.