IA en un minutoTopicsModels
Topic · 62 stories
Models
AI model launches: what they do better, what they cost and who can use them.

Runway opens the weights of Praxis-1, its robot control model
Runway opens the weights of Praxis-1, its 27-billion-parameter model for robots, including for commercial use.

Google launches Gemini 4 Hydrogen, its fast model from the Gemini 4 family, in Search, the app and the API
Gemini 4 Hydrogen, Google's fast model, reaches Search for everyone from today and beats Argon, according to the company.

Google introduces the Gemini Decisions API and its Gemini Decision 1 model to choose between options in milliseconds
Google presents Gemini Decision 1, a model that picks between options in milliseconds instead of writing text.

Google DeepMind presents AlphaGenome 2, its model that predicts how DNA variants affect specific human cell types
Google DeepMind presents AlphaGenome 2, a model that estimates the effect of DNA variants on each type of human cell.

Google DeepMind unveils Gemini Robotics 2.5, an AI model for controlling robots that is rolling out to its robotics partners
Google DeepMind unveils Gemini Robotics 2.5, a model for handling robots that has already been tested at Mercedes-Benz plants.

A thousand Claude Fable 5.2 agents formalize Fermat's Last Theorem in 30 hours
Claude Fable 5.2 needed 30 hours to formalize the Fermat proof with 1,000 agents; the first time it took 11 days.

Meta presents Muse Ember, a lightweight and affordable model for AI agent tasks
Meta presents Muse Ember, its smallest and most efficient model, built to cut the cost of the repetitive work of AI agents.

Anthropic tests Claude on electronics design: three prototypes that work
Anthropic has Claude Fable 5.2 design three electronic devices with Diode Computers, and the prototypes work without circuit fixes.

Anthropic launches Claude Fable 5.2 for everyone and reserves Claude Mythos 5.2 for verified organizations
Anthropic opens Claude Fable 5.2 to all users and keeps Mythos 5.2 for verified organizations, at a lower price.

OpenAI Says GPT-6 Astra, One of Its AI Models, Designed a Physical Device in Under a Week
OpenAI shows how far a model can go in designing a physical device, and where it still needs human review.

Tencent Unveils Hy4, the First in Its Model Family With Vision and Computer Use
Tencent launches Hy4 with vision and computer use, and says it trained it without offensive cybersecurity material.

NVIDIA Adds Four Open Models to Its Cosmos 3 Family to Simulate Worlds Interactively
NVIDIA opens up four models for generating driving and robotics environments that can be downloaded and adapted.

Google Launches Gemini Omni 1.2 Flash, Its Video Generation Model, With 15-Second Clips and a 30% Lower Price
Google makes the videos its model generates longer, lets users fix them by chatting and cuts the price, now for developers and businesses.

Phonic unveils Phonic-1, a voice AI that starts processing before the user finishes speaking
Phonic-1 answers in 1.1 seconds in its demo, against 15.7 for GPT-Live; the figures come from the company itself.

AssemblyAI launches AssemblyAI-TTS-1, its first text-to-speech model for voice agents
AssemblyAI releases a text-to-speech model at $15 per million characters, with no annual contract and more than 4,000 voices.

Liquid AI launches LFM3, a family of six open models for audio and multimodal use on-device
Liquid AI releases LFM3, six open models of about 2 billion parameters covering text, images, video and audio.

Giotto launches Solo and Alto, its Swiss-built sovereign AI reasoning models
Giotto presents two Swiss reasoning models, with Alto free for all, for a cloud in Europe or your own hardware.

Google DeepMind announces Genie 4, its new world model with native audio, for now only for trusted testers
Genie 4 simulates interactive worlds in real time with its own audio, but for now only Google's trusted testers use it.

NVIDIA releases Nemotron 3.5 Transcribe, an open transcription model for up to 40 languages
NVIDIA releases Nemotron 3.5 Transcribe, an open model that transcribes recorded audio in up to 40 languages.

Google unveils Gemini 4 Carbon, its strongest model, and opens it first only to cybersecurity defenders
Google's strongest model, Gemini 4 Carbon, reaches only cybersecurity defenders, with no date for anyone else.

Mistral AI releases the open weights of Mistral Large 4, its model with 1 trillion parameters for coding and cyber defense
Mistral AI releases the weights of Mistral Large 4, a model with 1 trillion parameters that any company can host on its own servers.

OpenAI says its models produced three results in math, computer science and physics and asks outside experts to shepherd them
If the scientific community confirms them, an AI would have settled conjectures that specialists considered open.

OpenAI launches GPT-6.1 Terra, a faster and cheaper model for using the computer
Automating office tasks with an agent that operates the computer is now considerably cheaper with OpenAI.

Cadence launches NVIDIA-backed AI models that chipmakers can run on their own premises
Chipmakers will be able to use AI specialized in design without their blueprints leaving their own premises.

Canva launches its own AI image generator to compete with Google and OpenAI
Canva now has its own image model, designed for layouts with legible text and open to all its users.

Jack Dorsey's Block unveils Dessa, a voice model trained only on human speech that is not yet public
An AI that understands voice without turning it into text could pick up emotion and rhythm, which today's voice assistants lose.

Moonshot AI launches Kimi K3.1 Pro, its new flagship model
An open-weight model lands within reach of the strongest closed ones in coding, and it comes with a faster version.

Hugging Face unveils Carbon-X, an open model you can ask questions about DNA
Hugging Face releases an open DNA model that comes close to the leading ones for far less money and can run on in-house machines.

ElevenLabs unveils Eleven Realtime Translation, its live voice translation model, in alpha for enterprises
A call between two languages may no longer need an interpreter, although for now only in testing and for enterprise customers.

KRAFTON, the company behind PUBG, releases Raon-1 mini, an open model built from scratch
KRAFTON enters open language models with one that fits on a single graphics card and puts Korean and agents first.

ElevenLabs launches Eleven Realtime, its real-time speech-to-speech models for building voice agents
Eleven Realtime offers voice agents with two levels of reasoning and a direct route away from OpenAI's Realtime API.

Mistral launches Voxtral 3, its new voice models with an SDK for building real-time voice agents
Mistral brings together, from a single provider, what an assistant needs to take calls, with control over where the audio stays.

NVIDIA releases Cosmos3-H-Dreams, an AI surgery simulator that generates video of the operation in real time
A simulator that learns from real surgeries lets teams rehearse surgical robot control without touching a patient.

Anthropic shows how Claude designed proteins that bind to scorpion neurotoxins
AI can already propose, within days, proteins that bind to hard toxins; turning them into a real treatment is what remains to be done.

OpenAI says its models have improved at COBOL and other legacy programming languages
Banks and essential services that still rely on COBOL or Fortran gain another way to maintain and modernize that software.

Cloudflare unveils Clairic, an AI model for pulling data out of text and images
Cloudflare proposes handing data extraction to a fast, cheap model and keeping language models for writing.

Inception unveils Mercury 3, a diffusion language model that generates 1,068 tokens per second
Diffusion is aiming to compete with models that write one token at a time in tasks where every second of waiting counts.

MiniMax launches M3.1 Flash, a fast and low-cost model for agents
A model for agents at a rock-bottom price changes the math on how much it costs to automate each task.

DeepSeek updates V4.1 Flash: better at coding and cheaper per task
DeepSeek improves its lightweight model at coding and lowers the cost of each task, with open weights that anyone can download.

NVIDIA launches Nemotron 3.5 Super VL, an open multimodal model for agents
A single open model that sees and reasons saves building two systems, and each company can tune it and deploy it wherever it wants.

Claude Opus 5.5 fast mode no longer requires an access request
Companies using Claude Opus 5.5 through the API can now pay more for faster answers without waiting for Anthropic's approval.

OpenAI launches GPT-Live-1.1, an update to its voice model with new voices and more natural accents in ChatGPT
ChatGPT gets voices in six languages and fixes accents, the weak spot OpenAI admitted in its voice model.

Meta releases Muse Spark Open, an open model designed with Apple to run on two Mac Studio computers
Meta returns to open models with a large one that, according to the company, already runs on a desktop with two Mac Studio computers.

Z.ai unveils GLM-Legal, an AI model for legal work that is only in beta for now
Small, specialized models are starting to be measured against the big ones on law firm work, though here the figures come from the maker.

Meta introduces Muse Wonder and Muse Wonder Full, its most capable models for agentic tasks
Meta is back in the leading group with models built for agents that remember across conversations and stay on a hard task.

Microsoft launches Microsoft-Decision-1, an AI model for fast decisions in applications and agents
Microsoft opens a category of cheap, fast models that decide instead of write, built to steer agents and workflows.

OpenAI starts rolling out Ultrafast, the fastest mode for GPT-6.1 Sol, in the API, Codex, and ChatGPT Work
Paying more for faster answers becomes one more option in OpenAI's API and its coding tools.

Google introduces Nano Banana 2.1, its new model for creating and editing images, based on Gemini 3.6 Flash
Google cuts the price of its API images and promises clearer text on posters, diagrams and infographics.

OpenAI brings GPT-6 to all ChatGPT users, including those on the free plan
GPT-6 is no longer just for paying users: it reaches all of ChatGPT with visual, interactive answers inside the chat.

Anthropic launches Claude Haiku 5.5, its cheapest and fastest small model yet
Anthropic is making its small model cheaper for repetitive, high-volume tasks, and is also lowering the cost of using Sonnet 5.5 with agents.

OpenAI publishes 722 math manuscripts from an internal model it has not yet released
Many results can already be checked by computer, but not all: with AI writing math, verifying matters as much as producing.

Google launches EmbeddingGemma 2, an open model that links text, images, audio and video on the device itself
It lets people search photos, audio and video without taking data off the device, useful for anyone handling sensitive files.

Mistral unveils Large 4, a one trillion-parameter multimodal model that is in public preview for now
According to Mistral, it already competes with the strongest open models in the world, and with open weights anyone will be able to download it.

TII unveils Falcon-Emirati, a language model focused on the Emirati dialect and culture
TII shows that model size alone is not enough to master a dialect: a small model trained on purpose beats much larger ones.

Reflection AI unveils Beam, its first open model, as a Western alternative to Chinese models
Until now, Chinese companies have led open AI, and Reflection AI wants to be the American alternative.

OpenAI makes GPT-6 Astra and GPT-6.1 Sol about 50% faster on ChatGPT subscriptions
People who code with these models through their ChatGPT subscription get faster answers without changing anything.

Aleph Alpha launches Kolibri-1, an open model with 78 billion parameters
It can run on in-house servers without sending internal data to third parties, but only in German and English.

Microsoft launches MAI-Transcribe-2-Streaming, its first real-time transcription model
A voice agent can start working out the request before the customer finishes speaking.

Google unveils Gemini 4 Argon, its new frontier model, with limited access for now
It cannot be used yet, so for now the only thing to do is follow it closely.

Anthropic launches Claude Sonnet 5.5
Everyday tasks get faster and cheaper without a change in price.

OpenAI unveils GPT-6.1 Sol at one-fifth of the price
A model this capable at a lower price makes automating tasks more affordable for a small business.

ElevenLabs launches Eleven v4, its most expressive voice model
A voice agent that answers calls can sound more natural without losing speed.