Claude Haiku 5.5 costs ten cents per million input tokens
Anthropic's smallest, fastest Claude got much cheaper: about 75% less to run than the last Haiku. Made for quick, high-volume jobs like summaries. anthropic.com

JetBrains opens Mellum2.1, a 12B model for coding agents
A small, free coding model from JetBrains, the makers of IntelliJ, that runs on your own computer. Trained to work as a fast helper inside coding agents. blog.jetbrains.com

Google DeepMind ships Nano Banana 2.1 for image editing
Google's model for making and editing pictures by just describing what you want. It's in the Gemini app, AI Studio and Search's AI Mode. deepmind.google

EmbeddingGemma 2 brings multimodal search to phones
A tiny Google model that lets apps search photos, video, audio and text by meaning. Small enough to run on a phone, and free to use. blog.google

Reflection announces Beam, a 501B open-weight model
A startup's first big open model, built for coding and agents. Sign up for early access now; the download comes later this month under Apache 2.0. reflection.ai

Liquid AI opens d1, decision models that answer in one pass
Instead of writing an answer word by word, these models pick one in a single quick step. Built for fast decisions on small devices. Free to download. liquid.ai

Perplexity halves the price of its decision model
Perplexity's model for quick choices, like which tool to use for a question, now costs half as much. Anyone can download it. community.perplexity.ai

Amazon Nova 2.5 Sonic is generally available
Amazon's voice model that you talk to in real time is now better at following instructions and using tools. Developers get it in Amazon Bedrock. aws.amazon.com

Reka previews Rho-1, one model for text, video and actions
One model that can read, watch video and steer a robot, instead of several models passing work between them. An early research preview. reka.ai

Aleph Alpha releases Kolibri, an English-German open model
A German open model that speaks English and German, for European companies that want to run AI on their own servers. Free to download. aleph-alpha.com

Ai2 open-sources AstaBrief for cited research reports
Ai2's free model turns a research question and a pile of papers into a short report with citations, in under a minute on average. allenai.org

Whistle fits speech recognition into 16.9 MB
Speech-to-text so small (17 MB) it runs offline on almost any phone or laptop, no internet needed. Understands seven languages. cactuscompute.com

Microsoft launches its first streaming transcription model
Microsoft's model writes down what people say as they talk, in 60 languages. Microsoft says it's the most accurate on a public leaderboard. microsoft.ai

Cloudflare launches Clef decision models on Workers AI
Cloudflare's models choose from answers you give them instead of writing text, which is handy for sorting and routing. Free to download. developers.cloudflare.com

Decagon's Voice 3 runs on its own speech model, Chord
Decagon's AI phone agent for customer support now speaks with its own voice model. Most people in a blind test couldn't tell it was AI, Decagon says. decagon.ai

Tavus previews Griffin, real-time video for AI agents
An AI that talks with you face to face on video and reacts in under half a second. Only a small group of testers can try it so far. tavus.io

Claude Code Projects opens to every Pro and Max user on the waitlist
Give Claude a goal and it splits the work into tasks that run side by side, then puts the result together. Everyone on the waitlist is now in. @ClaudeDevs

Claude Managed Agents gets dynamic workflows, in public beta
For developers: one Claude agent writes a plan, hands pieces to many other agents, then merges their work. Open to try in beta. @ClaudeDevs

Codex now predicts your next message
Codex can now guess what you'll ask it next and suggest the message for you, based on your chat so far. A beta for Pro users. @OpenAIDevs

Codex on Windows gets a new sandbox built on Microsoft containers
Codex on Windows now runs code inside a sealed-off box built by Microsoft, so it can't reach files or the network it shouldn't. @OpenAIDevs

Qwen-Image-2.1-Turbo makes and edits images in 8 steps
A faster version of Alibaba's free image model: it makes and edits pictures in just 8 steps and still outputs 2K images. @Alibaba_Qwen

Claude Dashboards and Claude Motion are in beta
Ask Claude to turn your data into a live dashboard, or an idea into a short animated explainer. Both are new and in beta. @claudeai

GPT-6 with Intelligent UI rolls out in ChatGPT
ChatGPT now runs on GPT-6 and can answer with interactive pieces, not just text. Paid plans got it first; Free and Go started on Oct 8. @OpenAI

SynthID Detector is open to everyone
A free Google tool that checks whether a picture, video or audio clip was made with AI from Google or partners like OpenAI and Nvidia. @GoogleDeepMind

Cursor lets you run your computer's agents from your phone
Start a coding agent on your computer, then check in, reply or give it new work from the Cursor iPhone app while it keeps going. @cursor_ai

Perplexity releases pplx-embed-v2-late retrieval models
Two free Perplexity models that help search find the right text, images and pages for a question: a big one (9B) and a small one (0.6B). @perplexity_ai

Cloudflare adds Clef-omni, which also hears and watches
A new Clef model that takes in audio and video as well as text and images. Cloudflare also cut Clef-flash's price and made Clef up to 2x faster. blog.cloudflare.com

Anthropic will report on Claude's behavior more often
The first report lists four ways Claude acted on real websites or systems in ways Anthropic didn't intend, like working around a block. All had little real-world impact. anthropic.com

You can now set up a ChatGPT dot from your phone
A dot is an always-on ChatGPT helper that keeps working on a goal between chats. Until today you could only create one on a computer. @ChatGPT

Grok Bot gets its own email address
Grok's assistant can now sign up for services, email businesses for you or book time with someone, using an inbox of its own. @grok

GPT-6.1 Sol gets an Ultrafast mode, up to 8x faster
A faster way to run OpenAI's GPT-6.1 Sol: close to its top model, up to 8x quicker than standard Sol. In the API, Codex and ChatGPT Work. @OpenAIDevs

The two biggest releases this week were about cost. A trillion-parameter model that only wakes 52 billion at a time, and a Haiku at ten cents.