AI news: week of November 5–12, 2026

Anthropic and Google released new top models, Claude gained spreadsheets and data dashboards, and several labs reported AI tools and results in biology.

Anthropic and Google release their most capable models

Two new top models arrived in one week: Claude Fable 5.2 takes on long, loosely described projects largely on its own, and Gemini 4 Carbon comes close to it at a much lower cost.

Claude Fable 5.2 ranks first in independent tests and did professional-level work on real freelance jobs about twice as often as Anthropic's everyday model Opus 5.5, which Anthropic still recommends for most daily work. Testers found expert-level work in many fields, but in Anthropic's own report biology and chemistry experts rated its research ideas no better than their own, and it made up answers more often than Gemini 4 Carbon when it did not know.
Gemini 4 Carbon, for Google AI Ultra subscribers and paying developers, came close to Fable 5.2 at a fraction of the cost, and in one independent test it was the least likely of the leading models to make up an answer.

DeepMind, Planet, Anthropic and others report AI tools and results in biology

New AI tools predict the effects of DNA changes and search satellite images, and AI systems did or sped up parts of several lab and field studies.

AlphaGenome 2: Google DeepMind's new model predicts how small DNA differences change gene activity in specific cell types, where the first version averaged over whole tissues; it is free for non-commercial research.
Satellite images: Planet's new map-based app answers plain-language questions about ten years of its satellite images, such as how a place has changed, and is free for now.
Vault particles: Researchers used Claude to help determine the first 3D structure of a vault particle from bacteria, cutting a three-day purification to about 90 minutes, though the model's mistakes also cost time.
Enzyme system: Anthropic reports early, not yet peer-reviewed results in which Claude found a new enzyme system linked to repeated DNA, resembling CRISPR (the bacterial defense system used in gene editing).
AutoDNA: In Nature, an AI-run lab carried out more than 11,000 steps of DNA work, with people only maintaining the hardware.
Elephants: Reuters reports that AI found the third of elephants that human counters had missed.
Dolphins: The Wild Dolphin Project uses AI to find patterns in decades of dolphin recordings, but has not shown a two-way conversation.
Lab tests: In a new test of reading scientific images such as protein gels, Gemini 4 Carbon did best but models often answered wrongly, and in OpenAI's new test on raw wet-lab data its top model handled about two-thirds of the tasks.

Anthropic adds spreadsheets and data dashboards to Claude

Claude now builds working spreadsheets with formulas and live data dashboards inside the chat.

Claude Sheets, in testing, builds a spreadsheet with formulas from a request or from an imported Google Sheets, Excel or CSV file; several people can work in it at once, and it exports to Excel.
Claude Data, now available to all, pulls in data from databases and other business systems as live feeds and turns figures into dashboards and visual data stories.

Google launches Gemini Decisions, a model for quick yes/no and scoring judgments

Google now offers a model that answers yes/no, multiple-choice or scoring questions about text, images and other media in bulk, with a confidence for each answer that is meant to match how often it is right.

Gemini Decisions, offered to developers in Google AI Studio, takes many questions about one input in a single request and returns scores instead of written answers.

People build a Norwegian speaking tutor, a family archive and other tools with AI

Seven worked examples show people building their own apps, archives and daily checks by describing them to AI.

Mural: A founder who had never built a phone app used OpenAI's Codex to build an open-source app for practicing spoken Norwegian.
Household records: A math teacher keeps his family's appliance details, warranties and to-do lists as plain text files that Claude Code reads and updates.
Letters archive: A reader built a searchable website of a relative's handwritten letters from 1942–43 entirely through prompts.
Running data: A reader built an app that analyzes a runner's Garmin watch data and has Claude explain the trends in plain English.
Care records: A parent used Claude to write an app in which caregivers check off her disabled son's daily tasks and which writes the required daily summary.
Reading history: A writer had Codex gather his Kindle and Spotify history, had Claude chart it, and turned his book history into a timeline.
Error alerts: A staff member who does not work in engineering described in plain English a daily Claude task that checks her company's error chart and alerts the team only on real problems.

AI systems rebuild software from the finished program, without its source code

AI systems can now study a finished program, or just watch how it behaves, and write new software that does the same.

MirrorCode: In a test where AI agents must recreate existing programs without seeing their source code, AI went from passing a small share of the tests on one program to nearly all of them within nine months.
REMEND: A small open model recovers the mathematical equations inside compiled programs (the runnable form of software).
OccupySaaS: Developers use AI to build free, open-source copies of paid online services such as Auth0 and Salesforce without copying their code.
Limits: Anthropic steers Claude Code away from recreating specific existing games and now offers to build an original game in the same genre instead; users also report refused reverse-engineering tasks.

Read more: MirrorCode paper · REMEND · OccupySaaS · Claude's game rule

Six labs release new models, four of them open

Several labs released open models whose files anyone can download and run, among them a mid-sized model that leads US open models at coding.

Ripple-65B Preview (Ricursive Intelligence), a mid-sized open model for coding and agent tasks, beats much larger US open models on a coding-agent test, the company says, and runs efficiently on local devices.
Kimi K3.1 Pro (Moonshot AI) is a very large open model that reads text and images and handles agent tasks such as coding.
Hy4 (Tencent) reads charts and documents and can operate computer programs; Tencent says the model files will follow.
Inkling 2 Small (Thinking Machines Lab) handles code and images, and its files became downloadable this week.
Muse Ember, Meta's smallest and cheapest model, matched Anthropic's Claude Haiku in an independent ranking at a lower cost per task.
Qwen4-Max-Preview (Alibaba) understands text, images and video, and is in preview through Alibaba's paid plan.

Read more: Ripple · Kimi K3.1 Pro · Hy4 · Inkling 2 Small · Muse Ember · Qwen4-Max-Preview

Tips

By hand three times first. A newsletter column argues for doing a task by hand at least three times before handing it to an AI agent, to learn what a good result looks like and where things go wrong. Source

Questions and sketches before building. One column has the assistant ask multiple-choice questions or sketch the plan in the chat before it builds anything, which saves work that would be thrown away. Google's free Elephant-Goldfish plugin has fresh agents without the background challenge a software plan before coding starts. Column · Elephant-Goldfish

Editing instead of arguing. A guide to Claude points out that the assistant rereads the whole chat before every answer, so a bad answer left in it keeps affecting later ones; editing the original prompt removes it. Source

Rating an answer shares the whole chat. Anthropic's privacy page says that pressing thumbs up or down under a Claude answer stores the entire conversation with Anthropic. Anthropic

Code for every number. A guide to checking bank exports has Claude use code for every number instead of estimating totals by reading. A guide to Claude in Excel adds a test row to see that totals update, since a wrong total shows no error. Bank exports · Excel

A short memory. Anthropic's research found that saved memory helps an assistant on long tasks, but its top layer has to stay small. A guide gives a prompt that has Claude list its stored memories, mark each to keep, fix or delete, and show a short cleaned version. Anthropic · Guide

The agent checks its own work. Anthropic's Claude Code team has Claude run the tests and build to check its own work before any person looks; Claude Code's creator calls this probably the most important step for good results. Anthropic

Same tasks, several assistants. A guide gives the free plans of six assistants the same tasks, including a factual question and a small spreadsheet, and scores each on accuracy, sources, file handling and usefulness. It finds free plans generous for chat but tight on uploads and research. Guide

In brief

e-means replaces k-means. Meta released e-means, a faster replacement for the k-means method of sorting data into groups, in its free Faiss library. GitHub

How scientists use Claude. An Anthropic study of research sessions finds that scientists mainly use Claude to analyze data and to write. Anthropic

New speech-to-text models. ElevenLabs and Microsoft released models that turn recorded speech into text with fewer errors. ElevenLabs · Microsoft

Coding agents still need cleanup. In a test on real-world tasks, the best coding agent fully completed only about a quarter of them. Paper

Chatbots led hikers into trouble. High Country News reports hikers who needed rescue after planning trips with AI chatbots. High Country News

AI agents run a radio telescope. The SETI Institute and Breakthrough Listen now steer the Allen Telescope Array and process its data with AI agents from a single plain-language request. NVIDIA

AI news: what is new in AI, once a week. Archive · Subscribe · Unsubscribe