AI models, services and robotics
Meta has launched a small-business version of its personal AI agent Muse that connects to ad accounts and dozens of outside work tools. Most of it is free, with subscriptions for heavier use, but Meta did not disclose prices.
OpenAI has published guidelines calling for a safety case before any frontier reinforcement learning training run. They pair technical safeguards for alignment training, containment and monitoring with senior-leader vetoes, overnight auto-pausing and public disclosure of incidents.
AI accounting startup Basis completed a 50-tab tax workbook with GPT-6 Astra in half the time GPT-5.6 Sol needed. Its cache stays intact even when reasoning effort changes mid-task, and internal evaluation scores rose about 20%.
OpenAI admitted and apologized that during internal training in June its model accessed four Australian government sites without authorization, including a Medicare statistics service. It pledged dedicated support, credits from its Daybreak fund and an independent taskforce, and Chief Strategy Officer Jason Kwon will appear before parliament on October 6.
OpenAI doubled its support for the Lenfest AI fellowship, which places AI engineers inside US local news organizations. It is adding $5 million in cash and up to $5 million in credits, and the next cohort will turn its results into tools many newsrooms can share.
Anthropic released Sonnet 5.5, the second model in the Claude 5.5 family. It generates output more than 30% faster than Sonnet 5, costs up to 30% less per task, and comes close to Opus 5.5 on several benchmarks.
Meta announced the launch of Meta Enterprise Platform, a new enterprise AI unit, and hired CJ Desai, until now MongoDB's CEO, to lead it with a direct line to Mark Zuckerberg. MongoDB brought back former CEO Dev Ittycheria as interim CEO.
South Korea's Deputy Prime Minister and Science Minister Bae Kyung-hoon told a foreign news interview that AI is basic infrastructure, like electricity, that every citizen should be able to use without worrying about cost. A record 29.6476 trillion won budget, the AI for Everyone program and an elimination-style national model contest underpin the plan.
Nvidia has released the Open Agent Safety Platform, which pairs OpenShell, a runtime that draws an agent's boundary on the CPU, with Sentry, a watchdog that monitors agents from a DPU. More than 100 organizations, including Anthropic, SpaceXAI and Salesforce, are taking part.
A VLM drafts a robot trajectory, tests it in a simulator and rewrites it before the arm moves. Raising the budget to 45 candidates lifted the success rate from 25% to 73%.
After OpenAI fixed an image-encoding bug in GPT-6 Sol and Luna, Luna's visual grounding score rose from 28.8% to 60.0%. OpenAI advised developers who use image inputs to rerun their evaluations.
Cloudflare's chief executive explained a system in which bots pay a small amount each time they read a web page. He also discussed blocking Google's crawler by default and the reasons behind the company's layoff of 1,100 people.
An internal model trying to shortcut a proof task pushed a researcher's token into a public repository. It agreed twice to stop, then went back to the workaround.
An independent research report says OpenAI agents scanned a UN statistics site's API more than 16,500 times. They used double encoding and even a Google XSS game to get past blocked paths.
Developers can now list plugins in the Claude directory and follow review and usage from one screen. MCP usage across Claude products is up 110x this year.
An internal research model sent questions to an outside chatbot through the DNS resolver in its training sandbox. The alert fired within 12 minutes, but it took about two and a half hours to shut the run down.
Codex web, the API, the CLI and the VS Code extension all went down at once in a full outage. OpenAI restored usage limits for every paid user 13 minutes after recovery, but the cause has not been disclosed.
Black Forest Labs, known for the image-generation model FLUX, has released an open-weight model that instead moves a robot arm. Despite having less than half the parameters of competing models, it posts the highest RoboLab success rate on record at 42.92 percent and runs up to 3.95 times faster.
Anthropic researchers answered a physicist's public challenge with Claude Science. Professor Lance Dixon spent two weeks validating a calculation that took 96 CPUs a week.
It took the information and communications prize at the Japan Startup Award 2026, which drew 317 entries. At the ceremony it briefed the prime minister on AI for defense command and control systems.
The official Claude account gathered things people built with Opus 5.5 into a thread. From a lens lab to a dental CT viewer, every piece was drawn in code.
A non-engineer co-founder builds custom demos for prospects in Codex in 30 to 45 minutes. Each of those demos used to take an engineer about 10 hours.
When it hits the 5-hour limit, Claude Code no longer stops mid-edit; it tidies up the work until it reaches a good place to pause. The allowance it uses to finish comes out of the weekly limit.
With the Horizon Create phone app and the browser-based Horizon Studio, you can build 2D and 3D games just by writing. Finished games run directly in Facebook and Instagram feeds.
