Sep 9 · AI Daily

Meta has launched its personal AI assistant Muse, featuring quick responsiveness, browser capability, and app integrations at https://muse.ai. Since it is not yet open to everyone, Testing Catalog demonstrated the web interface: Muse's animated avatar can change its appearance and generate matching animations upon request, supports WhatsApp integration right out of the box, lets users add payment methods to enable agentic shopping, and provides easy access and editing for Soul and Memory files.

Meta officially released its personal agent Muse (internal codename Hatch) on September 9, launching initially in the US. It can autonomously handle tasks like sending emails, selling cars, and booking trips, offering a free basic tier alongside monthly subscription plans of $20 and $100.

Claude Code v2.1.266 fixes a regression introduced in 2.1.265 where setting the CLAUDE_CODE_USE_GATEWAY environment variable on its own forced a Cloud gateway login, causing all configurations using API keys or custom auth headers to fail. This version restores the behavior of ignoring this variable when set independently, without requiring configuration changes.

The DeepSeek Open Platform announced adjustments to Flash series pricing effective September 10, 2026, at 12:00 Beijing time. During off-peak hours, input cache hits drop to 0.02 yuan per million tokens (a 60% reduction), cache misses drop to 1 yuan (a 33.33% reduction), and output drops to 4 yuan ( an 11.11% reduction). Peak hour prices are double the off-peak rates.

A source digging through tvOS 27 beta code discovered hardware traces for AppleTV18,1, estimated to be Apple's new Apple TV 4K. The device will feature an A19 or A19 Pro chip with RAM likely bumped to 8GB or 12GB to meet Apple Intelligence requirements. System code also shows the new Apple TV will support Siri AI and come with an upgraded Siri Remote.

Huawei Intelligent Automotive Solution BU CEO Jin Yuzhi posted that the Kirin 9050 Pro is the first high-performance chip using logic folding technology, debuting in the Huawei Mate XT Ultimate Design starting at 19,999 yuan. The chip arranges logic units in layers within a single die and adds vertical interconnect channels to shorten signal transmission paths and lower latency. Jin Yuzhi stated that from the Kirin 9000 to the Kirin 9050 Pro, it took six years for Kirin to complete its rebirth, marking a new milestone for domestic chips.

1/ We've built a bunch of connectors to make Muse fit seamlessly into your life. Muse runs in a secure virtual machine, and we never use your data for advertising or any purpose other than making Muse work better for you. You can connect Muse to Gmail, Google Calendar, Outlook, Plaid, Opentable, Google Docs, Spotify, Function Health, Withings, Tailscale, Peloton, and more.

Your coolest friend is already on Muse

If you want to try Muse on your phone, click here: https://muse.ai/join

Anthropic confirmed that attackers used infostealer malware to steal Claude login sessions, access user accounts, and drain quotas, such as generating unauthorized Claude Code OAuth tokens.

Sonos released the Sonos 27 audio system update on September 8, rolling it out for free to S2 series products. The new system supports integration with external AI models like ChatGPT and Claude via the early-access Sonos 27mcp, letting users adjust playback and sound controls directly inside the respective AI platforms. A new Portable Surrounds feature allows two portable speakers to pair with a soundbar to act as rear surround channels.

OpenAI announced on September 8 that its unclosed internal model (performing far beyond GPT-6 Astra) solved one of the Millennium Prize Problems—the Navier-Stokes existence and smoothness problem—generating a proof via about 10,000 collaborating AI agents in roughly 88 hours to show that smooth initial fluids can develop singularities in finite time.

OpenAI released ChatGPT Images 2.5 on September 9, reducing image generation latency by up to 50% while adding sketch references, a template library, and in-image comments and annotations, making it the most advanced image generation model currently available. DeepSeek announced price cuts for its Flash series models effective September 10 at 12:00, lowering output pricing to 4 yuan per million tokens.

OpenAI released the ChatGPT Images 2.5 image generation model on September 9, improving detail sharpness, natural lighting, and textures while cutting latency by up to 50% compared to Images 2 from April 2026.

OpenAI used an unreleased model in about 88 hours to provide a solution to the Navier-Stokes existence and smoothness problem (one of the seven Millennium Prize problems), completing 17 hours of Lean formal verification via GPT-6 Astra, sending 4.9 million messages and consuming about 300 billion output tokens in the process.

Mark Zuckerberg revealed that Meta is scaling training for post-Watermelon models using Prometheus, its 1-gigawatt AI compute cluster. Watermelon is the unclosed successor to Avocado/Spark and reportedly Meta's largest model to date, with training compute said to be roughly 10 times that of its predecessor. Sources cite the Sources Podcast and @alexeheath's YouTube channel.

Inception released Mercury 2.5, calling it the strongest diffusion large language model yet, with 40% higher intelligence than Mercury 2, speeds of 1,107 tokens per second on NVIDIA GPUs, a 260K token context window, and pricing of $0.20 per million input tokens and $0.75 per million output tokens, with a 20% launch discount.

The author has released an LLM attention visualization tool built on Transformers.js that lets users hover over generated tokens to inspect dependent historical tokens, with code open-sourced on GitHub. The visualization multiplies attention weights across all layers and heads by value vector magnitudes, aggregating them into a single value to control transparency, and pre-generates prompts to avoid downloading hundreds of megabytes of models. The author also exposed internal values by modifying an onnx file and uploading it to their Hugging Face repository for in-browser generation.

Lossless acceleration of large language models via discrete diffusion

Paper: https://huggingface.co/papers/2609.04010

Sierra open-sourced hyper-τ-bench today, a new long-horizon agent benchmark designed to measure a model's ability not just to act as an agent, but to build agents. The benchmark focuses on the task scenario of "agents building agents," providing evaluation standards for long-horizon autonomous agent development.

OpenAI introduced ChatGPT Images 2.5, stating officially that its image models have generated over 3 billion images across ChatGPT Images and GPT-Image API models.