Latest briefing ·

This week in AI

What changed, what it makes possible, and who can use it.

6 selected developments. Start here, then follow a tool or catch up on the month.

Selected by AI-Buzz. Source-review dates appear on each story.

Images 2.5 lets you sketch the change you want.

ChatGPT Images 2.5

OpenAI added sketches and annotated comments to image editing, alongside templates and shareable prompts. You can point to a change or suggest a composition instead of describing every detail in words.

Access & limits. OpenAI describes rollout to all tiers across ChatGPT, ChatGPT Work and Codex on web, desktop and mobile. Flare and Sunburst are separate API options. Account access and output quality have not been tested here.

Why it matters · AI-Buzz’s readingA rough drawing can make a revision easier to explain. That is useful for iterating on an image, even when the first result still needs work.

OpenAI’s Images 2.5 announcement

By AI-Buzz · Published
Sources reviewed · Eastern Time

Speed, precision and image provenance
API choices
OpenAI describes Flare as the faster option and Sunburst as spending more time on precision. This does not establish a fixed generation time.
Provenance
OpenAI says generated images include C2PA credentials and an invisible watermark. AI-Buzz has not tested whether either survives later editing.
This story and OpenAI’s tools

Meta puts an agent inside a conversation.

Muse

Meta says Muse can send email or book travel using its own cloud computer, asking for approval before emails or purchases. You can use Muse or WhatsApp to assign a task, then leave it running.

Access & limits. Meta says US access is rolling out on iOS, Android and the web. Reliability has not been tested by AI-Buzz.

Why it matters · AI-Buzz’s readingMessaging could become a way to delegate computer work. The useful test is whether the agent completes a task with less supervision than doing it yourself.

Meta’s Muse announcement

By AI-Buzz · Published
Sources reviewed · Eastern Time

Permissions, cost and what comes later
Your control
Meta says Muse asks before sending email or making purchases. You choose which services it can access and can disconnect them.
Cost
The announcement describes free use and subscription plans, without giving prices or usage limits.
Later
AI-glasses access and the Confidential VM are future plans. Neither is part of the announced US rollout.
This story and Meta’s tools

Lyria 3.5 gives music prompts more direction.

Lyria 3.5 in Gemini

Google added genre choices, vocal or instrumental selection and templates to music generation in Gemini. It also claims improvements to vocals and arrangements.

Access & limits. Google says the update is available to Gemini users globally on web and mobile, and names Flow Music, AI Studio, Vids and the API. The announcement does not specify quotas or a universal maximum track length.

Why it matters · AI-Buzz’s readingA template and a few musical choices can help turn a vague idea into a draft. The result still needs listening and revision for its intended use.

Google’s Lyria 3.5 announcement

By AI-Buzz · Published
Sources reviewed · Eastern Time

Controls and limits of the announcement
Length
Google describes shorter and longer track choices without establishing one duration limit for every surface.
Rights
Feature availability does not establish that a particular output is cleared for a reader’s intended use.
This story and Google’s tools

Astra takes on tasks across existing applications.

GPT-6 Astra

OpenAI introduced GPT-6 Astra for coding, document work and tasks across browser and desktop applications, including software without an API. Its September 9 follow-up describes access through ChatGPT Work, Codex and the API.

Access & limits. OpenAI’s initial rollout was limited. Its later announcement names ChatGPT Work, Codex and API access; Enterprise administrators must enable it. AI-Buzz has not tested an account or completed tasks with it.

Why it matters · AI-Buzz’s readingDelegation becomes more useful if it works in the software you already use. The important test is whether the task finishes correctly with manageable supervision.

OpenAI’s Astra launch

By AI-Buzz · Published
Sources reviewed · Eastern Time

Rollout and organizational controls
One release
September 3 introduced Astra; September 9 expanded the access description. These are supporting announcements for one development.
Controls
OpenAI describes Enterprise controls for allowed websites, applications and transfers. Model capability alone does not grant permission to use a work system.
Date limit
Both announcements were read on September 10. They do not establish the date any particular account gained access.
This story and OpenAI’s tools

Perplexity Computer adds local models on supported Macs.

Computer · Hybrid Compute on Mac

Perplexity added local models for Mac files and actions, while keeping cloud assistance for research, planning and harder reasoning. Some work can happen on the laptop without making the whole assistant offline.

Access & limits. Pro, Max or Enterprise required, plus an Apple silicon Mac with macOS 15 or later and at least 24 GB of unified memory. These feature requirements are stricter than a general Mac-app download requirement.

Why it matters · AI-Buzz’s readingWhere a task runs becomes a choice with practical hardware and privacy consequences. Local steps do not establish that every later step stays on the device.

Perplexity’s Mac hybrid-compute announcement

By AI-Buzz · Published
Sources reviewed · Eastern Time

What hybrid means here
Cloud help
Perplexity retains cloud assistance as part of the service. Its privacy checks are vendor claims, not independently tested guarantees.
Different hardware
Portable Computer is a separate DGX Spark offering with different platform and plan requirements.
This story and Perplexity’s tools

Runway shows generated worlds you can steer.

GWM Worlds 2

GWM Worlds 2 generates video and sound in response to character, scene and camera controls. Runway’s demos show a scene continuing as the player changes what happens.

Access & limits. Research preview. The announcement offers a contact form; it does not establish public access.

Why it matters · AI-Buzz’s readingGenerated video could become something you interact with. A convincing scene still leaves the harder question of whether a whole game can stay coherent.

Runway’s research and demos

By AI-Buzz · Published
Sources reviewed · Eastern Time

What the demos show—and their limits
Output
Runway reports continuous 720p video at 24 frames per second with generated audio. These are vendor-reported capabilities.
Limits
Runway says examples with actions written in advance produce better quality than its live demo, which maps keys and mouse input to premade prompts. Fast camera turns can degrade visuals, and the model’s memory is imperfect.
This story and Runway’s tools

End of this edition.

A selection of announcements, checked against primary sources. Product access and performance have not been independently tested.

Earlier in the month and the access comparison →

Earlier AI-Buzz research