This week in AI: Opus 5.5 arrives, OpenAI discloses image leaks

Good morning,

Anthropic and OpenAI released cheaper models this week, while OpenAI disclosed that its research agents had uploaded user images to outside websites. Meta announced plans for Muse on glasses, and Google made AI video creation free to try.

Let's dive in!

New Claude and GPT models lower the bill

Anthropic, OpenAI and SpaceXAI released new models for coding and other tasks you can hand to an agent. If you use their APIs, here's how the prices compare.

These are standard input/output prices per million tokens. Longer requests and faster processing can cost extra:

• Claude Opus 5.5: $4/$20, with cached input at $0.20. Anthropic says typical workloads cost 40% less than Opus 5. It is available across Claude and the major cloud platforms.
• GPT-6 Sol and Luna: $2/$10 and $0.10/$0.50 respectively. Both are in the API, ChatGPT Work and Codex for paid plans; Free and Go get Luna in the desktop app.
• Grok 4.7: starts at $2/$6, unchanged from 4.6, through its API, Grok Build and Cursor. SpaceXAI says the fast variant runs at twice the speed for twice the price.

Opus 5.5 is drawing enthusiastic early feedback for visual design and coding. Testers have shared games, animations and interactive demos, while Anthropic has raised five-hour usage limits on Pro, Max and Team. These aren't controlled comparisons, but it's worth revisiting a project that previously needed too much prompting or hit limits.

Artificial Analysis found mixed GPT results: Sol improved on coding-agent tests while Luna slipped. Both regressed on a professional-work evaluation. Anthropic's 40% saving is its own estimate.

Chinese tech company Xiaomi also released MiMo-V2.6 Pro and Flash, with downloadable weights and text, image, video and audio input. You'll need server-class hardware to run these yourself.

Before switching, rerun a few completed jobs with the cheaper models. Check whether they get the work right and what you spend on retries.

OpenAI's research agents exposed user images

OpenAI acknowledged that research agents uploaded user images to third-party hosting sites. Its statement described 53 cases, with images accessible to anyone who found the unlisted links.

The images came from accounts that allowed training use. OpenAI says they were separated from accounts and privacy-filtered; most have been removed. It told TechCrunch it cannot identify and notify affected users because it cannot reconnect images to accounts. It has not disclosed when they were posted.

Separately, Australia confirmed that an OpenAI research agent accessed public and non-public files in a Medicare statistics portal in June and wrote files to an internal server. Officials say no personal information is believed to have been accessed; investigations continue.

OpenAI's September 25 safety report says training, testing and tool-using runs of its most capable models remain paused after an internal agent bypassed internet restrictions. This is a research-safety restriction, not an announcement that ChatGPT or the API has closed.

These incidents do not establish malicious intent. For your own agents, limit upload destinations, keep sensitive data out of their workspace, and require approval before they share anything externally.

Meta wants Muse in your glasses

At Connect, Meta announced plans to bring Muse to its AI glasses. You could ask it about a product or a school-supply list you're looking at, without describing everything in a chat first.

Meta also announced connections to more shopping and work services, including Notion, GitHub and Box, plus an email address for Muse. These would let you pass work to the agent without copying information between apps.

Muse is due on AI glasses in the coming months. The separately announced Meta VR Glasses are due in spring 2027 for $1,299.99. Meta's roughly 100-gram weight figure excludes the attached computing-and-battery puck.

Readers in the US can try Muse through its app. I'd start there and wait for the glasses features to ship before buying hardware for them.

Quick hits 🗞️

• Microsoft announces Home, Code and Autopilot for Copilot. Home and Code enter its early-access program in the coming weeks. The persistent Autopilot agent expands to private preview at month-end; these features aren't generally available yet. Source

• Free AI video in Google Vids. Anyone with a Google account can now generate 1080p clips with Gemini Omni 1.1 Flash, extend scenes and choose durations. Clips carry invisible SynthID watermarks; paid plans give you more generation capacity. Source

• OpenAI's Muse rival could debut Tuesday. OpenAI is teasing DevDay on September 29. Leaks point to an always-on agent called "o", potentially competing with Muse and Grok Bot. Its name, launch and availability remain unconfirmed. Source

• US and China agree on an AI incident channel. The White House says the countries established an AI dialogue, with the next exchange by November, and agreed to create a channel for incidents. It has not detailed how that channel will work. Source

See you next week!

If this helped you catch up quickly with AI, forward it to someone who would find it useful. They can subscribe here.