Three things happened in AI this week worth paying attention to. Meta announced its next generation of agent models. xAI folded three separate subscription products into one plan with a new price structure. And Higgsfield, which makes AI video tools, used GPT-6 Astra to push its videos past the one-minute mark.
Meta Releases Muse Wonder and Muse Wonder Full for Agentic Tasks
On October 12, Meta announced Muse Wonder and Muse Wonder Full, calling them its most capable models to date. Both are built for what Meta describes as personal and professional agentic tasks: work where, in Meta’s words, a model “has to interpret the information in front of it and work through the problem.” The examples from Meta’s newsroom include working through an insurance claim, making sense of a lease, and interpreting someone’s finances. Explicitly named domains are law, finance, and health, along with coding.
Two features define the new models beyond raw capability. Active Memory carries what matters to a user across conversations and brings back the details relevant to the task at hand. Active Focus is the model’s ability to keep working toward a goal when a task turns difficult: trying different approaches, recovering from errors, and asking for clarification when needed, and checking in before following through on anything important.
Muse Wonder is available today through the Meta Model API and Muse Code. Muse Wonder Full is available to Muse Code subscribers. Meta says Muse Wonder will soon power experiences in Muse, its personal AI agent introduced on September 8, rolling out to subscribers first. Meta says the models build on the groundwork laid with Muse Spark, introduced on April 8, 2026.
On safety, Meta says the models are trained to avoid sharing private information without permission and to turn down requests to assist with harmful acts. A full safety evaluation report is listed as coming soon.
The UX question I keep coming back to: Active Memory is the kind of feature that sounds good in an announcement and reveals its design choices only when you use it. What the model retains, what it forgets, and what it surfaces in a given session is a design decision with real consequences for people using it in professional contexts. That detail is not in the announcement. It will be in the product.
Find out more: Introducing Muse Wonder and Muse Wonder Full (Meta Newsroom)
xAI Combines Grok, Cursor, and X Premium Into One Subscription
Also on October 12, xAI announced The X Subscription, which brings Grok, Grok Bot, Cursor, and X Premium under a single plan. The announcement describes it as one subscription covering “the full path from thinking and creating to building and sharing.” Globally available over the next few days.
The pricing table has five tiers. Free gives daily Grok access, a starter Cursor allowance, and standard X, with no monthly charge. Plus is $8 a month and adds more AI usage in both Grok and Cursor, a checkmark, creator tools, and fewer ads. Premium is $30 a month and replaces what would have been SuperGrok, Cursor Pro, and X Premium+ purchased separately, a combination xAI says previously cost $90 a month combined, with four times the Plus usage. Super is $100 a month at 15 times the Plus usage. Ultra is $200 at 50 times.
Paid tiers share a single usage pool for Grok models that works across Grok apps, the Grok tab on X, and Cursor. Each plan includes at least twice the coding usage with xAI’s own models of a comparable Cursor plan, according to xAI. Third-party models remain available in Cursor on Premium and above, but those are billed separately at API list prices through top-ups; they do not come from the shared pool.
Teams get Premium, Super, and Ultra at the same per-seat prices as individual plans. xAI says this is lower per seat than the old Cursor Teams pricing and now includes three times more coding usage, plus SSO, consolidated billing, usage analytics, and SOC 2 compliance.
Existing SuperGrok, Cursor, and X Premium subscribers can stay on their current plans. Switching is optional and comes with a prorated refund on the old plan.
The tier worth looking at carefully is Premium. Three products that previously cost $90 a month combined, now at $30, with what xAI says is more usage per dollar than the plans it replaces. If your team is already paying for Cursor and Grok separately, that math is worth running before the next billing cycle. The third-party model billing is the variable to watch. If you rely on third-party models inside Cursor for anything significant, “billed separately at API list prices” (Source) can add up fast and is worth modeling against what you currently pay.
Find out more: Introducing The X Subscription (SpaceXAI News)
Higgsfield Uses GPT-6 Astra to Break the One-Minute Wall for AI Video
On October 13, OpenAI published a customer story on how Higgsfield AI used GPT-6 Astra to extend AI-generated video from roughly one minute to over ten minutes.
Before GPT-6 Astra, Higgsfield videos topped out at around a minute. Higgsfield’s creative agents handle repetitive production tasks, such as creating multiple variations of a shot. With GPT-6 Astra, those agents can now call external tools to handle complex work in 3D modeling. According to CEO and co-founder Alex Mashrabov, this helps address one of creative AI’s hardest problems: maintaining consistency across longer videos.
Higgsfield has already applied the capability to its own YouTube channel, which Mashrabov says has close to half a million subscribers. The company now uses an AI workflow to generate instructional videos over ten minutes in length. That is content that previously required substantial manual work.
Higgsfield also uses GPT-Image-2.5, which the story ties to promotional images for small businesses. Mashrabov described it in the OpenAI story as “the first model of its kind that can reliably render the text and infographics on top of the imagery.” That precise text rendering matters to small business content, where readable text and precise infographics can be just as important as the image itself.
Looking further out, Higgsfield wants to build voice-to-video with GPT-Live-1. The concept is that a creator describes an idea aloud and the AI renders it in real time. Mashrabov calls this vision “voice to movie.” It is a stated goal, not a shipped feature.
Ten minutes of coherent AI video is not the same product as ten-second clips stitched together. Short-clip tools have been adequate for a specific kind of creator work. Long-form instructional video, tutorial content, branded explainers: all of these have required human editors to hold the piece together across its runtime. If Higgsfield’s consistency holds across external use cases and not just its own internally built workflows, that changes what a small team can realistically produce. The fact that Higgsfield is using its own tool to build its own content library is the most useful signal in this announcement. That is a real test. The open question is whether the workflow transfers.
Find out more: Higgsfield takes AI video from 1 minute to over 10 (OpenAI)