Lead story / launch

Meta launches its newest assistant model with paid-only access

Meta launched Muse Wonder on October 12, calling it its most advanced model for digital assistants. Access in the Muse app requires a $20 or $100 monthly subscription, even though the app itself remains free. The release pairs a promise of better task completion with what Meta describes as its strongest cyber-misuse safeguards yet.

Meta AI chief Alexandr Wang said the model was developed to improve personal and professional agents. His examples were familiar errands: booking dinner reservations and checking availability. The training goal, he said, was to get an agent to try another approach when it hits a roadblock, rather than stall and ask the user for help.

Developers have two routes. Muse Wonder Full is limited to Muse Code's $50 or $200 monthly plans, while Muse Wonder is also available through the paid Meta Model API. Meta has not disclosed API pricing. Wang said it would cost more than standard subscriptions because the model is much larger, but be priced “meaningfully below” leading OpenAI and Anthropic models.

Meta says its strongest protections against cyber misuse are currently restricted to the paid model. The paid rollout is therefore a test of both AI monetization and access to those protections. It also gives Meta a potential source of AI revenue beyond its core advertising business.

enterprise adoption

Red Bull says its AI now recommends actions during Formula 1 races

Red Bull says its strategy agent, developed with Oracle, now recommends race-day actions instead of just finding rules. At the 2025 Italian Grand Prix, it advised giving back Max Verstappen's position after he cut a corner, to avoid a potential penalty. The team followed the advice, and Verstappen later overtook his rival. Executive Jack Harington says human engineers and strategists remain involved as decisions weigh track events, off-track developments and the championship.

culture

Andrej Karpathy says casual chatbot use misses stronger AI capabilities

Andrej Karpathy argues that brief chatbot exchanges reveal little about AI used for demanding, long-running work. He says capability is strongest in digital tasks with verifiable results, especially coding and mathematics. His rough estimates put professional users of leading models in those fields at about 20 million and people with internal frontier-system access at about 5,000. These figures are his “napkin math,” and his descriptions of internal capabilities are not independently verified.

benchmark

Artificial Analysis launches a 3D scene benchmark with four models tied at the top

Artificial Analysis's AA-3D compares 21 models turning written briefs into explorable 3D scenes, with each run capped at three hours or 250 agent turns. Four models from Anthropic and OpenAI are statistically tied at the top. Among the top five, GPT-6 Sol had the lowest mean cost at $3.46 per scene and was quickest across the six tasks shown. Rankings reflect a panel's scene preferences, not a universal measure of quality.

Also worth knowing

•  Anthropic says attackers extracted 90–99% less of Claude's reasoning

•  Reddit seeks to bar its former law firm from defending Perplexity

•  X adds Cursor to paid plans, but third-party AI models cost extra

•  Meta unveils liquid-cooled network switches for larger AI clusters

•  Army sets out a plan to use AI to spot biological threats earlier

•  Blueprints AI says software that updates building plans together launches this week

•  Meta defends explicit sexual chats in its adults-only Muse app

Daily tool drop

6 AI tools worth knowing today

The Internet Had a Point

Meme from the stashFrom “Show Me the Code” to “Show Me the Prompt”

Reader check-in

Help shape tomorrow's briefing

One click tells us what to keep, improve, or tighten.