Meta’s first frontier model keeps its strongest tier off the API

Meta’s Superintelligence Lab released Muse Wonder and Muse Wonder Full on October 12, its first frontier model. It runs on the Meta Model API at $4 per million input tokens and $20 per million output tokens. Input can be text, image, speech or video. Meta TechCrunch

The stronger of the two is not sold by the token at all. Meta sells the Full variant only through a monthly subscription to Muse Code, its coding agent, and never on the API. Muse Code runs $50 or $200 a month, and the Muse app itself costs $20 or $100. One family. Two payment paths. CNBC

Artificial Analysis scores Muse Wonder at 52 on its Intelligence Index and the Full variant at 56, which ranks it fourth of 699 models. Meta’s own model card claims benchmark scores comparable to Fable 5.1 and GPT-6 Astra, but those are the vendor’s numbers. Artificial Analysis

The API-reachable model is the only one your product can call

For an indie builder, what changed is procurement. If your plan is to route your product to whichever model scores highest on the leaderboards, that plan now has a hole in it, and the hole sits at the top of the family. The one in fourth place is not something your app can invoke at any price. Assume the top tier is always on the token menu, and this week proves you wrong. When capability follows the subscription tier, someone else’s price list sets your ceiling.

Muse Wonder is the variant worth testing, and the claim to check is memory. The company says Active Memory carries what matters across conversations, which would cut down the save-summarize-reinject code that plenty of products hand-roll. For now that is a description, not a measured result. Take a few tasks your product already runs, point both models at them, and see whether your own context plumbing gets shorter.

The rest of today’s news

  • Microsoft Maia 300: third-generation in-house AI accelerator, and from 2027 the primary inference system for Microsoft’s products. Microsoft
  • NVIDIA opens Vera Rubin rack designs: rack and tray designs opened to the Open Compute Project under MGX 1.2, with liquid cooling and 250kW racks. NVIDIA
  • Anthropic warns on IoT self-replication: small open-weight models are close to self-replicating on cheap IoT devices, the Frontier Red Team says. Anthropic
  • Torvalds publishes a Claude Code kernel branch: the page cache patch’s code, tests and cover letter draft came from Claude Code, and small-read throughput tripled. Hardware Busters
  • Ondas and Palantir form Regna AI: a company fine-tuning open-weight models to run inside customer-controlled environments, starting with defense. Ondas
  • HCLTech and Anthropic sign a Global Premier Partnership: HCLTech will stand up a dedicated Anthropic business unit and embed Claude and Claude Code into enterprise products. HCLTech
  • LeCun: AI takeover fears are science fiction: at a Seoul briefing he pointed to the limits of LLM-based safety approaches and pitched world models instead. Digital Today