Meta Launches Muse Wonder, Its First Top-Tier AI Model

Meta has released its first frontier-level AI models, Muse Wonder and Muse Wonder Full. According to TechCrunch and VentureBeat, they launched on 12 October. CNBC and Axios also covered it. Meta says they perform on a par with the best models from OpenAI and Anthropic, with a focus on agent-style work and memory.

What actually happened

The facts that matter, as reported by those outlets:

• There are two models. Muse Wonder is available to developers through Meta's Model API at $4 per million input tokens and $20 per million output tokens. TechCrunch says Meta describes that as roughly the cost of OpenAI's GPT-6 Sol or Anthropic's Opus 5.5.
• Muse Wonder Full, the more capable one, is not available through the API. VentureBeat reports it is only offered through Meta's Muse Code subscriptions, at $50 or $200 a month with usage limits.
• Neither model can be downloaded or run on your own servers, which is a change from Meta's earlier open approach. Meta has not said whether it will open source any Wonder models.
• Meta stresses "active memory": keeping and using information you give it over time. It also says it trained the models to try another approach when stuck rather than stalling and checking back in.
• The benchmark claims come from Meta's own testing. VentureBeat, citing Artificial Analysis, reports Wonder Full tying Claude Sonnet 5.5 on one index, with Opus 5.5 still ahead.
• On data, VentureBeat reports that API use is not used for training, while Muse Code subscriptions default to a mode that lets Meta use submissions to improve its models. You can switch that off, but only for future submissions.

What it means for a small business or recruiter

For most small firms, nothing changes today. You don't need a new tool because a new model exists. What it does give you is more competition on price and capability, and a reminder that the model behind a product matters less than how the product handles your data.

The memory feature is the part to think about. An assistant that remembers what you tell it is useful for repeat work. If you hold CVs, salary details or client information, it is also a place that information could end up, and the training default on subscriptions is exactly the kind of setting people miss.

What to do

1. Don't switch tools because of a launch. Wait for independent testing on tasks like yours.
2. Check the data setting before you paste anything in. If you try a Muse Code subscription, look at the training option first, and keep candidate and client details out until you've read it.
3. Test on a low-risk job. Drafting, summarising or research, with a person checking the result. Meta itself says the models can make mistakes on unfamiliar multi-step tasks.
4. Compare cost per job, not per token. VentureBeat cites Artificial Analysis estimates of about $4.52 per average task for Wonder against $5.98 for Opus 5.5.
5. Keep your workflows tool-agnostic. If your process only works with one vendor, you'll feel every price change.

The caveat

This is days old and mostly vendor-reported. The performance comparisons are Meta's own, and the independent figures cover a single index. Wonder Full has no API price yet. CNBC also notes some users have complained that Meta's Muse agent accessed personal messages without authorisation, which Meta disputes. We haven't tested these models ourselves.

If you're deciding where AI fits in your recruitment or sales process, talk to us.