Meta released Muse Wonder and Muse Wonder Full on 12 October, describing them as its most capable models to date. Fourteen months after Mark Zuckerberg rebuilt the lab under Alexandr Wang, the result is two closed models. Wonder is on the Meta Model API at $4/$20 per million tokens, Meta's own list price, and a match for Claude Opus 5.5. Wonder Full is only available inside Muse Code subscriptions. There are no open weights, a change of direction for the company that gave the industry Llama.
For most of our readers, a fourth lab reaching the frontier is not news in itself. You already have Copilot, Claude or Gemini in the stack, and the leaderboard changes hands every few weeks. The key point is that these memory features could help project managers and quantity surveyors retain information across long projects. But before testing Muse Code on live projects, firms should check how it handles their data by default.
We covered Muse for Small Business in ‘Curated Links’ on 5 October and Muse haggling on Marketplace as ‘One More Thing’ the day after. This is the model underneath both.
What Meta says it built
Meta's research post leads with two ideas. 'Active Memory' is carrying what matters to you across conversations and retrieving the relevant detail when a task needs it. 'Active Focus' is working through blockers, trying alternative approaches and checking in before taking important actions.
Kavya Kopparapu, a research scientist at Meta, put the memory ambition plainly in the launch video, as reported by TechCrunch.
We put a tremendous amount of effort into making sure your model can remember what's important to you over days, weeks, months, and beyond.
For anyone running an eighteen-month programme, that is the feature you actually want. It is also the one that raises every data governance question, because memory that persists across conversations is data that persists somewhere.
Where it sits on the independent chart
Meta's own table shows Wonder Full leading on SWE-Marathon, a long-running coding benchmark, and on Harvey's legal agent benchmark. The card admits some rows mix harnesses and vendor-reported figures: Meta's chart, not a controlled test. The independent figure is Artificial Analysis, which scores Wonder Full at 56 on its Intelligence Index, fourth of 693 models. VentureBeat, citing the same index, has Claude Opus 5.5 at 58, GPT-6 Astra at 53 and the standard Wonder at 52.
Carl Franzen, writing for VentureBeat, flagged the number a finance director will care about more than any index score. Artificial Analysis estimates the cost of its average index task at $4.52 for Wonder, against $5.98 for Opus 5.5 and $3.26 for GPT-6 Astra, because the models burn tokens at different rates.
Lower token prices also do not necessarily translate into lower costs per completed task.
The forest, and the rate limits
Developer Dishant Sharma documented a launch-day run of Wonder on PokeBench, a leaderboard of agents playing Pokemon Red. By his account the model reached Viridian City by turn 59, on pace with the fastest agents, then entered Viridian Forest and spent the next 792 turns there, costing $96.06 without beating the first gym leader. He reports Opus 5.5 finished in 271 turns for $8.44.
His conclusion was that "scores are a starting point, not a review". A model that tops a long-horizon benchmark can still spend six hours going nowhere on a task nobody benchmarked.
Access was also uneven. Stork.AI reports that reviewer Matthew Berman upgraded through the $50 and $200 Muse Code tiers and hit HTTP 429 quota errors within minutes of each, and that Wang acknowledged the problem as known. Not a verdict on the model, but a reason not to pilot anything the week it ships.
Before anyone pastes a tender into Muse Code
Meta's own documentation states that Muse Code subscriptions default to 'Contributor' mode, in which you get more usage in exchange for letting Meta use your content, defined as the prompts, code, files and context you send plus the responses, to train future models.
Standard mode does not train on your content, and the pay-as-you-go Wonder API is Standard only. VentureBeat adds that switching Contributor mode off applies only to future submissions, and that Meta's terms prohibit submitting confidential, sensitive or personal information while it is on. So the default on the only route to Wonder Full is one Meta's own terms say you should not use for client data.
Meta's model card also states that, without mitigations, it does not rule out the model reaching its high-risk threshold in the chemical and biological domain, and that it deploys with safety post-training and classifiers in place. That is a candid disclosure, not a scandal, but a client's procurement team may ask you about it.
Check which mode the subscription is in before the first prompt, and record who turned Contributor mode off.
Decide whether the Standard-mode API, rather than Muse Code, is the only acceptable route for live project work.
Ask Meta about zero data retention, which VentureBeat reports is available to qualified organisations but disables the server-side memory the model is sold on.
Takeaway
Muse Wonder deserves attention. A year ago, few would have expected the company to compete at this level. It is priced against Opus 5.5 and focuses on memory and continuity, both useful for long-running projects. But that does not mean firms should switch. Independent tests still place it behind Opus, and the actual cost of completing a task may be higher than the token price suggests.
The more capable model also comes with data terms that Meta itself says are unsuitable for confidential work. If you want to try it, use the API in Standard mode for a limited task and check the actual cost before expanding its use.
Treat Meta's leaderboard as Meta's chart, and use Artificial Analysis for the comparison.
Assume Contributor mode is on until someone in your firm has proven otherwise.
Judge agents on tasks that run for hours, because that is where the forest is.
Project Flux reads every frontier release so you only have to read the ones that change your week. The newsletter is free at projectflux.ai.
Links and Stuff
All content reflects our personal views and is not intended as professional advice or to represent any organisation.
