Muse Wonder has two doors, and only one is public
Meta's October 12 launch splits one model family into an API tier and a subscriber-only tier. The benchmarks are self-reported and the safety report is still pending.
Facts checked October 13, 2026
On October 12, 2026 Meta introduced Muse Wonder and Muse Wonder Full. Per Meta, Wonder is available through the Meta Model API and Muse Code; Wonder Full only to Muse Code subscribers. Meta says a full safety report is coming soon. Meta post
What Meta says it released
Meta's research blog describes the Muse Wonder series as its most capable models for personal and professional agents. It names four themes: domain expertise, Active Memory, Active Focus and frontier coding. Meta says the models were trained at its Prometheus data center and that Muse Wonder will later power experiences in the Muse assistant, first for subscribers.
Meta research post, October 12, 2026Two models, two doors
The access split is the most concrete fact in the launch. Meta states that Muse Wonder is available today through the Meta Model API and Muse Code, while Muse Wonder Full is available to Muse Code subscribers. TechCrunch and VentureBeat both report that Wonder Full is not served through the API and that neither model ships with downloadable weights. VentureBeat contrasts this with the earlier Llama models, which organizations could run themselves.
VentureBeat reportPrices and benchmark numbers: who said what
Both outlets report API pricing for Wonder of $4 per million input tokens and $20 per million output tokens; VentureBeat adds $0.20 per million for cached input. Those are press-reported figures, so check Meta's pricing page before budgeting.
Meta's own benchmark claims come from its model card. VentureBeat reports that Artificial Analysis scored Wonder Full 56 and Wonder 52 on its Intelligence Index, with Claude Opus 5.5 at 58. The third-party index is more independent than a vendor chart, but it is a single index from launch day.
TechCrunch reportAnalysis: my inference
My reading: Meta sells the cheaper model to developers and keeps the strongest one inside its own product. Engineers can only test what they can reach. If you can only reach the weaker model by API, any claim about Wonder Full has to be taken from Meta or from subscribers, not from your own harness.
I read the lack of open weights as a change in how Meta competes. That comes from the press reports; Meta's post does not discuss licensing.
What is not established
Meta says a full safety report is coming soon. Until it is published, I treat safety claims as pending. The agent-memory features also raise a question the post does not answer in the text I read: what is stored, for how long, and who can inspect it. I found no independent replication of the benchmark results yet, and I did not run either model.
A small test I would run before adopting it
Use your own tasks, not Meta's. Keep a fixed set of ten real jobs, record the model and date, run each twice, and score by a rule written before the run. Add one task where the right answer is to ask a question. Meta describes Active Focus as asking for clarification and checking in before important actions, so that behavior is testable.
Illustrative evaluation ledger, not a benchmark result
task model/date run1 run2 rule-score asked-when-needed
T01..T10 ________ __ __ __ yes/no
rule: written before the run; never edited after
route used: API (Wonder) | Muse Code (Wonder Full) | not reachable