Meta's Muse Wonder remembers across conversations. For estimators, that makes bid-data retention the first question
Meta released Muse Wonder and Muse Wonder Full on October 12, with memory that carries across conversations and an API priced at $4 and $20 per million tokens. Before an estimating shop tries it, the retention terms matter more than Meta's benchmark charts.
Meta released Muse Wonder and Muse Wonder Full on October 12, and the feature that matters to an estimating shop is memory that carries across conversations. A model that remembers is handy for a long bid, and it also means pricing, sub quotes and scope notes may persist somewhere you haven't looked. Ask where before you upload anything.
What did Meta actually release?
Two models, described by Meta as its most capable to date, built on its earlier Muse Spark. Reported details from launch coverage:
| Item | What was reported |
|---|---|
| Models | Muse Wonder and Muse Wonder Full |
| Access | Muse Wonder through Meta's model API and Muse Code; Muse Wonder Full for Muse Code subscribers |
| API price | $4 per million input tokens, $20 per million output tokens |
| New features | "Active Memory" (retains relevant information across conversations) and "Active Focus" (keeps working toward a goal, tries other approaches, recovers from errors) |
| Safety | Meta says the models avoid sharing private information without permission; a fuller safety report is expected later |
What does this mean for estimators?
Estimating is repetitive across projects: the same sub lists, the same scope gaps, the same bid-day clean-up. Memory and goal persistence point at that kind of work, such as a model that remembers how your shop levels HVAC bids or which alternates you always carry. That is the plausible upside.
The risk sits next to it. Bid pricing, sub quotes and owner documents are confidential, and many are covered by NDAs or bid instructions. A remembering model raises concrete questions:
- What is stored, and for how long?
- Is memory per user, per company, or shared?
- Can you view and delete it?
- Is it used for training?
- Does the memory follow a departing estimator?
Launch coverage does not answer these, and Meta says more safety detail is still coming. Treat that as unknown, not safe.
Should a mid-size GC or sub switch because of the benchmarks?
No. The charts circulating come from Meta. One chart reportedly shows Muse Wonder Full ahead of GPT-6 Astra on a legal-agent benchmark, but an independent ranking cited at launch put Anthropic's Opus 5.5 at 58 to Muse Wonder Full's 56. None of these tests measure takeoff accuracy, spec reading or bid leveling.
What should an estimating lead do this week?
- Don't load live bid files into any new model with memory until the retention terms are in writing.
- If you want to test, use a closed, awarded job with pricing stripped or old enough not to matter.
- Run the same scope-gap check you use today and compare against the answer your senior estimator found.
- Price a typical bid in tokens first. A spec book is large, and output priced at $20 per million adds up faster than the input rate suggests.
The takeaway: a cheaper or smarter model matters less than knowing what it keeps. Settle retention first, then compare accuracy on your own past bids.
- What is Meta Muse Wonder?
- Muse Wonder and Muse Wonder Full are two AI models Meta released on October 12, 2026. Meta calls them its most capable so far, and they add a memory feature that retains information across conversations.
- How much does the Muse Wonder API cost?
- Coverage of the launch lists the API at $4 per million input tokens and $20 per million output tokens. Confirm current pricing with Meta before budgeting, since reports are secondary.
- Is it safe to put bid documents into an AI model that remembers across conversations?
- Not until you know what the memory stores, where, and who can see or delete it. Ask the vendor in writing before loading bid pricing, sub quotes or owner documents, and test with non-sensitive files first.
- Should an estimating team switch models because of Meta's benchmark results?
- No. The headline numbers come from Meta's own comparisons, and an independent ranking reported at launch put Anthropic's Opus 5.5 slightly ahead of Muse Wonder Full. Test on your own past bids instead.