On 2026-10-12 Meta published "Muse Wonder" on its research blog. It is a new model family with two members, Muse Wonder and Muse Wonder Full, and Meta is pitching both at the work most of us do every day: writing code and running agents. The Hacker News thread reached 570 points and 312 comments, and most of those comments asked the same thing: is this a real option for coding, or another chart?
This post covers what Meta shipped, how you can reach it, what the first independent numbers and hands-on tests say, and what nobody knows yet.
What happened
Meta's post makes three claims, in its own words. The pair "deliver frontier coding performance", "supporting projects from small applications to complex software". Muse Wonder is "A frontier model that leads on personal and professional agents". And the whole thing is the result of a reset: "Fourteen months ago, we began a ground-up overhaul of our artificial intelligence efforts."
The access is the part that matters for anyone who builds things:
- Muse Wonder is "available today through the Meta Model API and Muse Code".
- Muse Wonder Full is "available today for Muse Code subscribers".
- In the consumer app, Muse Wonder "will soon power experiences in Muse, and will roll out first to subscribers."
So the smaller model has an API, and the bigger one, for now, lives inside Meta's own coding product. That split shapes everything below.
How it works, as far as Meta says
The post does not describe an architecture. What it describes is behaviour. Meta calls its agent behaviour Active Focus: "trying alternative approaches, working through blockers, and adapting to new information". It also says that "Working directly in a browser and across apps, Muse Wonder builds a running plan".
That is the same direction every lab is moving in. The model is not only asked for an answer, it is asked to keep a plan, act in a browser, notice when something failed and try again. The demo Meta chose fits that story: "Muse Wonder builds a playable 3D kitchen game from customizable image inputs."
The post names its benchmarks: Harvey's Legal Agent Benchmark (LAB), FocusBench and SWE-Marathon. It shows their scores only in a chart, not in the text, which makes them hard to quote and harder to check. A full report with more safety evaluations is promised later.
What the first independent numbers say
The one outside source with numbers so far is Artificial Analysis, and its page is about Muse Wonder Full:
- A score of 56 on the Artificial Analysis Intelligence Index.
- A rank of #4 of 689 models there.
- A 1M tokens context window.
- It used 140M output tokens to run the index, against a median of 88M.
The last number is the one I would read twice. A model that spends far more output tokens than the median to reach its score is thinking out loud a lot. On a per-token bill, that is a cost you pay on every call, not just on the benchmark. We cannot price it yet, though, because the same page lists no public API price for Wonder Full.
Why someone building web apps with AI should care
Most new models arrive as a chat app first. This one arrives as an API and a coding product. That puts it straight into the market that frontend developers already pay for: the assistant in the editor, the agent that opens a pull request, the model behind a feature you are shipping.
What matters to us is not the index but how it handles our kind of work, and the first hands-on reports in the HN thread were exactly that:
- One commenter, jjcm, ran an image-to-HTML test (rebuild an LCARS-style UI from a picture). Their verdict was "much worse than Opus 5.5", placing it between Sonnet 5.5 and Haiku 5.5.
- Another, ekojs, used it for React UI polish and some Node changes. They called it "a step below" Opus 5.5 and Fable, and said it got stuck in a loop of reviewing and fixing its own work.
- On the other side, jiscariot uses it as a daily personal agent and said it "exceeded expectations out of the gate".
These are three people, not a study. But the pattern is worth noticing: the praise is for agent work, and the complaints are about visual frontend work. That matches Meta's own pitch, which leads with agents.
If your product is a UI, the honest reading today is that the index rank and the day-one UI tests disagree, and the UI tests are closer to your job.
What is not known yet
A lot, and it is worth being exact about it.
- Price. Artificial Analysis lists no publicly available API price or speed. A per-token price was quoted in the HN thread with no source, so I am not repeating it.
- Which model the headline scores belong to. HN commenters say Meta's benchmarks are for Wonder Full, the one limited to Muse Code subscribers. The blog's text does not say.
- Weights. The post says nothing about weights or a licence. One HN commenter pointed out that open weights promised at the earlier Muse Spark launch have not appeared.
- Independent evaluations. Beyond Artificial Analysis, there are none yet. As Nathan Lambert put it on Bluesky: "Will be interesting to see independent evaluations / user testing."
- Routers. It was not on OpenRouter at the time of the HN thread, so you cannot yet swap it into an existing setup with one line.
My take
This section is my opinion, Amar's, not part of the news.
I have not used Muse Wonder yet, so I will not tell you it is good or bad. What I can say is how I would decide. I build React and Next.js interfaces with AI every day, and the tests that matter to me look like jjcm's and ekojs's: give it a screenshot and ask for the component, give it a messy form and ask for the polish. A rank of #4 on a general index tells me the model is serious. It does not tell me it will get my spacing, my states and my accessibility right.
The part I find more interesting is the split in access. The model with an API is the smaller one, and the bigger one sits inside Meta's own product. That is a choice to win developers through Muse Code first, the way other labs grew their coding tools. Until Wonder Full has a public price and a place on the routers, I would treat it as something to try on a few real tickets, not something to rebuild a workflow around. If it is still "a step below" on UI work after the report comes out, the agent strengths will not be enough for a frontend team.
