The best coding model isn’t necessarily the one that writes the cleverest function. It’s the one that survives three hours inside your repository without turning a small migration into an architectural hostage situation.
That’s the interesting bet behind Muse Wonder, not the “first frontier model” headline. Meta released Wonder and Wonder Full on October 12, 2026, moving beyond Muse Spark’s lightweight, low-latency pitch toward sustained engineering work.
And there’s a catch before you even write an API call: the strongest version isn’t available through the API.
Two models, one awkward shopping decision
Both Wonder models have a one-million-token context window and accept text, images, video, and audio, with text output. But access splits sharply:
- Wonder: Meta Model API and Muse Code; $4 per million input tokens, $0.20 for cached input, and $20 for output.
- Wonder Full: Muse Code only, through eligible $50/month Power Usage or $200/month Maximum Usage subscriptions.
Those are different products, not interchangeable leaderboard labels. If you’re building a GitHub agent with Meta’s OpenAI-SDK-compatible client, you get Wonder. You don’t get to borrow Full’s scores for your pitch deck.
Also, “Meta’s first frontier model” is historical amnesia. Llama had already been regarded as frontier-class. Wonder is the frontier-focused arrival from the reorganized Muse program, following Spark’s April debut and successive developer releases.
The leaderboard has a specific shape
Meta’s [published scorecard](https://dev.meta.ai/models/muse-wonder/) gives Wonder Full 57.5 on SWE-Marathon, versus 46.3 for Claude Opus 5.5 and 42.5 for GPT-6 Astra. On DeepSWE 1.1, Full reaches 79.0; regular Wonder is close behind at 78.8.
That’s encouraging for repository work: multi-file changes, testing, revisions, and recovering when the first approach fails.
But Opus leads the Artificial Analysis Intelligence Index, 58 to Full’s 56, and wins ProgramBench, 70.5 to 59.0. Meta hasn’t swept the board. Its strongest showing is narrower—and more useful—than a universal intelligence victory lap.
<> The compelling promise isn’t “better code in one shot.” It’s “less babysitting after the fifth failed test.”/>
These comparisons also have limits. Meta describes competitor runs as best-effort, and SWE-Marathon uses different agent products rather than a common harness. The scores are a reason to run your own evaluation, not replace it.
The Elephant in the Room
Meta built developer goodwill around downloadable Llama weights. Wonder is a hosted service. The separately released open-weight Muse Glimmer doesn’t change that.
So this launch is also a business-model shift: paid API usage, premium coding subscriptions, and persistent personal context inside Meta’s ecosystem.
The memory pitch deserves particular scrutiny. Active Memory retains relevant context across conversations; Active Focus is intended to keep agents working through blockers and asking for confirmation before important actions.
Wonder scores 85.6 on MemoryBench. But Meta’s internal test uses conversations of 64,000–100,000 tokens, including conflicting and updated facts. That measures contextual recall, not months of dependable personal memory.
Remembering your preferred test runner is convenient. Remembering private documents and calendar details creates obligations around consent, deletion, and access controls. “It remembers me” stops sounding cute when nobody can explain what it forgets.
WIRED’s Will Knight also flags incomplete agent-safety disclosure. Meta says a fuller safety report is forthcoming. Shipping powerful action-taking software before complete public documentation deserves scrutiny, not applause for moving fast.
Give it a branch, not the production keys
I’d test Wonder on an ugly, bounded repository task: a migration with flaky tests and several interacting files. Measure completed-task quality, token spend, recovery, and unauthorized actions—not how polished its progress updates sound.
Keep execution sandboxed and require approval for irreversible changes, as Meta itself recommends.
Muse Wonder looks like a serious engineering contender. But the prize isn’t a frontier badge. It’s a pull request you can merge without spending the afternoon discovering what the agent quietly broke.
