Meta's Muse Wonder Is Good at the Boring Parts. That's the Point.

By Lyle Heartman · October 13, 2026

Where to put it to work this week

If you have a Muse Code seat, give Wonder Full the jobs where being steady matters more than being brilliant. That's the short version. Meta shipped it on October 12, and Theo Browne spent four or five days pushing it before launch. His verdict, roughly: it's a little less smart than Opus 5.5, and a lot less likely to do something stupid.

That trade points at a specific set of jobs.

  • Bug triage before you write a line. Point it at a bug report and ask what's wrong. In Theo's test on his T3 Code repo, Opus and GPT-6 Sol both wrote fixes. Wonder Full wrote none. It explained the cause in plain sentences, then pointed out that someone else had already opened a smaller, cleaner PR with a one-line fix. Neither of the other two noticed it.
  • The conductor in a multi-agent setup. It writes tight briefs for sub-agents: read the actual code before you cite it, don't edit anything, report back. It handled steering well and kept a lot of parallel workers straight. Theo says labs outside OpenAI and Anthropic usually struggle with exactly this.
  • Second reviewer on PRs. On eight open T3 Code PRs, Opus flagged 23 issues, Sol flagged 9, and Wonder Full landed in between at 16, with five that nobody else caught. Fewer nitpicks than Opus, more coverage than Sol. Pair it with a deeper model rather than using it alone, since it caught fewer high-severity bugs.
  • Volume work under a stronger lead. On Theo's toy compiler project, Opus (running the show) gave Wonder Full the fixers, standard-library work, fuzzing and reviews. Opus kept the parts where a wrong design is expensive. Opus graded Wonder Full a B+ and itself an A-.
  • Repo audits when you're short on time. One codebase audit took Wonder Full 16 minutes. That was about half of Sol's time and well under half of Opus's. Token use was lower than Opus too, at about 54K tokens per task.
  • Long-running assistants that need to remember things. Memory was Meta's headline feature. They trained the model on how memory works during post-training instead of bolting it on.

Keep it away from front-end work and anything 3D. More on that below.

Two models, one good name, and a pricing gap

Meta shipped two models, and the one you want isn't the one you can call by API. Muse Wonder (muse-wonder-1.0) is on the Meta Model API. Muse Wonder Full only lives inside Muse Code, Meta's coding tool, on the paid plans. Neither has downloadable weights. (VentureBeat)

They're clearly different sizes. Theo clocked Wonder at roughly 120 tokens per second and Full at roughly 50. On Artificial Analysis's Intelligence Index, Full scores 56 and Wonder scores 52. That puts Full level with Sonnet 5.5, ahead of GPT-6 Astra and Fable 5.1 (both 53), and behind Opus 5.5 (58).

ModelWhere you get itInput / 1MOutput / 1MCached / 1MAA Index
Muse Wonder FullMuse Code, $50 or $200/monthNo API priceNo API price—56
Muse WonderMeta Model API, Muse Code$4.00$20.00$0.2052
Claude Opus 5.5Anthropic API$4.00$20.00$0.2058
GPT-6 AstraOpenAI API$10.00$50.00$1.0053

One correction to the early chatter: Wonder isn't cheaper than Opus on the rate card. It's the same price. Where it can come out ahead is per task, because it burns fewer tokens. Artificial Analysis puts an average index task at $4.52 for Wonder and $5.98 for Opus. Web search through the API adds $2.50 per 1,000 queries.

The $50 Power plan gets you Full with 20 times the Everyday plan's usage. The $200 Maximum plan gets 80 times. Nobody outside Meta has said what that means in actual tokens, Theo included.

Where it falls over

Front-end work is the big one. Theo didn't call the UIs broken. He called them generic, the kind of interface Sonnet 3.7 used to turn out. If your agent ships screens, use something else.

3D is weak too, and Meta says so in its own limitations section. In Blender the models came out rough, with poor coloring and lighting. The game it built still moved and played fine, which is more than Theo could say for Grok 4.7. Theo's broader point is worth keeping in mind: 3D skill comes from training environments built for it, not from general smarts. A model can be bright and still bad at it.

The last one is the trap. The API model is not the model people are raving about. On Theo's T3 Code architecture audit, plain Wonder scored roughly half of what Opus did, a drop of more than 30% from Full. On his Lakebed audit it landed at 5.75 out of 10, where GPT-6 Sol scored 9. A lot of the bad takes online come from people wiring the API model into their usual tools and assuming they're testing the flagship. They aren't. At Opus prices, Theo doesn't think plain Wonder makes much sense.

It's also quieter on the scary bugs. In the PR review test it caught fewer high-severity issues than either Opus or Sol.

Quirks other people have hit

Wonder is a day old, so most of the field reports are about Muse Spark, the smaller line. Spark 1.3 shipped about six weeks before Wonder, and Wonder runs on the same API and harness. A lot of the plumbing quirks probably carry over. Check before you build on any of them.

From the Spark line

These come from eesel's Spark 1.1 review, DataCamp's Spark 1.3 tutorial and eesel's Spark 1.3 review.

  • tool_choice only accepts "auto". You can't force a specific tool, and "required" or "none" returns a 400.
  • Structured output is off by default. With strict: false, tool arguments aren't guaranteed to match your schema. Validate before you execute.
  • On Chat Completions, Meta redacts the reasoning field before you see it. Each turn thinks from scratch, and Meta's own coding-agent guide warns it can repeat or contradict earlier steps. Use the Responses API with previous_response_id for multi-turn agent work.
  • Pointing Claude Code at it takes three changes: a base URL with no /v1, every model alias repointed, and ANTHROPIC_AUTH_TOKEN instead of ANTHROPIC_API_KEY. Miss an alias and part of your run quietly falls back to Claude.
  • Spark 1.3 is wordy. It used about 120 million output tokens to finish the Artificial Analysis index, against a field median of 72 million.
  • Muse Code can update itself in the middle of a session. Fine for daily use, annoying if you're benchmarking.

From Wonder's first day

  • Theo found it token-efficient in his harness. Artificial Analysis measured the API model differently: 130 million output tokens to run its index against an 85 million median, and 19.27 seconds to first token against a 5.52-second median. In an agent loop that wait repeats on every call. (Gabriel Anhaia on DEV)
  • Theo leaned on his own AGENTS.md (keep it short, vary sentence length, use plain technical English). Without it, he says, the model gets noticeably more verbose. Bring your style rules.
  • Usage limits are tight. Matthew Berman's review ran into strict quotas and uneven visual quality (4sysops). Another reviewer climbed from the $15 plan to $50 and then $200 just to get enough time on Full (Stork).
  • The memory claim is mostly Meta's own. MemoryBench is internal and judged by an LLM. In Kingy AI's independent recall test, Wonder and a GPT model both got all 1,152 checked values right, so no visible edge there (Kingy AI).
  • Muse Code defaults to Contributor mode. That means Meta can train on your prompts, code and outputs unless you switch it off, and switching only covers future submissions. The pay-as-you-go API is Standard mode only. Check this before you point Wonder Full at a client repo. (VentureBeat)

The cheap tier is still there, but not for Wonder

The discount people remember is real and still live, and it's much bigger than 50%. It just doesn't cover Wonder. Meta calls it Contributor mode. You pay a fraction of the normal rate, and in return Meta can train future models on your prompts, code, files and outputs. (Meta pricing page, checked October 13, 2026)

ModelModeInput / 1MOutput / 1MCached / 1M
Muse Spark 1.3 or 1.2Contributor$0.10$0.20$0.002
Muse Spark 1.3, 1.2 or 1.1Standard$1.25$4.25$0.15
Muse Wonder 1.0Standard only$4.00$20.00$0.20

That works out to about 92% off input and 95% off output on Spark. You call it with muse-spark-1.3-contributor or muse-spark-1.2-contributor. Spark 1.1 has no Contributor version. Meta's page says outright that Contributor mode isn't available for Wonder on the API.

There are strings attached. Contributor endpoints are capped at 150 requests per minute, against 3,000 on Standard. Meta's terms also bar confidential, sensitive or personal data while you're in it. So it's a prototyping and evals tier, not a production tier for anything you'd mind Meta reading.

One wrinkle for Wonder Full: Muse Code subscriptions run in Contributor mode by default, and Meta frames that as discounted pricing or more usage. If you see someone getting a lot out of a $50 plan, that may be part of why.

How I'd slot it in

Treat Wonder Full as a second opinion with good manners, not a replacement for your main model. Run your hardest design work through the strongest model you trust. Let Wonder Full do the triage, the reviews and the bulk fixes around it. Have each one check the other.

If you only have API access, run your own bake-off before paying Opus prices for plain Wonder. One real ticket, both models, same afternoon. And if cost is the whole point, Spark 1.3 on Contributor mode is still about the cheapest capable thing around, as long as Meta reading your prompts is fine.

Theo admits his own bias here. He reviews a lot of models, and something that simply behaves differently feels good after months of sameness. That's fair. But a model that reads the bug, finds the existing fix and tells you to merge it is useful whether or not it's novel.

Sources