Meta Muse Wonder Launched Two Models Under One Name

Meta Muse Wonder went live on October 12, 2026. Two models, one label. That’s six words of context, and they change how you should read every score attached to this launch.

The benchmark results in circulation don’t describe one consistent product. They describe Muse Wonder and Muse Wonder Full. Separate models, different numbers across coding, command-line work, legal tasks, medical questions, and memory. Which variant produced which number isn’t always visible from a headline. That’s the trap.

Meta positions both as frontier models built for personal and professional agentic tasks.

Direct quotes from the launch: “our most capable models” and “a major leap toward personal superintelligence” (Meta, eesel). I’m not knocking the launch. Worth watching. Just don’t treat “Muse Wonder” as a single benchmarked thing until you know which model generated the result.

Meta’s model variant naming makes that harder than it should be.

Two models behind one product name

Vendors publish benchmarks constantly.

Normal. What isn’t normal is two separate models sharing one product name, which is exactly the arrangement with Muse Wonder and Muse Wonder Full (eesel).

Here’s the failure mode.

A headline says Muse Wonder leads a benchmark.

The model you actually pull for evaluation might be the other variant. Name stays familiar. Capability profile shifts underneath it.

Meta’s launch materials claim Muse Wonder Full tops coding benchmark leaderboards.

DeepSWE and SWE-Marathon among them.

In independent third-party runs (Meta). One reported comparison gives Muse Wonder Full a 79.0 on DeepSWE 1.1, against 74.2 for Claude Opus 5.5 (Compute Leap). Another puts plain Muse Wonder at 78.8 on the same benchmark, Claude Opus 5.5 still at 74.2 (Dev.to).

Both strong on that reported comparison. Not interchangeable.

Meta Muse Wonder benchmarks: a split story

The strongest case for these frontier agentic models isn’t dominance. It’s uneven strength by task type. Full table of the reported numbers, one row per benchmark:

| Benchmark | Muse Wonder | Muse Wonder Full | Claude Opus 5.5 |
|—|—|—|—|
| DeepSWE 1.1 | 78.8 | 79.0 | 74.2 |
| SWE-Marathon | 53.1 | 57.5 | 46.3 |
| ProgramBench | not reported | 59.0 | 70.5 |
| Terminal-Bench 4.0-AA | 52.0 | 57.1 | 59.6 |
| HealthBench Pro | 60.5 | 62.3 | 65.6 |
| MemoryBench | 85.6 | 85.5 | 81.5 |
| Harvey LAB-AA | 10.0 | 13.1 | 4.2 |

Three footnotes, because agentic AI benchmarks love asterisks. Muse Wonder Full’s 57.5 on SWE-Marathon came at maximum reasoning effort (eesel, Compute Leap). The reported ProgramBench comparison covers Muse Wonder Full only, so Muse Wonder’s column stays empty (Dev.to). And Muse Wonder Full’s Harvey figure is 13.1 on Artificial Analysis’ version of Harvey’s Legal Agent Benchmark, where passing means meeting every criterion and clearing hallucination checks (SiliconANGLE). Close enough to sit in the row, other enough to flag.

Now the split story in prose.

SWE-Marathon favors both Muse variants over Claude Opus 5.5’s 46.3.

ProgramBench goes the other way, hard: 70.5 for Claude Opus 5.5 versus 59.0 for Muse Wonder Full.

Terminal-Bench 4.0-AA also goes to Claude Opus 5.5 at 59.6, Muse Wonder Full at 57.1, Muse Wonder at 52.0 (DataNorth).

HealthBench Pro too: 65.6 for Claude Opus 5.5, 62.3 for Full, 60.5 for Muse Wonder.

Legal-agent work flips it back. Muse Wonder scored 10.0 on Harvey LAB-AA against 4.2 for Claude Opus 5.5 (Dev.to). Ten versus 4.2. Nobody rounds their way into a gap like that.

Then memory, where the two-model problem gets practical. Muse Wonder scored 85.6 on MemoryBench, Muse Wonder Full scored 85.5, and Claude Opus 5.5 scored 81.5 (eesel). The API-accessible model beating the flagship by 0.1 tells you more than another “most capable” claim ever could. Kinda the whole argument in one decimal. Benchmark variant confusion costs real decisions when the wrong model wins by a tenth of a point and nobody notices.

There’s a measurement problem too.

Meta’s model card reportedly states the SWE-Marathon comparison isn’t controlled, since some competing models ran inside their own coding tools (DataNorth).

Doesn’t make the result useless. Makes it narrower. An uncontrolled comparison can tell you a model performed well in that reported setup. It can’t tell you how the model behaves inside your own automation, repository, prompt flow, or tool configuration. For anyone building AI automation, that’s the line between a useful lead and an expensive assumption.

What Meta Muse Wonder results mean for small operators

Small teams and solo builders don’t need to settle the frontier-model debate. You need to know whether the exact Muse variant available to you handles your actual workload better than what you’re already running.

Start small. Build an evaluation set from real tasks. The coding, memory, command-line work, document handling, and agent steps your automation actually performs. Hold the prompt, tools, and inputs fixed. Then compare outputs from exact model variants, not brand names.

Track more than whether an answer looks impressive. Record whether the task completed, whether it followed the required steps, whether it introduced errors, whether you had to repair the output. A model that wins a leaderboard but creates cleanup isn’t winning your workflow. Benchmark names multiply faster than anyone can audit them, which is its own quiet problem.

Separate three questions while you’re at it, as they don’t collapse into one.

Whether Muse Wonder performs well on your task.

Whether Muse Wonder Full performs better on it.

And whether you can access the specific variant and setup that produced whichever result convinced you.

The facts here answer the first question only in selected benchmark settings. They show Muse Wonder and Muse Wonder Full hold other results, and that neither dominates every listed comparison.

My hot take: the shared name is a bigger operational risk than the benchmark dispute itself. Builders are trained to question scores. Product names slide through uninspected. When two models share one label, the first line of your evaluation notes should name the exact variant.

Muse Wonder FAQ

Does Meta Muse Wonder beat Claude Opus 5.5?

Not across the board.

Reported comparisons put Muse Wonder and Muse Wonder Full ahead on DeepSWE 1.1, SWE-Marathon, MemoryBench, and Harvey LAB-AA. Claude Opus 5.5 leads on ProgramBench, Terminal-Bench 4.0-AA. And HealthBench Pro. “Depends on the task” is the only honest answer.

Is Muse Wonder the same as Muse Wonder Full?

No. Two separate models under one product name.

Full tops the coding leaderboards in Meta’s launch materials; the regular variant still scored 78.8 on the reported DeepSWE 1.1 comparison and edges Full on MemoryBench, 85.6 to 85.5.

Are the SWE-Marathon numbers a fair comparison?

Partially. Meta’s model card reportedly says the comparison isn’t controlled, since some competing models ran inside their own coding tools. Treat the score as evidence about that setup, not a verdict.

Which Muse variant should you actually test?

Whichever one you can call, inside your own model evaluation setup.

The result that matters comes from your tasks, your prompts, your tools.

Not from whatever number is attached to the shared name.

Verdict: Muse Wonder is a family

Treat it that way.

Credible launch, real strengths on selected coding, legal-agent, and memory benchmarks.

Not one model with one stable score. And the available results don’t support a blanket claim that it beats Claude Opus 5.5 at everything.

Read the model name.

Inspect the evaluation setup. Test the exact variant against your own work before you change a production workflow.

The launch hit October 12, 2026, and same-day coverage was already pulling the comparison setups apart (SiliconANGLE). That pace won’t slow down. If you build small and ship often, that discipline saves you more time than any leaderboard headline. Use the benchmark results as a reason to run your own comparison, not as permission to skip it.