Meta Muse Wonder Benchmarks: Which Claims Hold Up Against Rivals
Meta says its first frontier models lead on memory and long coding runs. Independent scores put them just behind Claude Opus 5.5, and the memory test is Meta's own.
In this story 9 sections
Meta's Muse Wonder and Muse Wonder Full, released October 12, 2026, look strong on Meta's own tests for memory and long coding projects. Independent scores are more modest. Artificial Analysis puts Muse Wonder Full at about 56 on its index, behind Claude Opus 5.5 at 58, and several of Meta's best numbers cannot be rerun by outsiders.
This guide is for developers and teams deciding whether to try Meta's new models. Emergent Wire covers AI models, capabilities and the industry behind them, and this piece tests Meta's claims against the numbers that have been published. We cover what launched, where the scores are solid, where rivals still win and what to check before you build on it.
What Did Meta Actually Launch?
Meta released two models built for long-running agent work. TechCrunch reported that Meta's Superintelligence Lab shipped Muse Wonder and Muse Wonder Full as the company's first attempt at a frontier model. Meta calls them a major leap toward personal superintelligence.
The two are separate models. According to DataCamp's breakdown of Muse Wonder and Muse Wonder Full, both accept text, image, video, audio and PDF input and return text only, with a 1,048,576-token context window. Setting regular Muse Wonder to maximum reasoning does not turn it into Full.
Access differs. Muse Wonder is available through the Meta Model API in public preview and in the Muse Code product. Muse Wonder Full is limited to Muse Code's Power and Maximum plans, at $50 and $200 a month, and is not served through the API.
How Strong Is Meta's Memory Claim?
Memory is the headline feature, and it is the hardest to check. Meta's Active Memory feature is meant to carry important details across conversations and retrieve them when they become relevant. Research scientist Kavya Kopparapu said the model can remember what matters to you over days, weeks and months.
The evidence is MemoryBench, a test Meta built. Muse Wonder scored 85.6 and Wonder Full 85.5, ahead of Claude Opus 5.5 at 81.5 and GPT-6 Astra at 81.4. Those margins look good, but nobody outside Meta can rerun the test.
The score also says little about recall over months. DataCamp notes that MemoryBench shows nothing about long-term recall, and that apps built on the API must store facts themselves. The API itself does not remember a user from one call to the next.
That leaves a gap between the marketing and what you can build today. If your product depends on durable memory, plan to build your own storage layer, and treat Meta's feature as a convenience rather than a guarantee.
Where Do the Coding Numbers Hold Up?
Meta's coding numbers are strong on long projects. In Meta's table, Muse Wonder Full scores 79.0 on DeepSWE v1.1 and 57.5 on SWE-Marathon, a 20-project test of long software tasks. Claude Opus 5.5 scored 46.3 and GPT-6 Astra 42.5 on SWE-Marathon.
A third-party check is less dramatic. DataCamp cites Abundant AI results of 53.1% for Muse Wonder Full and 52.5% for Muse Wonder on SWE-Marathon, with Claude Sonnet 5.5 at 51.3% and Haiku 5.5 at 50.0%. The lead shrinks from double digits to a couple of points.
The table also shows where Meta loses. On ProgramBench, Opus 5.5 scores 70.5 against Muse Wonder Full's 59.0. On Terminal-Bench 4.0-AA, Full scores 57.1 against 59.6 for Opus 5.5 and 59.1 for GPT-6 Astra.
The lesson is the usual one. A model can lead one benchmark and trail another. Read the row that matches your workload, not the average.
What Do Independent Rankings Say?
Independent indices give a more even picture. On the Artificial Analysis Intelligence Index, DataCamp reports Muse Wonder Full at 55.8, which Meta's table rounds to 56. Muse Wonder scores 52.4, Claude Sonnet 5.5 scores 56.0 and Claude Opus 5.5 leads at 58.
That puts Muse Wonder Full close to the top, not past it. The API model, Muse Wonder, sits about four points lower. Meta's frontier claims rest on the Full model, which most API users cannot reach.
The table also shows mixed results on other tests. Muse Wonder Full trails Opus 5.5 on six of 14 tests in Meta's own table, and GPT-6 Astra beats it on three. Nearly all scores use each model's highest reasoning setting, which raises cost and latency.
| Test | Muse Wonder Full | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|---|
| SWE-Marathon | 57.5 | 46.3 | 42.5 |
| ProgramBench | 59.0 | 70.5 | 52.5 |
| Terminal-Bench 4.0-AA | 57.1 | 59.6 | 59.1 |
| HealthBench Pro | 62.3 | 65.6 | 65.3 |
| AA Intelligence Index | 56 | 58 | 53 |
The table above pulls five rows from Meta's own comparison, as reported by DataCamp. Because Meta ran some of the rival comparisons itself, treat the gaps as indicative rather than final.
How Does the Price Compare?
Muse Wonder's API price is $4 per million input tokens and $20 per million output tokens, with cached input at $0.20. That matches Claude Opus 5.5's rates, according to DataCamp. It is well below GPT-6 Astra at $10 and $50, and above GPT-6.1 Sol at $2 and $10.
DataCamp's worked example uses 3 million input tokens, 2.4 million of them cached, and 0.2 million output tokens. Muse Wonder and Opus 5.5 each cost $6.88, GPT-6 Astra $18.40 and GPT-6.1 Sol $3.44. Prices fall when caching works well, so your own workload matters more than the list rate.
Price trends matter here too. Our look at why AI model prices keep falling explains why a premium model's rate is rarely stable for long. Muse Wonder costs more than its predecessor Muse Spark 1.3, which was $1.25 and $4.25.
What Are the Known Weak Spots?
Meta's model card is candid about limits. DataCamp lists weaknesses in multi-step reasoning in unfamiliar scenarios, spatial and 3D reasoning, and lower performance in some languages. An informal test found Muse Wonder Full produced weaker 3D output than Opus 5.5.
Privacy needs attention too. Muse Code subscriptions start in Contributor mode, where content may be used to train Meta's models. Switching to Standard mode applies only going forward. The pay-as-you-go API runs in Standard mode only.
Safety claims are partial. Meta says that without safeguards it does not rule out the high-risk level for chemical and biological threats, and that with safeguards all risk areas are rated moderate or lower. A fuller safety report has been promised, so wait for it before relying on any safety claim.
Why Does Meta's First Frontier Release Matter?
The launch matters because Meta now competes directly with OpenAI and Anthropic on agent models. TechCrunch notes that this is the first frontier release since Alexandr Wang restructured Meta's AI efforts in August 2025. Since then, Meta had shipped Muse Spark and the Muse agent, but nothing that sat at the top of the benchmark tables.
The models were trained at Meta's Prometheus data center in central Ohio, which came online earlier in 2026. TechCrunch gives no spending figures, so the cost of the training run is unknown. What is clear is that Meta wants assistants that carry out tasks over long periods and remember what users tell them.
Meta also describes an Active Focus feature, meant to keep the model working through blockers and to try other approaches when one fails. The evidence for it is Meta's FocusBench and SWE-Marathon, so the same caution applies. The independent SWE-Marathon result is the more useful signal.
For buyers, competition is good news. A third strong provider gives teams leverage on price and terms, and it gives engineers a real option for failover. Whether Meta keeps pace with rivals' release cadence is the open question.
How Should You Evaluate It Before Building?
Run your own tests. Public benchmarks are a screen, not a decision. Use this short process.
- Pick five to ten real tasks from your product, with known good answers.
- Run them on Muse Wonder, Opus 5.5 and your current model.
- Record cost, latency and pass rate at the same reasoning level.
- Check how each handles a long session with interruptions.
- Review the data terms before sending customer content.
The steps above give you a result tied to your own work. They also catch issues benchmarks hide, such as quota limits. Reporting after launch noted that Muse Code quotas drained unusually fast on day one, so check whether that has settled.
Benchmark skepticism is a good habit in general. Our guide to how AI benchmarks work covers why internal tests, contamination and settings can move a score, and why release timing matters. See also our note on why labs ship faster than ever.
At Emergent Wire, we treat first-week numbers as provisional. Scores shift as third-party testers run their own checks, and a model that looks second-best on day one can look different a month later.
Conclusion
Muse Wonder is a credible first frontier entry from Meta. It is strong on long coding projects in Meta's tests, competitive on independent indices and priced like Claude Opus 5.5, but it trails the leader and its biggest claim, memory, rests on an internal benchmark.
Your next step is to test the models on your own tasks at the same settings as your current model. Check the data terms, wait for independent results on memory, and revisit once Meta publishes its fuller safety report.