Meta Muse Wonder: The First Time Meta Labs Actually Reached the Frontier
Quick verdict
Meta Superintelligence Labs landed two models, Muse Wonder and Muse Wonder Full, and for the first time the independent numbers put a Meta model next to Opus 5.5 instead of a tier below it. Artificial Analysis scores Full at 56, with only Opus 5.5 higher at 58. Full debuts at #5 on the Vals Index, above every OpenAI model. The catch is the gating: the strong one is locked to Muse Code subscribers, and the standard Wonder people can actually buy is, by several early accounts, a lot weaker.
What actually shipped
Two models. Muse Wonder is the fast frontier model, sold on the Meta Model API and wired into Muse Code. Muse Wonder Full targets what Meta calls "personal and professional agentic work" and is limited to Muse Code subscribers for now. Meta's own framing is that it got here about 14 months after standing up the lab.
- Wonder pricing: $4 per million input tokens, $0.20 per million cached input, $20 per million output
- 1M context window, with text, image, video and audio in and text out
- No per-token price announced for Wonder Full yet
- Active Memory for recall across days and weeks, and Active Focus for working through blockers mid-task
- Meta claims domain strength in finance and legal, and says it built an internal Personal Superintelligence Index to steer development
On Meta's own benchmarks, Wonder Full scores 57.5% on SWE-Marathon against 46.3% for Opus 5.5 and 42.5% for GPT-6 Astra. On DeepSWE it posts 79.0% against 74.2% for Opus and 74.1% for Astra. A researcher on the team says the model was trained for persistence and long-horizon instruction following, and that reward hacking under heavy optimization pressure took real alignment work to contain.
What the independent evals say
Vendor benchmarks are marketing until someone else runs them. Here the outside numbers hold up better than Meta's recent launches would predict.
| Benchmark | Muse Wonder Full | Context |
|---|---|---|
| Artificial Analysis Index | 56 (Wonder: 52) | Opus 5.5 leads at 58; Muse Spark 1.3 was 48 |
| Terminal-Bench 4.0 | 57% | Up from 33% for the prior Meta model |
| Coding Agents Index | 68 | Full's agentic coding score |
| Vals Index | #5 | Above every OpenAI model |
| ProofBench | 94 / 100 | Up from 58, and with fewer tokens |
The cost line is less flattering. Artificial Analysis measured about 54k output tokens per task and heavy input use, which lands Wonder at $4.52 per task in its harness. That is next to $5.98 for Opus 5.5 but a world away from $0.72 for GPT-6.1 Sol. Wonder is priced like a frontier model and burns tokens like one.
Perplexity's Aravind Srinivas said Wonder compares favorably to Fable 5.1 on their WANDR evals. A security team running RSI evals reported zero refusals, with standard Wonder oddly outperforming Full on cyber tasks.
Why it matters
For most of the past two years Meta's open and hosted models sat a clear step behind the OpenAI and Anthropic frontier. A second-place finish on an independent aggregate, within two points of Opus 5.5, is a different conversation. If you route work across models, Muse Wonder is now a real candidate for the hard coding and agentic jobs rather than a cheaper fallback.
The pricing changes the math in two directions. The headline per-token rate undercuts Opus, but the token appetite eats most of that gap on long agentic runs, so the per-task cost is the number to watch, not the sticker. For anyone comparing spend across providers, that is the usual trap: a low per-token rate on a model that writes 54k tokens a task is not actually cheap.
The gating fight
The loudest complaint is not about quality, it is about access. Theo ranked Wonder Full in the same S tier as Opus 5.5, then said standard Wonder is "significantly worse" and called out the subscriber-only wall on the good one. That is the version of Wonder most people can pay for, and the one that scores 52 rather than 56. Meta also shipped a launch bug that drained Muse Code quota too fast; Alexandr Wang said they fixed it and reset usage.
Not everyone is sold. One reviewer with disclosed preview access put Meta roughly 3 to 6 months behind OpenAI and Anthropic based on SVG and 3D output, and Wang acknowledged the team did not train on 3D demo games. Whether "3dmaxxing" is a real weakness or just a benchmark nobody optimized for is still an open argument.
Video: Meta Muse Wonder explained
A walkthrough of what Meta shipped, the benchmark claims, and how Wonder stacks up against the OpenAI and Anthropic frontier.
FAQ
Can I use Muse Wonder Full right now?
Only through a Muse Code subscription. Standard Wonder is on the Meta Model API and in Muse Code, and Meta says the Muse personal agent will get Wonder later, starting with subscribers. If you want to weigh it against other frontier options first, see our best AI models for coding in 2026 rundown.
Is Muse Wonder better than Opus 5.5?
On Artificial Analysis, no: Opus 5.5 leads at 58 to Full's 56. On Meta's own SWE-Marathon and DeepSWE runs, Wonder Full comes out ahead. The honest answer is that they are close enough that your workload decides, which is why running both through one interface beats committing to one. Our Opus 5.5 launch breakdown has the other side of that comparison.
Why does Wonder cost more per task if the per-token price is low?
It writes a lot of tokens. Artificial Analysis clocked about 54k output tokens per task, so the $4/M input and $20/M output rates add up to roughly $4.52 per task, far above GPT-6.1 Sol at $0.72. Per-token rates hide this; per-task cost does not.
Sources
- @finkd - introducing Muse Wonder
- @AIatMeta - availability and the two models
- @alexandr_wang - Active Memory, Active Focus, finance and legal
- Artificial Analysis - independent index, pricing, and per-task cost
- SWE-Marathon results summary
- @Yuchenj_UW - DeepSWE scores
- @ValsAI - Full debuts at #5 on the Vals Index
- @theo - standard Wonder is weaker and the gating is the problem
- @scaling01 - Meta 3 to 6 months behind on SVG and 3D
- @alexandr_wang - Muse Code quota bug fixed and usage reset
Further reading
Try all the models mentioned in this article
Admix gives you GPT-5, Claude, Gemini, and 350+ AI models in one app. Compare responses side by side. Try it for 7 days for $1.99.
Try Admix for $1.99