Meta put its best car in the showroom this week. Then it told you that you can only drive the slightly slower one home.
On 12 October 2026, Meta announced two new models, Muse Wonder and Muse Wonder Full, and called them its most capable to date. They are Meta's first real entry into the frontier tier, the handful of models at the top of the industry that everyone else is measured against.
The one Meta ranks highest is not something you can build software on.
What Meta actually announced
Meta's own announcement leads on two features, both aimed at AI agents rather than chat.
The first is Active Memory, which Meta describes as "the ability to carry what matters to you across conversations and bring back the details that are relevant to the task at hand." In plain terms, the model remembers you between sessions instead of starting cold every time.
The second is Active Focus, "the ability to keep working toward your goal when a task gets hard, try other approaches, recover from errors, and follow through." That is the gap between a model that quits halfway through a long job and one that finishes it.
Availability is the interesting part, and this bit comes straight from Meta. Muse Wonder is live through the Meta Model API and Muse Code.
Muse Wonder Full is available only to Muse Code subscribers.
Worth noting what the announcement does not contain: no benchmark scores, no prices, no context window. Everything in the next two sections is secondary-sourced, and labelled that way.
Meta's agent, Muse, is already a named partner in Shopify's agent checkout rollout, so these models are heading into live commerce.
What the benchmarks say, and who measured them
None of these numbers come from Meta. They come from Artificial Analysis, an independent benchmarking group, via its Intelligence Index v4.3, reported by VentureBeat.
On that index, measured at maximum reasoning effort, Wonder Full scores 56 and Wonder scores 52. For context, Claude Opus 5.5 sits at 58 and GPT-6 Astra at 53.
Meta's API model lands below Astra. Its subscription-only model lands above.
Coding is where Wonder Full looks strongest. It takes 68 on the Coding Agents Index, tying Claude Code running Sonnet 5.5, leads DeepSWE 1.1 at 79.0 and SWE-Marathon at 57.5, and places top three on 6 of the index's 10 benchmarks.
Then there is Harvey LAB-AA, a legal benchmark, where Wonder Full scores 13.1 against Opus 5.5's 4.2.
A 3x lead sounds enormous. Look at the scale.
LAB-AA measures an "all-pass rate", meaning the share of 120 real legal tasks where the model satisfies every single criterion on the grading rubric.
So 13.1 means Wonder Full fully completes roughly 13 jobs in every 100. Opus 5.5 manages about 4.
Both are poor. Meta's is less poor.
The benchmark's own authors say frontier legal work is far from solved, and a 13% finish rate is not a lawyer replacement.
Sources differ on the framing: TechCrunch described Wonder as "comparable" to Astra and Claude Fable 5.1, while the Artificial Analysis figures put it a point below Astra.
Why only one Muse Wonder model is on the API
Back to the showroom.
Muse Wonder, the one you can actually deploy, costs $4 per million input tokens and $20 per million output, with cached input at $0.20 per million. Those prices are secondary-sourced too, and land roughly in line with GPT-6 Sol or Opus 5.5.
Muse Wonder Full has no public API price, because it has no public API. The only way in is a Muse Code subscription: Power Usage at $50 a month, Maximum Usage at $200.
So the model Meta is marketing on and the model you can build on are two different products.
Here is the wider shift.
Meta built its name on open weights, and its first frontier models are closed, with the stronger one behind a consumer subscription rather than a developer API. Only the smaller Muse Glimmer still ships weights.
That is a different Meta from the Llama era.
What this means for you
If you build software: you cannot ship on the good one. Customer-facing apps, internal automation and production agents all run through the API, which gives you Wonder at 52, below Astra.
If you already pay for Muse Code: go and check your data settings today. The next section explains why.
If you are just choosing an assistant: nothing urgent changes. On the independent index Claude Opus 5.5 is still ahead, and our ChatGPT vs Claude comparison still holds. Meta's pitch is memory and follow-through, so it is worth a look if you run long, multi-step jobs.
The catch Meta did not lead with
Muse Code defaults to Contributor mode, a setting that lets Meta use the prompts, code, files and outputs you submit to train future models.
Opting out is manual, and it is per session.
So the only route to Meta's strongest model is a subscription that learns from your work unless you flip a switch every time you sit down.
For a solo developer that is a judgement call. For an agency on a client's codebase under an NDA, it is a contract problem.
Meta's API, by contrast, is reported to exclude your content from training by default.
Almost nobody is leading with this.
What to do about it
If you pay for Muse Code, check Contributor mode in your settings before your next session. That is the one urgent action here.
If you build, test Wonder against Astra on your own workload rather than trusting either index. A one-point gap on a general benchmark says little about your specific job.
Everyone else can wait. As of 13 October 2026, Meta finally has a genuinely competitive frontier model.
It just has not decided to sell it to you yet.
Explore the tools featured in this article
What is Meta Muse Wonder?
Is Muse Wonder Full available through an API?
Does Muse Code train Meta's models on my code?
Is Muse Wonder better than GPT-6 Astra?
