Meta's Muse Wonder launches with split access, unconfirmed pricing and regional waitlists

Meta's first frontier-class model, Muse Wonder, arrived on October 12, 2026 — and the days since have been defined less by benchmarks than by access confusion. According to a widely shared setup guide on dev.to, the launch splits the model family across three channels, ships without an official price list, and leaves the consumer app rollout undated. A second dev.to post, written from Spain, captures the rougher edge of that rollout: a user completing a full sign-up flow only to be told the product is unavailable in their region.

What actually launched

The dev.to guide describes Muse Wonder as the first true frontier LLM from Meta's Superintelligence Labs, the product of training runs at a gigawatt-scale data center Meta calls Prometheus. Two variants shipped: a standard model for everyday agentic and coding work, and Muse Wonder Full, a maximum-reasoning version aimed at the hardest, longest tasks. Both carry two capabilities Meta highlights: Active Memory, which holds context about a user across conversations so agents working over days don't lose the thread, and Active Focus, which keeps an agent pushing through blockers and pausing before consequential actions. Meta claims state-of-the-art results on agent, memory and coding evaluations — figures the guide rightly flags as vendor-reported until independent runs land.

Access is split three ways

The single biggest source of confusion, per the guide: the two variants are not distributed through the same channels. The standard model is available now on the Meta Model API as muse-wonder-1.0. The endpoint is OpenAI-compatible — existing code mostly needs a new base URL and model name — and it also accepts the Anthropic Messages format. The model is listed on OpenRouter as meta/muse-wonder. There is no free tier: a payment card and active billing are required, or calls fail with authentication errors.

Wonder Full, the variant behind the headline benchmark scores, is not served through the API at all. The only announced route is a Muse Code subscription, with tiers reported at $50 and $200 per month. The third channel, the Muse app, has no arrival date: Meta says the model will reach the app's experiences soon, with paying subscribers first in line, on plans reported at $20 or $100 monthly.

Pricing without a price list

Meta has not published an official pricing page, according to the guide. The circulating figures — roughly $4 per million input tokens, $20 per million output tokens and $0.20 per million cached input tokens — are reported across outlets rather than confirmed by the company. The guide also passes along a cautionary data point: one user reported that a single Wonder Full run consumed 38% of a weekly Muse Code allowance. Per-token prices look modest, but long-running agents on reasoning models burn through them fast.

Where developers are getting stuck

The guide catalogs the launch's most common failure points. API and Muse Code availability is region-dependent, enforced at the account level — a VPN won't reliably restore billed API access. The reasoning_effort parameter accepts values from minimal through max, but not none; sending none returns a 400 error. On the Chat Completions endpoint, reasoning doesn't persist between turns, which makes multi-turn agents inconsistent — the fix is the Responses API with previous_response_id to carry reasoning context. And the absence of Wonder Full from the API is by design, not a bug: no hidden endpoint exists.

The waitlist nobody expected

A separate dev.to post, dated October 15 and written by a user in Spain, documents the consumer-facing side. After reaching muse.ai and seeing Meta branding, the author verified an email address, entered a first name, surname and date of birth — and only then hit a screen saying Muse isn't available in their country or region, with an option to join a waitlist. The author explicitly avoids calling it a scam and notes they can't determine what data was retained, but argues the sequence is the problem: if availability is region-restricted, that restriction should surface before personal details are collected, not after.

Benchmarks, for now, are Meta's word

On launch-reported numbers, Wonder Full scores 56 on the Artificial Analysis v4.3 index — second globally, behind Claude Opus 5.5 at 58 — while leading long-horizon coding evals such as DeepSWE 1.1 (79% versus 74.2%) and SWE-Marathon (57.5% versus 46.3%, both figures Meta-reported). The standard model scores 52 on the index. The guide also notes that Meta's safety-evaluation report was not published at launch.

Why it matters

A frontier launch is also a distribution decision, and this one fragments it. The variant with the best scores is subscription-only, the API model's pricing is unconfirmed, and the consumer app is undated and region-gated — so what Meta's frontier model means in practice depends entirely on which door you walk through. For teams, that translates into three cautions: treat the benchmarks as vendor claims until independent runs appear, wait for an official price list before budgeting agent workloads, and check the supported-regions list before building around the API. The Spain account adds a softer but real cost: a sign-up flow that collects personal data before revealing a regional block erodes trust at exactly the moment Meta is asking users to hand agents long-running, consequential work.