Meta has officially entered the million-token price war. On October 12, 2026, the company announced Muse Wonder and its bigger sibling Muse Wonder Full, calling them its most capable models to date — and priced the standard version at $4 per million input tokens with a 1M-token context window. If you publish, code, or build workflows on top of frontier models, this launch changes the math you've been running all year.
What Meta actually shipped
Unlike the consumer-facing Muse releases we've covered — the personal agent inside WhatsApp, or Muse Spark's day-to-day assistant upgrades — Muse Wonder is aimed squarely at professional work. According to the official announcement, it delivers state-of-the-art performance on personal and professional agentic tasks, with expert-domain capabilities spanning law, finance, and health, plus what Meta describes as frontier-level coding.
Two capabilities carry the generation:
- Active Memory: the model carries what matters to you across conversations and recalls relevant details, instead of resetting context every session.
- Active Focus: when a task gets hard, the model keeps working toward the goal — trying alternative approaches and recovering from errors rather than stalling.
There are two distinct tiers of access. muse-wonder-1.0 is available today through the Meta Model API and the Muse Code editor. muse-wonder-full-1.0, the stronger variant, is subscription-only inside Muse Code's higher plans. Meta also says the new model will "soon power more of Muse, rolling out first to subscribers."
The launch shipped with an official researcher video from Meta introducing Muse Wonder, published alongside the announcement — the best starting point before digging into the spec sheet.
The official pricing, in plain numbers
| Model | Context | Input / Mtok | Cached input | Output / Mtok | Access |
|---|---|---|---|---|---|
| muse-wonder-1.0 | 1,000,000 tokens | $4.00 | $0.20 | $20.00 | Model API + Muse Code |
| muse-wonder-full-1.0 | 1,000,000 tokens | subscription | subscription | subscription | Muse Code subscription only |
To reach Muse Wonder Full you need a Muse Code plan: Power Usage at $50/month (includes the most powerful Muse models and 20x the everyday plan's usage) or Maximum Usage at $200/month (80x usage for the heaviest coding workloads). Both tiers also carry early feature access and higher file upload limits. The standard model is additionally listed on OpenRouter, useful if you route multiple providers through one account.
The performance claims, and how to read them
Meta says Muse Wonder Full topped DeepSWE and SWE-Marathon in independent third-party runs, scoring 81.4 on SWE-Marathon versus 79.0 for Claude Opus 5.5 at max settings, alongside 59.1 on the Personal Superintelligence Index. Treat manufacturer benchmarks the way you'd treat a resume: promising, but wait for the wider independent evaluations before you migrate a production pipeline. The pattern we track across this year's launches says the real gap between rivals usually shows up in weeks two through six, not on announcement day.
What this means for your workflow
- Long-context work becomes routine. A 1M-token window means entire content archives, full documentation sets, or a complete product catalog fit in one request. At $4/Mtok input, ingesting a 300,000-token knowledge base costs about $1.20 — cheap enough to re-run on every update instead of maintaining a separate index.
- Memory changes editorial workflows. If you produce serialized content — newsletters, tutorial series, client reports — Active Memory means the model holds your established framing and terminology instead of being re-briefed each session. Cached input at $0.20/Mtok makes repeated work on the same corpus dramatically cheaper.
- Price pressure keeps compounding. With xAI selling Grok 4.7 at $2/$6 per Mtok and Meta now at $4/$20 with a larger window, frontier capability keeps getting cheaper per unit of work. Anyone still budgeting 2025 prices for AI-assisted production is overpaying.
The lowest-friction evaluation path: create a key on Meta's developer portal, run one request against a sample of your own content — twenty product descriptions, two long articles — and compare the output to your current tool on identical inputs. If the delta is visible, move your reference corpus into the cache and let the $0.20/Mtok tier compound the savings from there. Teams that need the Full variant specifically have exactly one road in: a Muse Code subscription, which only pays for itself above a steady daily usage volume.
One honest caveat for non-English publishers: the announcement does not specify language coverage for the API model, and we don't confirm language claims without testing them ourselves. Meta's consumer apps serve massive Arabic- and Asian-language audiences, but API-grade quality per language is a separate question.
Quick comparison
| Muse Wonder | Muse Spark 1.1 | Grok 4.7 | |
|---|---|---|---|
| Launched | Oct 12, 2026 | Jul 2026 | Sep 2026 |
| Input / Mtok | $4 | lower tier | $2 |
| Context | 1M tokens | smaller | smaller |
| Positioning | agentic + frontier coding | daily personal agent | general performance |
| Where | API, Muse Code, OpenRouter | Meta apps | X, Grok apps |
In practice: Muse Spark remains the in-app assistant layer, Grok competes on raw price, and Muse Wonder is the build-things-on-top option for professional work.
A worked cost example
Abstract pricing means nothing until you map it onto a real project. Say you run a content site and want an assistant that knows your entire back catalog:
- Step one — load the corpus. Ingesting a 250,000-token archive (roughly 400 product descriptions or 150 long articles) costs about $1 in input tokens. Because cached input runs at $0.20/Mtok, every subsequent session that references the same archive pays roughly five cents for the cache — not another dollar.
- Step two — daily production. A session generating 20 new pieces at ~15,000 output tokens costs around $0.30 in output plus trivial input fees. Even heavy months of batch production stay within a few dollars.
- The takeaway: the architecture rewards you for keeping a stable knowledge base and iterating on top of it. The bigger and more fixed your reference corpus, the more the cached-input economics tilt in your favor.
Compare that against the $50/month Power Usage plan in Muse Code: the subscription makes sense if you're running coding agents continuously all day or specifically need Muse Wonder Full. For intermittent content production, metered API usage stays cheaper in most scenarios we can construct.
Limits and honest drawbacks
- No pay-as-you-go path to the Full variant. The strongest model sits behind a $50/month minimum subscription. Evaluators and small teams are effectively priced out of the top tier.
- Staged rollout. "Rolling out first to subscribers" means the consumer Muse experience won't change overnight — if you're waiting for it inside the apps you already use, you'll wait.
- Vendor-reported numbers. Most headline benchmarks come from Meta's own reporting. Fine for direction, insufficient for a migration decision.
- Text and code only. This is not an image or video model; creators expecting multimodal generation should look elsewhere in Meta's stack.
And if your production line is multilingual publishing rather than raw model access, purpose-built editors still beat wiring APIs together yourself — tools like ARWriter's Arabic-first writing and publishing platform handle the workflow layer so the model layer stays optional.
Frequently asked questions
What is Meta Muse Wonder?
A new AI model family Meta announced on October 12, 2026, positioned as its most capable to date, focused on agentic personal and professional tasks, expert domains like law, finance, and health, and frontier-level coding.
How much does Muse Wonder cost?
The standard muse-wonder-1.0 model costs $4 per million input tokens, $0.20 per million cached input tokens, and $20 per million output tokens via the Meta Model API. Muse Wonder Full requires a Muse Code subscription at $50 or $200 per month.
Is Muse Wonder available on OpenRouter?
Yes, Meta's developer portal lists OpenRouter among the ways to use the model alongside the Meta Model API and Muse Code.
What is the difference between Muse Wonder and Muse Wonder Full?
The standard model is open to everyone via API and Muse Code with per-token pricing, while Full is the higher-performance variant available only inside Muse Code's Power Usage and Maximum Usage subscription plans.
Does Muse Wonder replace Muse Spark?
No. Muse Spark powers the everyday in-app assistant experience, while Muse Wonder targets professional agentic and coding work through the API and Muse Code, and will gradually power more of the Muse experience for subscribers.
Is Muse Wonder good for marketing and product copy?
The 1M-token window is built for exactly that pattern: load your catalog or article archive once, then work on top of it with Active Memory holding your brand voice. The low input price makes experimentation cheap — but test it against your current tool on a sample of your own content before committing a workflow to it.
Official sources: Meta's Muse Wonder announcement and the official pricing and model card.
