Meta's Muse Wonder is out at $4/$20 - the model you can call trails Claude Opus 5.5

Meta released Muse Wonder on 12 October, its first frontier model sold through a paid API. The API model, `muse-wonder-1.0`, costs $4 per million input tokens and $20 per million output, with cached input at $0.20. It has a context window of just over a million tokens with no surcharge for long requests. It accepts text, images, audio, video and PDFs and returns text. Meta says it was built for long-running agent work.

There is a second model, Muse Wonder Full, and most of Meta's launch charts lead with it. You cannot buy it through the API. VentureBeat reports it is only available inside Meta's Muse Code subscription, on the $50 and $200 a month plans. If you are building an automation, Full is not the model you will be running.

What is actually new

The price is not the story. Claude Opus 5.5 has the same $4/$20 rate card. Alexandr Wang said the API would come in "meaningfully below" OpenAI and Anthropic. As Handy AI pointed out, that holds against Claude Fable 5.1 at $10/$50, but not against Opus.

Three things are worth noticing.

**Long context at a flat rate.** A million tokens with no long-context premium is useful for work that means reading a lot at once: contract bundles, a year of support tickets, a large policy manual.

**A lower hallucination rate.** On Artificial Analysis's AA-Omniscience test, Handy AI reports Wonder at 26.5%, about half of Opus 5.5's 58.6%. For extraction and question-answering over your own documents, wrong-but-confident answers are the expensive failure. This is the number I would test first.

**Cheaper per task than the rate card suggests.** Artificial Analysis estimates Wonder costs $4.52 per average task on its index, against $5.98 for Opus 5.5, according to VentureBeat. Same per-token price, fewer tokens used. GPT-6 Astra comes in lower again at $3.26.

Meta also ships "Active Memory", which carries context across conversations, and "Active Focus", which is meant to keep an agent working after a step fails.

What it means for a business

Put the marketing aside and look at the benchmark closest to business automation. On AutomationBench-AA, which tests business workflows, Meta's own table has the API model at 39.1%. Opus 5.5 scored 41.2% and GPT-6 Astra 41.6%. Only Full leads, at 45.4%, and Full is the one you cannot call from your code.

On the independent Artificial Analysis Intelligence Index, Wonder scores 52. Opus 5.5 scores 58, and Claude Sonnet 5.5 scores 56. Sonnet 5.5 costs $2/$10, half of Wonder's price. Meta's comparison table leaves Sonnet out.

So for a typical agent that reads an email, checks a CRM, updates a record and drafts a reply, Wonder gives you no reason to switch today. You would pay Opus prices for a model that scores below Opus, and double Sonnet's price for a model that scores below Sonnet.

Where it could earn a place is narrower: high-volume document work where you need the whole file in context and wrong answers are costly. Think pulling terms out of long contracts, or answering staff questions from a big internal knowledge base. The flat long-context price and the low hallucination rate both point that way. That is a hypothesis to test, not a result.

Switching to test it is cheap. The API supports OpenAI- and Anthropic-compatible interfaces, so in most stacks it is a base URL and a model name.

The honest caveats

Most of the headline numbers are Meta's own evaluations at maximum reasoning effort. Several coding comparisons ran rivals in different tools. VentureBeat notes Meta's table mixes third-party and internal results.

The data handling needs reading before anything sensitive goes near it. The API runs in Meta's standard mode, where submissions are not used for training. Muse Code subscriptions default to "Contributor mode", where Meta may train on your prompts, code and files. Zero data retention is available to approved API customers. VentureBeat reports it switches off server-side conversation state and built-in web search, which undercuts the memory features Meta is selling.

The safety documentation is incomplete. Handy AI reports the model card says Meta cannot rule out the unmitigated model reaching its high-risk threshold for chemical and biological misuse, and the full safety report is still pending. Meta itself recommends human confirmation before agents take consequential actions. That is good advice for any model, and it matters more for one this new.

There are no open weights and no self-hosting. If you were hoping Meta's name meant a Llama-style download, it does not.

And it is day two. Prices, limits and scores on launch-week models tend to move. We watched DeepSeek quadruple a price three days after launch in August.

What to do about it

If you run document-heavy automations where accuracy matters more than speed, take 20 real cases you have already checked by hand. Run them through your current model and through `muse-wonder-1.0`, and compare wrong answers, not average scores. Track the total cost of each run, not the per-token price.

If your automations are mostly routing, updating records and drafting replies, leave them alone. Nothing here beats what you already have at a lower price. Check again when Wonder Full gets an API price, or when independent workflow benchmarks catch up with Meta's claims.