Meta released Muse Wonder on 12 October, the model TechCrunch calls Meta's first frontier model. The API price is $4 per million input tokens and $20 per million output tokens, exactly what Claude Opus 5.5 charges. The first independent test from Artificial Analysis gives it a score of 52, six points behind Opus. This article covers what Meta actually shipped, what a finished job costs once you count the tokens the model uses, how Wonder compares with Claude Opus 5.5, Claude Sonnet 5.5 and GPT-6 Astra, and who should switch to it. The short version: it is not the cheapest way to buy a score of 52, and the better version is not on the API yet.
What Meta released on 12 October
Meta announced two models on the same day. According to Meta's own announcement and model card:
- Muse Wonder. The model you can call through the Meta Model API as
muse-wonder-1.0. Meta's developer documentation labels it "in preview". - Muse Wonder Full. The stronger version. It is available to Muse Code subscribers only, with no pay-as-you-go API price.
- Inputs and context. Text, images, audio and video go in, text comes out, with a context window of about 1 million tokens, according to the model card.
- What Meta emphasises. "Active Memory" (keeping your context across conversations), "Active Focus" (trying alternatives when it hits a blocker and checking in before important actions), expertise in law, finance and health, and long coding projects.
- No downloadable weights. VentureBeat reports that neither model can be downloaded, so this is not an open-weight release like Meta's Llama line.
The Meta API is compatible with the OpenAI and Anthropic software kits, so if your tools already talk to one of those, trying Wonder is a change of address and key rather than a rewrite. That is useful for a tester. It does not make Wonder better.
What Muse Wonder costs, and what the price list leaves out
From Meta's pricing page, checked on 13 October 2026:
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| Meta Muse Wonder | $4.00 | $20.00 |
| Claude Opus 5.5 | $4.00 | $20.00 |
| Claude Sonnet 5.5 | $2.00 | $10.00 |
| GPT-6 Astra | $10.00 | $50.00 |
Three details matter for a budget:
- Cached input is $0.20 per million. If your agent re-reads the same long instructions or documents on every call, that is a 95% discount on the repeated part. Artificial Analysis lists the same discount.
- Web search costs extra. Meta charges $2.50 per 1,000 search queries on top of the tokens.
- One flat rate at any length. Meta says the token price does not change with how much of the 1M-token window a request uses.
On the face of it, Wonder is priced like Opus. As we showed with cost per task, though, the price per token tells you very little. What you pay is price multiplied by how many tokens the model burns to finish the job.
What Artificial Analysis measured: a score of 52 and $4.52 per task
Artificial Analysis is an independent tester that runs the same set of tasks on every model. For Muse Wonder, its model page shows a score of 52 on its Intelligence Index (ninth of 216 models), $4.52 per average index task, and 130 million output tokens used to run the index, against a median of 85 million. It also lists a speed of 87.9 output tokens per second. VentureBeat reports 56 for Wonder Full.
Read it in three steps:
- Against Opus 5.5. Same price per token, but Wonder used 130 million output tokens where Opus used 260 million. That makes a task about 24% cheaper ($4.52 against $5.98) for a score six points lower.
- Against Sonnet 5.5. Sonnet charges half as much per token but used 420 million output tokens. Sonnet costs about 21% more per task and scores four points higher.
- Against GPT-6 Astra. This is the awkward one. Astra's tokens cost two and a half times as much, but it used only 60 million of them. It scores one point higher and costs about 28% less per task ($3.26).
Muse Wonder is priced like a premium model, scores like a mid-tier one, and costs more per task than the model one point above it.
Where Meta's own benchmarks say Wonder wins
Meta's model card compares Wonder with GPT-6 Astra and Claude Opus 5.5. VentureBeat reproduced the table. These are Meta's numbers, and Meta describes its method: where a third-party or provider result exists it uses that, otherwise it takes the highest of its own validated run, a self-reported score or an official leaderboard score. It also says its runs of competitor models are "best-effort". Treat this as a vendor claim with some independent sources inside it.
| Benchmark | Muse Wonder | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|
| Harvey LAB-AA (legal work) | 10.0 | 8.6 | 4.2 |
| SWE-Marathon (long coding projects) | 53.1 | 42.5 | 46.3 |
| DeepSWE 1.1 (software engineering) | 78.8 | 74.1 | 74.2 |
| AutomationBench-AA (business workflows) | 39.1 | 41.6 | 41.2 |
| JobBench (professional tasks) | 70.5 | 66.1 | 75.4 |
| Terminal-Bench 4.0-AA | 52.0 | 59.1 | 59.6 |
| HealthBench Pro | 60.5 | 65.3 | 65.6 |
The pattern is clear. Wonder leads on legal work and long coding projects, and trails on terminal tasks, health questions and the business-workflow test most relevant to an operations manager. Meta's card also notes that the SWE-Marathon row is not a controlled comparison, because the rivals were run through their own coding products. Seven rows are a sample chosen by the vendor, not a verdict.
The data terms matter more than the price
This is the part of the launch most coverage skips, and it affects a business more than a three-point score gap.
- Standard mode. Meta says content you submit is not used for training. The pay-as-you-go Wonder API is available only in this mode, according to VentureBeat.
- Contributor mode. This is the default for Muse Code subscriptions, which is where Wonder Full lives. In it, Meta may use your prompts, code, files and outputs to improve future models. You can turn it off, but only for future submissions.
- Zero data retention. Offered to qualified API organisations after separate approval. It does not support file uploads, background requests, server-side conversation state or built-in web search.
In plain terms: the stronger model comes with the weaker privacy default. If you paste customer lists, contracts or financial statements into a coding or agent tool, check that setting before anyone on your team does. That is true of every vendor, but here the default is the one that shares.
A warning for Urdu and Arabic users
Meta's model card says the model was trained on multilingual data, but also that it "has not been evaluated on all languages" and that performance "may degrade on languages outside the strongly supported set". It does not list which languages are strongly supported. If your customers write in Urdu, Arabic or Roman Urdu, test with your own messages before you move anything. We have not tested Wonder in these languages, and nobody has published an independent result we could cite.
We also could not find a list of supported countries on Meta's developer pages, so check that your company can open an account and pay before planning around it.
Who should switch, and who should wait
- Teams on Claude Opus 5.5 paying for long, token-heavy jobs. Worth a test. Same price per token, about a quarter less per task on Artificial Analysis' index, and better scores on Meta's legal and long-coding rows. You give up six index points.
- Teams choosing a new default model for an app. Look at GPT-6 Astra first. It scores a point higher and costs less per task. Claude Sonnet 5.5 is the pick if you want four more points and can afford about a fifth more per task. See our Opus 5.5 pricing guide and the ChatGPT vs Claude vs Gemini comparison for Pakistan for the full price lists.
- Anyone who wants Wonder Full. Wait. It has no API price today. VentureBeat reports Muse Code plans at $50 and $200 a month. Its 56 would put it level with Sonnet 5.5, but you cannot put it behind your own product yet.
- Small businesses buying a ready-made product. Do not choose by model name. Our AI Chat Assistant and Accounts modules let you work with AI without managing API keys, and the model underneath can change without you rebuilding anything.
What we would do this week
Pick ten real jobs from your own business. For example, ten customer messages, ten invoices to code, or ten contract clauses to summarise. Run them through Wonder, Astra and Sonnet, and add up the tokens each one used. That one-afternoon test will tell you more than any leaderboard, because the cost per task depends on how wordy the model is on your work, not on the benchmark's.
Meta's earlier product, the Muse agent that acts inside your email and calendar, is a different thing from this API model; we covered its business risks in OpenAI Dots vs Meta Muse vs Copilot Autopilot. To talk through which model fits your workload, contact us.
This is a launch-week article. Wonder is in preview and Artificial Analysis will likely re-run it, so the scores and costs here are as of 13 October 2026.
