Meta AI chief Alexandr Wang told CNBC the Muse Wonder API would be priced "meaningfully below" leading models from OpenAI and Anthropic. Meta's price sheet, published the same Monday as the company's first frontier model, lists $4 per million input tokens and $20 per million output tokens, exactly what Claude Opus 5.5 costs.
A finished Wonder task costs more than one on GPT-6 Astra, even though Astra's list price is much higher, because Wonder writes far more tokens before it arrives at an answer. Wonder is also twice the list price of GPT-6.1 Sol. On the sticker, Wang's claim holds against that top tier, Astra and Claude Fable 5.1. It doesn't hold against Opus or Sol, or against Astra once you pay for the tokens a task actually uses. Per finished task, Wonder costs less to complete than Opus or Fable.
Wonder Full isn't on the API, so there isn't a second token rate to shop (TechCrunch). Contributor mode's discount, the cheaper rate if you let Meta train on your prompts, isn't offered for Wonder either, which leaves the published $4 and $20 as the price you pay.
The sticker matches Opus
Prices below are standard list rates, and the index and per-task cost are Artificial Analysis's.
| Model | List price, input / output per 1M tokens | Intelligence Index | Cost per index task |
|---|---|---|---|
| Meta Muse Wonder (max) | $4 / $20 | 52 | $4.52 |
| Claude Opus 5.5 (max, with fallback) | $4 / $20 | 58 | $5.98 |
| Claude Fable 5.1 (max, with fallback) | $10 / $50 | 53 | $7.63 |
| GPT-6 Astra (max) | $10 / $50 | 53 | $3.26 |
| GPT-6.1 Sol (max) | $2 / $10 | 52 | $0.72 |
| Gemini 4 Argon (high, limited access) | $2 / $10 intro, then $4 / $20 | 53 | $1.99 |
Opus 5.5 costs the same as Wonder and scores higher on the index, so leaving Opus doesn't save you money on the rate card. Put one job you already run on both, and see whether Wonder's memory and a second attempt are worth a matching price with a lower score. Sol matches Wonder's score for much less. There, the test is whether it finishes that same job. A lower rate only helps if the work comes back done.
Per task, Astra costs less
Astra's list price sits well above Wonder's, and Astra still costs less per task. The index run shows how. Wonder generated about 130 million output tokens, somewhat verbose next to about 60 million for Astra. You're paying Wonder's output rate on a much longer reply, and that is how the lower sticker loses. Output-heavy agent work is where you feel it: a long trace, a document the model keeps revising, a task that has to try again. Cached input is $0.20 per million tokens, which helps when the same prompt comes back. It doesn't offset the output tokens on a verbose reply.
Wonder ranks ninth of 219 models on that index, so this isn't a bargain model buried down the board. Vals AI puts it tenth of 47 overall and first on CyberBench v1.1 at 80.36%, the area where Wonder has its strongest safeguards to date. A fuller safety report is on the way, and until it shows up you don't have the lab's full account of those safeguards.
Meta's internal tests are what it uses to argue the new models can rival GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1 (The Verge). Those comparisons are the company's own, and they aren't the index in the table, so they don't settle the price question. On MemoryBench, Meta's own testing puts Wonder ahead of Fable, Opus, and Astra. A lead in Meta's own testing is worth trying on a job you already run. It still isn't a reason to switch on price.
If you already pay for Opus 5.5, price isn't a reason to move. Switch only if Wonder's memory, and the way it sticks with a long task, wins on your own work. If the cost of a finished task is what you care about, run your evals on GPT-6.1 Sol first. Keep an eye on the safety report Meta has promised, and on the day Gemini 4 Argon's intro price ends.
Photo: Nokia621 / Wikimedia Commons, CC BY-SA 4.0.
