Meta priced Muse Wonder at Claude Opus 5.5's $4 and $20 to the cent. Opus on high effort outscores it for 40% of the cost per task, and the Wonder that beats Opus at coding isn't sold by the token.
Meta's first frontier model went on sale yesterday. Every price you can pay for it is a price Anthropic already charges, and the scores it buys are lower.
Photo by Aryo Yarahmadi on Unsplash
Short version: if you were going to pay $4 and $20, Opus 5.5 gets you more for it, and on a batch job it gets you the same tokens for half. Wonder's case rests on Meta's coding numbers, and the model those numbers flatter most, Wonder Full, is the one you can't call.
The rate card, line by line against Opus 5.5
Meta Superintelligence Labs released Muse Wonder on October 12 as muse-wonder-1.0, in preview on the Meta Model API and in Muse Code. OpenRouter carries it as meta/muse-wonder at the same rates. We didn't find it on Bedrock, Azure, Vertex or Together.
| Per 1M tokens | Muse Wonder | Claude Opus 5.5 |
|---|---|---|
| Input | $4.00 | $4.00 |
| Cached input | $0.20 | $0.20 |
| Output (incl. reasoning) | $20.00 | $20.00 |
| Cache write, 5 min | none listed | $5.00 |
| Batch input / output | no batch tier | $2.00 / $10.00 |
| Context window | 1,048,576 | 1,000,000 |
Meta Model API pricing page and Anthropic pricing page, both read October 13, 2026. Web search grounding on Wonder is $2.50 per 1,000 queries. Rate limits 3,000 RPM and 4M TPM per team.
Two lines differ, and they pull in opposite directions. Meta doesn't charge to write a cache, where Anthropic bills 1.25x input for a 5-minute write. If you rebuild a large prompt prefix often, that's real money in Wonder's favor. Meta also has no batch discount. It offers "background responses" for async work, capped at 600 submissions a minute, but bills them at the full rate.
The batch gap is bigger than the cache gap for most teams. Any eval run, backfill or nightly summarization job that can wait a few hours costs half as much on Opus 5.5.
An agent loop that mostly reads
Say an agent reads 80 million tokens a month and writes 8 million, a 10-to-1 ratio that's normal for code agents re-reading files. No caching, prompts under 200K so nobody's long-context tier kicks in:
| Model | Input / output | Month |
|---|---|---|
| Muse Wonder | $4 / $20 | $480 |
| Claude Opus 5.5 | $4 / $20 | $480 |
| Claude Opus 5.5, batch | $2 / $10 | $240 |
| Claude Sonnet 5.5 | $2 / $10 | $240 |
| GPT-6.1 Sol | $2 / $10 | $240 |
| Grok 4.7 | $2 / $6 | $208 |
| Muse Spark 1.3 | $1.25 / $4.25 | $134 |
Our arithmetic. Different tokenizers count the same text differently, so treat cross-vendor rows as approximate.
Now cache 64 million of those reads. Wonder drops to $236.80: $64 of fresh input, $12.80 of cache hits and $160 of output. Caching cut the bill in half, and what's left is two-thirds output, so the next lever is the effort setting. Meta bills reasoning tokens at $20 too, and if you leave effort unset the model picks its own level. Meta doesn't say what that level is.
What a finished task costs
The rate card ties. The work doesn't. Artificial Analysis ran Wonder at max effort through its Intelligence Index and published a cost per task. Here it is next to the models people will cross-shop:
Artificial Analysis Intelligence Index v4.3, cost per task at list price, read October 13, 2026. Effort level in brackets.
Opus 5.5 on high effort scores 54 for $1.82 a task. Wonder on max scores 52 for $4.52. So for the same rate card you get two more points at 40% of the spend. GPT-6.1 Sol matches Wonder's 52 at $0.72, about a sixth of the price.
One thing surprised us. Wonder wrote about 54,000 output tokens per task, under half the roughly 119,000 Opus 5.5 writes at max. Yet its per-task bill is 76% of Opus's, not half. Working backwards from AA's figures, about three-quarters of Wonder's $4.52 is input, against under two-thirds for Opus. That split is our arithmetic, not a number AA prints, but AA does report Wonder reading around 7.7 million input tokens per task. For agent loops that re-read files, cache everything you can.
Also worth knowing before you build on it: AA measures time to first token at about 19 seconds and output at around 88 tokens a second. It's not a chat model you put in front of impatient users at max effort.
Meta's coding case, and the model it's really about
Meta published a 14-row scorecard. These are the rows people are quoting:
| Benchmark | Wonder | Wonder Full | Opus 5.5 | GPT-6 Astra |
|---|---|---|---|---|
| DeepSWE 1.1 | 78.8 | 79.0 | 74.2 | 74.1 |
| SWE-Marathon | 53.1 | 57.5 | 46.3 | 42.5 |
| Terminal-Bench 4.0 | 52.0 | 57.1 | 59.6 | 59.1 |
| ProgramBench | 50.0 | 59.0 | 70.5 | 52.5 |
Meta's Muse Wonder model card, max effort (Wonder Full's SWE-Marathon at xhigh). DeepSWE scores for the Muse models and Opus are Datacurve's runs; the rest are vendor-reported.
DeepSWE is the one Meta leads with, and the API model really does beat Opus there, 78.8 to 74.2. On long-running repo work (SWE-Marathon) it's ahead too. Then it loses Terminal-Bench and loses ProgramBench by 20 points. Across all 14 rows, the Wonder you can buy beats Opus 5.5 on 5 and loses 9.
Wonder Full is the stronger model: 56 on AA, tied with Sonnet 5.5, and ahead of Wonder on every row above. You can't buy it by the token. It ships only inside Muse Code, on the Power plan at $50 a month and Maximum at $200. Meta's plan page doesn't spell out what the $5 and $15 plans include. If Wonder Full is what you want, the subscription is the only price there is, and Meta measures it in prompts per five hours, not tokens.
The data question
Meta sold Muse Spark at 12.5x off input and 21x off output if you let it train on your prompts. There's no such deal for Wonder. Meta's pricing page says Contributor mode isn't available for Wonder on the API, so every API call runs in Standard mode, where Meta says it doesn't train on prompts or completions.
Muse Code is different. VentureBeat reports subscriptions default to Contributor mode, which lets Meta use your prompts and code. We couldn't find that stated on a Meta page, so check the toggle in your own settings before pointing it at a private repo.
Who should pay $4 and $20 to Meta
Not many teams, yet. If your workload looks like DeepSWE (long agentic fixes across a real codebase), Wonder is worth an A/B test against Opus, and the free cache writes help if your prefixes churn. Run it at high or xhigh before you assume max is needed.
For general reasoning, Opus 5.5 at high effort beats it on score and per-task cost at the same card, and GPT-6.1 Sol ties it for a sixth. If you're already on Muse Spark 1.3, Wonder is four points up on AA for about 2.8x the cost per task.
It's a preview, and preview prices can move. We'll update the Muse Wonder model page if it changes. To price your own mix, try the calculator.
Sources
- Meta: Model API pricing and rate limits - Wonder rates, Contributor availability, limits
- Meta: models and reasoning docs - context window, effort levels, reasoning billing
- Meta: Muse Code plans - subscription tiers, Wonder Full access
- Meta newsroom: introducing Muse Wonder - release date, availability
- Anthropic: pricing - Opus 5.5 and Sonnet 5.5, cache and batch
- Artificial Analysis: Muse Wonder - Intelligence Index, cost per task, tokens, speed
- VentureBeat: Muse Wonder and Muse Wonder Full - Contributor default in Muse Code