Muse Wonder and Wonder Full in Muse Code: plans, prices, and the launch-day bug
Meta’s first frontier models arrived on October 12 with a new $200 Maximum plan. Here’s which Muse Code plan gets Wonder or Wonder Full, what Wonder costs on the API next to Spark 1.3, and whether an upgrade is worth it.
What launched on October 12
On October 12, 2026, Meta released Muse Wonder and Muse Wonder Full, which it calls its first frontier models and its most capable to date. They are larger than Muse Spark, which Meta built for low latency, and were trained at Meta’s Prometheus data center in Ohio. Both are available in Muse Code from day one.
The two models are not sold the same way:
- Muse Wonder is available in Muse Code and on the Meta Model API as
muse-wonder-1.0. - Muse Wonder Full is available only to Muse Code subscribers on the Power and Maximum plans. It is not on the API and has no per-token price.
Meta highlights two behaviours. Active Memory carries relevant context across conversations. Active Focus keeps working toward a goal: it tries other approaches, recovers from errors, and checks in before important actions. Meta aims the models at expert domains such as law, finance, and health, as well as at frontier coding.
Which Muse Code plan includes Wonder Full?
Only the two most expensive ones. Alongside Wonder, Meta added a fourth plan, Maximum Usage at $200 a month:
| Plan | Price | Usage vs. Everyday | Muse Wonder | Muse Wonder Full |
|---|---|---|---|---|
| Everyday Usage | $5/month | 1× (10–50 prompts every 5 hours) | Yes | No |
| High Usage | $15/month | 5× | Yes | No |
| Power Usage | $50/month | 20× | Yes | Yes |
| Maximum Usage | $200/month | 80× | Yes | Yes |
If you read our comparison of Meta Muse and Muse Code earlier, note two changes. First, the multipliers went up: at launch in August, High was 3× and Power 10×. By September 21, Meta’s subscription page showed 5× and 20× at the same prices. Meta made no announcement, so the exact date of the change is unknown. Second, both products now have a plan called Maximum — $100 a month for the consumer Muse agent, $200 a month for Muse Code. Check which product a receipt or forum post refers to before comparing numbers.
Upgrades take effect immediately and are prorated. Downgrades apply from the next billing cycle. There are no paid top-ups: when you hit your limit, you upgrade or wait for the allowance to refresh, and on Maximum you can only wait.
What Wonder costs on the API
Through the Meta Model API, muse-wonder-1.0 costs well above Spark 1.3, and it has no cheaper Contributor variant:
| Per 1M tokens | muse-wonder-1.0 |
muse-spark-1.3 |
muse-spark-1.3-contributor |
|---|---|---|---|
| Input | $4.00 | $1.25 | $0.10 |
| Cached input | $0.20 | $0.15 | $0.002 |
| Output | $20.00 | $4.25 | $0.20 |
The rate limit is 3,000 requests per minute. TechCrunch placed the price in a similar band to GPT-6 Sol and Claude Opus 5.5, and third-party listings report that Wonder is still labeled preview on the API.
To see what this means in practice, take a hypothetical agent task that reads 200,000 uncached input tokens and writes 20,000 output tokens:
- Spark 1.3: 0.2 × $1.25 + 0.02 × $4.25 = $0.25 + $0.085 = about $0.34
- Wonder: 0.2 × $4.00 + 0.02 × $20.00 = $0.80 + $0.40 = $1.20
That is roughly 3.6 times the cost for the same token count. Output-heavy work widens the gap, because Wonder’s output rate is nearly five times Spark’s. Whether Wonder uses fewer or more tokens than Spark for the same job is something Meta hasn’t published, so measure it on your own tasks. If you hit your API keys hard, a subscription may be cheaper; the pricing page has an estimator.
How to switch models in Muse Code
Use /model. Since CLI 1.4.5 you can pass a model ID directly, for example:
/model muse-wonder-1.0Running /model on its own opens the picker. Meta’s documentation doesn’t publish a separate model ID for Wonder Full, so if you’re on Power or Maximum, choose it from the picker rather than guessing an ID. If /model <model-id> isn’t recognised, check muse --version and update first.
The launch-day usage bug
On launch day, users reported that Wonder and Wonder Full were draining their plan allowances unusually fast. Alexandr Wang, Meta’s Chief AI Officer, confirmed that Wonder usage was being counted against limits in error. The same day, Meta said the issue had been identified and fixed, and that usage had been reset. Some users reported lingering problems afterward.
What to make of it: launch-day posts along the lines of “three prompts used half my week” are not a reliable guide to how far a plan stretches with Wonder. Meta has not published a per-model usage cost for its plans, so the honest answer is that nobody outside Meta knows yet. If your allowance still drops faster than it should, check whether the reset reached your account before drawing conclusions.
How good is it?
Most of the numbers so far come from Meta. Meta says Wonder is comparable to Fable 5.1 and GPT-6 Astra, and that on MemoryBench it scores higher than Fable, Opus, and Astra in Meta’s own testing. Its launch materials also cite Harvey LAB, FocusBench, and SWE-Marathon.
Independent figures are thinner. According to reported Artificial Analysis Intelligence Index results, Claude Opus 5.5 scores 58, Wonder Full 56, and Wonder 52. That puts Wonder Full close to the top of the field rather than clearly ahead of it. Meta says the models were developed under its Superintelligence Scaling Framework and that a safety report is “available soon”; it was not out at launch.
Contributor mode is on by default
Muse Code subscriptions default to Contributor mode: in exchange for more plan usage, Meta may use your content to train future models. CLI 1.4.6 made this the default for subscribers. You can turn it off in /settings, after which future content isn’t used, but your usage allowance goes down — Meta doesn’t say by how much. Wonder and Wonder Full work in either mode. Our Contributor vs. Standard breakdown covers the trade-off in more detail.
Should you upgrade to Power or Maximum?
- Stay on Everyday or High if Spark 1.3 already handles your work. You get plain Wonder on every plan, so you can try it on hard tasks without upgrading.
- Move to Power ($50) if you want Wonder Full or regularly run out of allowance on High. Upgrades are prorated, so trying it for the rest of a cycle costs less than a full month.
- Consider Maximum ($200) only if you already hit Power’s limits regularly. It gives the same models as Power — the only difference is 4× more usage.
- Use the API for scripted or CI work with Wonder, keeping in mind that Wonder Full isn’t available there at any price.
Give the post-launch metering a week or two to settle before judging any plan on its usage. The consumer Muse app is still on Spark 1.3; Meta says Wonder is coming there “soon,” starting with subscribers. For Spark 1.3, still the cheaper default, see our Muse Spark 1.3 write-up.