Meta's Muse Wonder costs 2.8 times as much per task as Muse Spark 1.3
Meta's Muse Wonder costs $4.52 per Intelligence Index task, about 2.8 times the $1.60 of Muse Spark 1.3, according to Artificial Analysis. The extra $2.92 a task buys 4 index points, a price that OpenRouter's search tests suggest only hard tasks repay.
Not a builder's beat, but builders have a standing stake in it.
Muse Wonder (max) costs $4.52 per Intelligence Index task, against $1.60 for Muse Spark 1.3 (max). Muse Wonder (max) gains 4 Intelligence Index points over Muse Spark 1.3 (max) and lands at 52.
| Figure | Value | Note | Claim |
|---|---|---|---|
| Muse Wonder (max), per Intelligence Index task | $4.52 per task | 2 | |
| Muse Spark 1.3 (max) | $1.6 per task | 3 | |
| Muse Wonder (max) Intelligence Index gain | 4 index points | over Muse Spark 1.3 (max), lands at 52 | 4 |
What happened
- Muse Wonder Full scored 56 and Muse Wonder 52 on the Intelligence Index at max effort, six weeks after Muse Spark 1.3 scored 48.
- Artificial Analysis puts Muse Wonder above most similarly intelligent models on cost per task, with Gemini 4 Argon at $1.99 and GPT-6.1 Sol at $0.72, though below Claude Fable 5.1 at $7.63.
- OpenRouter's search benchmarks found that going from 1 turn to 25 roughly doubles the score at 2.5 to 7 times the cost per question.
Why it matters
- What it costs cost Moving from Muse Spark 1.3 to Muse Wonder costs about $0.73 per index point, and the team running the workload pays it on every task.
- What it limits constraint Muse Wonder Full, the higher scorer, has no per-token price and is open only to Muse Code users, so API buyers cannot call it or budget for it.
- Who has to decide decision For simple workloads, OpenRouter's HLE result argues for staying shallow: similar scores at triple the cost.
- Who's exposed exposure Workloads that mostly fail carry the worst bill, since OpenRouter saw an 81-search attempt that was still graded incorrect.
The sources do not report OpenRouter traffic by model, so they cannot show how much usage has stayed on older, cheaper models. They do show where Muse Wonder's price comes from. Artificial Analysis attributes the cost to premium API pricing and to heavy input use, 7.69M tokens per task [6]. Output is not the driver. The model uses around 54k output tokens per task, down about 10% from Muse Spark 1.3 [13].
Divide the cost gap by the score gap and the upgrade has a per-point price that a team can test against its own tasks [16]. The yardstick is built on Artificial Analysis's tasks, and a team's own invoice scales with its own token counts.
OpenRouter's search tests give the shape of the decision, though they vary search depth and not the model, and the 1-to-25 turn result comes from BrowseComp runs only [12]. In our view each workload needs two answers: how hard its tasks are, and how often the model fails them. Hard tasks that the model usually completes are where more depth or a frontier model pays. Easy tasks, and tasks that mostly fail, belong on a capped, cheaper setup, since the HLE and WideSearch examples show spend rising without the score following [10][11].
A team can run a sample of its own tasks through both models, divide the extra cost by the extra score, and compare the result with Artificial Analysis's per-point figure [16]. If its own gain is smaller per dollar, the older model is the better buy for that workload.
What to watch
- Whether OpenRouter usage rankings show Muse Wonder taking traffic from cheaper models.
- Whether Meta puts a per-token price on Muse Wonder Full or opens it beyond Muse Code.
- Whether OpenRouter's live leaderboards change the frontier-versus-budget results at 25 turns.
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Muse Wonder and Muse Wonder Full scored 52 and 56 on Artificial Analysis Intelligence Index v4.3 at max reasoning effort, against 48 for Muse Spark 1.3 (max), with a release gap of only 6 weeks.
- [2]
Muse Wonder (max) costs $4.52 per Intelligence Index task.
- [3]
Muse Spark 1.3 (max) costs $1.60 per Intelligence Index task.
- [4]
Muse Wonder (max) gains 4 Intelligence Index points over Muse Spark 1.3 (max) and lands at 52.
- [5]
Muse Wonder is priced at $4 input, $0.20 cached input and $20 output per 1M tokens.
- [6]
Artificial Analysis says premium API pricing and high input token use (7.69M per task) push Muse Wonder (max)'s cost to $4.52 per Intelligence Index task.
- [7]
Muse Wonder Full is not priced per token, so no Cost per Task figure is available, and it is only available to Muse Code users; Muse Wonder is available on Meta's first-party API and Muse Code.
- [8]
Muse Wonder (max) cost per task is higher than most similarly intelligent models: lower than Claude Fable 5.1 ($7.63), but higher than Gemini 4 Argon ($1.99) and GPT-6.1 Sol ($0.72).
- [9]
OpenRouter reports that increasing search from 1 turn to 25 roughly doubles the score while costing only 2.5-7 times more per question.
- [10]
On HLE, GPT-5.6 Sol with Perplexity scored similarly between 1 turn and 25 turns, for triple the cost.
- [11]
OpenRouter found that models would exhaust their search budget attempting to find an answer they would eventually fail; the deepest attempt, 81 searches on a WideSearch table, was still graded incorrect, and worst-case cost is driven by failure rate.
- [12]
OpenRouter's budget runs cover BrowseComp only, using the server tool with ten results per search, no page fetching or code execution, and the latest qualifying run per configuration.
- [13]
Muse Wonder (max) uses around 54k output tokens per Intelligence Index task, down about 10% from Muse Spark 1.3 (max) at 60k.
- [14]
Muse Wonder (max) costs about 2.8 times as much per Intelligence Index task as Muse Spark 1.3 (max).
- [15]
The cost gap per Intelligence Index task between Muse Wonder (max) and Muse Spark 1.3 (max) is $2.92.
- [16]
The upgrade from Muse Spark 1.3 to Muse Wonder costs about $0.73 per Intelligence Index point.
Sources & coverage · 2 publishers
The reporting this story was synthesized from, earliest first. Every link goes to the original.