Muse Wonder: 52.4 at $4.52, with $3.43 of it on the input side

Meta released Muse Wonder on October 12, 2026 at $4.00 per million input tokens and $20.00 per million output tokens, beside a subscription-only Muse Wonder Full that has no API price. In the Intelligence Index snapshot retrieved Oct 19, 2026, 4:38 AM UTC, Muse Wonder (Max) scores 52.4 at $4.52 per task, ninth of 116 comparable configurations, meaning the rows with a measured cost per task. Four configurations score at least as high for less, so the row is off the cost frontier. Input-side cost is 76% of the $4.52, and the row’s cache-write line is the largest of any row on the chart. The coding-agent chart stores no Muse Wonder row, so this note has no coding-agent score to set beside the Index.

The evaluations behind the row

The Artificial Analysis Intelligence Index scores a model through its API under one harness shared by every model, across 10 evaluations at version 4.3.2: AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience, and AA-LCR v1.1. The categories weigh 30% agents, 20% coding, 20% scientific reasoning, and 30% general capability. Cost per task is the weighted average API bill for one task across those evaluations. The chart keeps five components of it: non-cached input, cache reads, and cache writes on the input side; reasoning tokens and answer tokens on the output side.

A row on the chart is one configuration, the model at one effort level with the settings Artificial Analysis names in the row. The snapshot stores one Muse Wonder row, Muse Wonder (Max), so unlike the Claude Opus 5.5 and Claude Haiku 5.5 notes on this site, there is no effort ladder to tabulate; every figure below is that single row.

Ranked ninth of 116, off the cost frontier

Among the 116 comparable configurations in the snapshot retrieved Oct 19, 2026, 4:38 AM UTC, Muse Wonder (Max) ranks ninth on the index at 52.4. The leader, Claude Opus 5.5 (Max, Default Fallback), scores 57.6 for $5.98 per task, 5.2 points higher at 1.3x the cost. The row is off the chart’s cost frontier. Four configurations score at least as high for less, cheapest first: Claude Opus 5.5 (High, Default Fallback) at 53.6 for $1.82 per task (0.4x the Muse Wonder cost), Gemini 4 Argon (High) at 52.6 for $1.99 per task (0.4x the Muse Wonder cost), GPT-6 Astra (Max) at 52.7 for $3.26 per task (0.7x the Muse Wonder cost), and Claude Opus 5.5 (Xhigh, Default Fallback) at 56.0 for $3.46 per task (0.8x the Muse Wonder cost).

On cost alone the row sits near the top of the chart: 110 of the 116 cost less per task, and five cost more. Within one index point of the 52.4 sit seven configurations: GPT-6.1 Sol (Max) at 51.8 for $0.724 per task, Gemini 4 Argon (High) at 52.6 for $1.99 per task, Claude Sonnet 5.5 (Xhigh, Default Fallback) at 51.9 for $2.01 per task, GPT-6 Astra (Xhigh) at 52.4 for $2.31 per task, GPT-6 Astra (Max) at 52.7 for $3.26 per task, Claude Fable 5.1 (Xhigh, Default Fallback) at 53.2 for $5.98 per task, and Claude Fable 5.1 (Max, Default Fallback) at 53.4 for $7.63 per task.

The chart prints one decimal, and at one decimal GPT-6 Astra (Xhigh) shows the same 52.4. At stored precision GPT-6 Astra (Xhigh) scores 0.01 points lower, so it is not among the rows that beat Muse Wonder on both axes; it costs $2.31 per task, 0.5x the Muse Wonder cost, for a score that prints the same.

The table lists the five highest-scoring configurations after it, with each cost as a multiple of the $4.52. Every one of them is cheaper than the Muse Wonder row, and the widest gap is 1.2 points.

Where the $4.52 goes

Of the $4.52 per task, $3.43 is input-side cost and $1.08 is output. Largest first, the five components are cache writes at $1.92, cache reads at $1.44, reasoning tokens at $0.574, answer tokens at $0.511, and non-cached input at $0.071. Cache writes alone are 43% of the bill, and the row’s cache-write bill is the largest of any of the 116 comparable configurations: no row on the chart pays more for that line. Its input-side bill is the fifth-largest and its output-side bill the ninth-largest, so the row is costly on both sides but far more so on input.

Input-side cost is 76% of the total. Every configuration ranked above it spends a smaller share of its bill on input; the highest share among the eight configurations above it is 69%. The model writes 54,232 output tokens per task, 53% of them reasoning tokens, so the output side is not where the $4.52 comes from.

The table splits the bill of each row that beats Muse Wonder on both axes, plus the row that prints the same score, into its two sides, cheapest first. GPT-6 Astra (Max) and Claude Opus 5.5 (Xhigh, Default Fallback) spend more than Muse Wonder on output and still cost less in total, because the input side is where the gap is: $1.90 against $3.43 for GPT-6 Astra (Max) and $2.15 against $3.43 for Claude Opus 5.5 (Xhigh, Default Fallback).

Artificial Analysis’s model page, captured October 21, 2026 UTC, lists the prices behind these figures as $4.00 per million input tokens and $20.00 per million output tokens with a 95% cache discount, and rounds the row to an index score of 52 at $4.52 per task. It counts 130M output tokens to run the whole index, which it calls “somewhat verbose in comparison to the median of 82M,” measures 152.2 output tokens per second, lists the model as proprietary and released October 2026, writes that it “supports text, image, speech, and video input, outputs text, and has a 1M tokens context window,” and sums the row up as “amongst the leading models in intelligence, but somewhat expensive when comparing to other models of similar price” and “notably fast, however somewhat verbose.” The page’s pricing notes add that “cache write and cache storage are billed separately and vary by provider.” Meta’s pricing page lists no cache-write price, so the $1.92 cache-write line is taken from the snapshot as Artificial Analysis measured it and is not derived here from a list price.

The same page gives the row four ranks: #9 / 214 on intelligence, #27 / 186 on speed, #104 / 113 on cost, and #85 / 115 on verbosity, each computed within the page’s own comparison group rather than across the whole index. This chart ranks every configuration with a measured cost per task, which puts the same row ninth of 116 on the index. The denominators differ because the cohorts differ, and a reader should quote the rank with its cohort.

Muse Wonder against the other Meta rows

The snapshot stores seven configurations from Meta with a measured cost per task. Muse Wonder (Max) is the highest-scoring of them; the next, Muse Spark 1.3 (Max), scores 4.3 points lower at 0.4x the cost. The table measures every other Meta row against Muse Wonder (Max) and marks whether each is a vertex of the cost frontier drawn through Meta rows alone.

Walking the Meta-only frontier from the top, the vertices are Muse Wonder (Max), Muse Spark 1.3 (Max), Muse Spark 1.3 (Xhigh), Muse Spark Open (Xhigh), Muse Glimmer (High), and Llama 4 Maverick. Muse Wonder (Max) is the top vertex, and the step up to it from Muse Spark 1.3 (Max) at 48.1 for $1.60 per task buys 4.3 points for 2.8x the cost. Within Meta’s range on this chart, that is the price of the last 4.3 points. The snapshot dates Muse Spark Open (Xhigh) October 15, 2026, after Muse Wonder (Max), so the Meta range on the chart has already moved since the launch.

Meta’s prices and its benchmark table

Meta’s launch post of October 12, 2026 opens with “Today we’re excited to launch Muse Wonder on Muse Code and Meta Model API at $4 input / $20 output per million tokens (USD).” Its model page prices the API model at $4.00 per million input tokens, $0.20 per million cached input tokens, and $20.00 per million output tokens, with a 1M token context window, and names it muse-wonder-1.0. The same page lists a second model, muse-wonder-full-1.0, with no API prices: Muse Wonder Full is “available exclusively by subscription in Muse Code,” on the Power Usage plan at $50.00/month and the Maximum Usage plan at $200.00/month, which Meta describes as “80x more usage than the Everyday Usage plan.” Meta calls the API model’s performance comparable to the subscription model (“Muse Wonder brings comparable performance to Muse Wonder Full”) and says that “Muse Wonder Full is our most advanced model yet, bringing SOTA coding performance, long-running agents, and advanced computer-use capabilities to tackle developers’ most critical workflows.”

The launch post’s benchmark table has five columns, Muse Wonder Full (max), Muse Wonder (max), Claude Fable 5.1 (max), GPT-6 Astra (max), and Claude Opus 5.5 (max), and its last row is labelled Artificial Analysis Index, reading 56, 52, 53, 53, and 58. Four of those five columns are API models with a max-effort row on this chart: Muse Wonder (Max) at 52.4, Claude Fable 5.1 (Max, Default Fallback) at 53.4, GPT-6 Astra (Max) at 52.7, and Claude Opus 5.5 (Max, Default Fallback) at 57.6. Each rounds to the whole number in Meta’s row, so on the Intelligence Index line Meta’s table and this chart agree. The rest of the table, including Meta’s sentence that “Muse Wonder Full tops coding benchmark leaderboards, including DeepSWE and SWE-Marathon in independent third-party runs, outperforming every other model evaluated,” reports vendor-selected runs that this site does not chart, and the Muse Wonder Full column has no counterpart on any aicharts chart because the model has no public API for Artificial Analysis to run.

No Muse Wonder row on the coding-agent chart

The coding-agent chart is a daily snapshot of the public Artificial Analysis coding-agents comparison. A row there is a model inside a named agent harness at one effort setting, scored on three coding benchmarks (DeepSWE v1.1, Terminal-Bench 4, and SWE-Atlas-QnA) and costed as the mean API bill for one task in that harness.

In the coding-agent snapshot retrieved Oct 5, 2026, 2:10 PM UTC, no row runs Muse Wonder in any harness, so this note has no coding-agent score for the model and does not estimate one. The Meta rows it does store are Muse Code · Muse Spark 1.3 (max) at 54.3 on AA Index for $3.98 per task and Muse Code · Muse Spark 1.3 (xhigh) at 48.3 on AA Index for $3.47 per task. DeepSWE appears in Meta’s table as a vendor-reported row and on the coding-agent chart as one of the three benchmarks Artificial Analysis runs inside a harness; the two are different measurements, and this note carries neither across to the other.

Limits

• The chart score, cost, and token counts are Artificial Analysis measurements of the Muse Wonder row on the retrieval date under Intelligence Index version 4.3.2. They say nothing about other tasks, prompts, or harnesses.
• The rank, the rows that beat it on both axes, the nearest-score table, the component ranks, the input shares, and the Meta-only frontier are computed from the snapshot by aicharts. A new, removed, or rescored configuration moves them, and the snapshot updates on a schedule.
• The cache-write line depends on how much context the harness writes to cache and on a price Meta does not publish on its pricing page. It is reported here as measured, not explained.
• The per-evaluation scores behind the composite are not stored, so no sentence here says which of the 10 evaluations the row does well or badly on.
• Muse Wonder inside Muse Code or another harness is a configuration the coding-agent snapshot does not store, so this note says nothing about it.
• The prices, the subscription plans, the benchmark table, and every sentence about Muse Wonder Full are Meta’s. aicharts did not run Muse Wonder.