Muse Wonder $4/$20 per million tokens vs Sume's agent rate table

Meta's Muse Wonder (muse-wonder-1.0) costs $4.00 per million input tokens, $0.20 per million cached input tokens and $20.00 per million output tokens (read 2026-10-13). Those are the same three numbers Sume's agent rate table holds for Claude Opus 5.5, so a turn of 100,000 input and 5,000 output tokens costs $0.50 on either, and $0.196 when 80,000 of the input is a cache read.

Muse Wonder is not a row in Sume's agent rate table, and an Agent Completion accepts only model: "sume-agent", so this page is a price comparison, not a switch you can flip. The arithmetic is useful anyway: it tells you what a long agent turn costs before you decide how large a generation cap to set.

What does Meta list for Muse Wonder?

Meta's pricing page lists input at $4.00, cached input at $0.20 and output at $20.00 per million tokens. It adds web search at $2.50 per 1,000 queries on top of tokens, and rate limits of 3,000 requests and 4,000,000 tokens per minute. The page did not state a context window, so this post does not either.

• Per search query: $2.50 / 1,000 = $0.0025.
• Per 100,000 input tokens uncached: 100,000 x $4 / 1,000,000 = $0.40.
• Per 5,000 output tokens: 5,000 x $20 / 1,000,000 = $0.10.

How does one turn price against Sume's rate-table rows?

The table uses the per-million rates in Sume's provider-pricing package for each row (read in the repo on 2026-10-13). Scenario A is 100,000 uncached input tokens and 5,000 output tokens. Scenario B is 80,000 cache-read tokens, 20,000 fresh input tokens and 5,000 output tokens. Cache-read rates were only compared where the repo lists one.

What does that add up to over a day of turns?

Multiply by volume. A thousand scenario A turns is 1,000 x $0.50 = $500 on Muse Wonder or Opus 5.5, $250 on Sonnet 5.5, $146.25 on Muse Spark 1.3 and $12.50 on GPT-6 Luna. With the cache in scenario B, the first two fall to $196 and Sonnet 5.5 to $106. A thousand search queries add $2.50 on Muse Wonder.

Where does the Sume spend cap fit in?

Sume's Agent Completions page says generation_spend_cap_credits is required, has no default, and is the most credits that generation can use on the run; a missing value returns 400 invalid_request. Size it from the generation steps on the API rate card, as the page instructs, and treat the token figures above as a separate estimate of the language-model turn, since the cap is documented as a generation limit.

If you only want a cheap model for tool-driving turns, the table shows the spread is 40x between the top and bottom rows, which matters more than the 0.196-versus-0.50 cache effect on a single row.