Muse Wonder vs Claude Opus 5.5 vs Gemini 4 Argon

Muse Wonder and Claude Opus 5.5 cost the same per token: $4 per million input tokens and $20 per million output (OpenRouter model list, read 2026-10-13). Gemini 4 Argon costs half that until 2026-11-30, then the same. All three hold about a million tokens of context.

Three frontier models went on sale within weeks of each other, and the "muse wonder vs opus 5.5" searches started the day Meta launched. Most answers so far repeat vendor benchmark claims. This post sticks to what two sources agree on: the price of each token, the context window, who can call each model and since when, and what 1,000 agent runs cost on each. Where we could not confirm a claim twice, we say so and leave it out.

Price per million tokens, side by side

Each vendor's own page and the OpenRouter list give the same input and output prices.

The cached input column is the price of a cache hit. Anthropic also bills cache writes, and the three vendors cache in different ways, so compare cached costs on your own traffic before you rely on them.

Prices read from the OpenRouter model list through looot on 2026-10-13.

Anthropic's pricing page says Opus 5.5 gets its full 1M token context at the standard rate. Google describes Argon's limit as 1M tokens shared by input and output. Meta's pricing page states no context window, so the Muse Wonder figure comes from OpenRouter's list and VentureBeat's launch report.

Muse Wonder cost for 1,000 agent runs, against the other two

Take one agent run as 20,000 input tokens and 2,000 output tokens. That is an example size, not a measured average. Here is the arithmetic for 1,000 runs from the tables above.

1,000 runs = 20 million input tokens + 2 million output tokens

Muse Wonder:                 20 x 4.00 + 2 x 20.00 = 80 + 40 = 120 USD
Claude Opus 5.5:             20 x 4.00 + 2 x 20.00 = 80 + 40 = 120 USD
Gemini 4 Argon, to 11-30:    20 x 2.00 + 2 x 10.00 = 40 + 20 =  60 USD
Gemini 4 Argon, from 12-01:  20 x 4.00 + 2 x 20.00 = 80 + 40 = 120 USD

With 15,000 of the 20,000 input tokens read from cache:
Muse Wonder or Opus 5.5:     15 x 0.20 + 5 x 4.00 + 2 x 20.00 = 3 + 20 + 40 = 63 USD
Gemini 4 Argon, to 11-30:    15 x 0.10 + 5 x 2.00 + 2 x 10.00 = 1.50 + 10 + 20 = 31.50 USD

On list price, the choice between Muse Wonder and Opus 5.5 is not about money. Argon is the cheapest until November 30, and a project that runs into December should budget at the higher rate. Cache hits cut the bill about in half in this example for all three, which matters more than the model you pick if your agent resends the same instructions every turn.

The per-token bill is not the whole bill. Each model's built-in web search has its own price, covered for Meta in Muse Wonder API pricing and for Google in the Gemini 4 Argon MCP guide.

Access: who can call each model today

Only Gemini 4 Argon and Muse Wonder are this week's launches. Opus 5.5 was already on sale.

Meta lists no free tier and no Contributor discount for Muse Wonder; the cheaper Contributor mode on its pricing page covers only Muse Spark. What Muse Wonder Full is and why it is not on the API is in what is Meta Muse Wonder.

Muse Wonder benchmarks: what we could not confirm

All three vendors publish benchmark tables, and each table is the vendor's own selection. Meta says Muse Wonder scores comparably to other frontier models, and Google says Argon leads several boards. We found no independent head-to-head of all three models that we could confirm on two sources, so this post repeats no benchmark number. Run the same 20 tasks from your own work on each model and compare the answers and the bills. That test costs a few dollars and tells you more than a leaderboard.

Built-in search or a data tool

Each of these models can search the web through its vendor, at the vendor's price. An agent often needs more than search: a page read as markdown, a company record, a set of Google results with positions. looot gives the agent one key for those data tools, whichever of the three models it runs on, and shows the price of each call before it runs. The live price of a web search through looot, read from the catalog when this page loads:

The public catalog lists the other data tools. looot does not run these three models: its chat tool, `openrouter-chat`, accepts 15 listed models and none of the three is among them. It does read the OpenRouter model list for free through `openrouter-models`, so your agent can check today's price of each model before a long job. Every launch this week is priced on the model price page, and Meta's model has its own Muse Wonder page.

To try it from a terminal first:

npx looot init

What this comparison does not cover

• Quality. We ran no benchmark and repeat none.
• Speed. None of the three vendors' pages we read states a latency figure we could check.
• Rate limits for Opus 5.5 and Argon. They depend on your account tier with each vendor.
• Prices outside the list. Enterprise contracts and regional pricing can differ.

Last checked 2026-10-14. Vendor pricing pages were read on 2026-10-14 and the OpenRouter list on 2026-10-13.