Pick the cheaper Muse Wonder bill with $4/$20 math

Start with the task you actually ship

You have a long coding task sitting in your editor: a multi-file refactor, a failing test suite, or a postponed migration. A short prompt can make a model look cheap, but a long agent session can flip the bill.

Pick that task, not a toy prompt. Write down the files, the expected result, and the hours you would spend if the agent stalled. That time is part of the bill. Note whether it needs web search, touches many files, or runs long. Measure your own workload, not a benchmark. Use a task with a clear finish line, because an open-ended agent run can hide the cost. Done looks like a short brief: the task, the files, the acceptance check, and the prompt you would actually paste.

Count the tokens from one real task

• Run the task through Muse Wonder on the Meta Model API and end with a token log: input, output, cached input, prompt, and model settings.
• Count the web-search queries, because the API charges $2.50 for every 1,000 queries. Keep a search count you can turn into a dollar line.
• Check whether the session stays inside the API limits: 3,000 requests and 4,000,000 tokens per minute, with a 1,000,000-token context window. Note whether the task fit in a single context or needed to be split.
• Save the numbers in a small table: input, output, cached input, searches, and total time. Reuse that row next week.

Running a short prompt and extrapolating is the usual mistake. Long agent work changes the mix. Input tokens can pile up while output stays modest, and cached input can change the bill. Count retries too. A failed attempt still consumes tokens, and the bill does not care whether the result was useful. If the task needs a new context window, note the extra input tokens from re-reading files.

Turn the token log into a dollar bill

Muse Wonder's bill uses $4 per million input tokens and $20 per million output tokens on Muse Code and the Meta Model API, $0.20 per million cached input tokens through the API, and $2.50 per 1,000 web-search queries through the API. Multiply your token log by those rates. When your client does not report cached input, leave that line out rather than guessing. The bill is only as good as the token log behind it.

For example, a task that uses 1,000,000 input tokens and 100,000 output tokens has a model bill of $4 plus $2, or $6 before cached input and searches. Treat that $6 as a floor, not a full bill. Add the cached-input and search lines when your log has them.

Do not compare that number to a leaderboard rank. The average cost of an Intelligence Index benchmark task is $4.52 for Muse Wonder, $5.98 for Claude Opus 5.5, and $3.26 for GPT-6 Astra. The SWE-Marathon result for Muse Wonder Full is sourced from Abundant AI and was not run under a single controlled harness for all models, so leaderboard rank is not a universal proof. Per-token price and per-task cost can diverge. A model swap driven by leaderboard rank can move your bill in the wrong direction. Your own task is the better judge.

Let the flat plan lose unless your week is full of long sessions

Muse Wonder Full is reserved for Muse Code subscribers and is unavailable through the Meta Model API, so long-session access depends on a Muse Code plan. It is accessible through Muse Code's Power Usage and Maximum Usage plans; the Maximum Usage plan has a limit 80 times the Everyday Usage plan and uses five-hour and weekly resets.

Put your one-task bill next to the monthly plan price you would pay. The flat plan is a commitment to a usage pattern you have not measured. If the per-token bill for your real task is far below the flat plan, stay on the API. When your week is full of long sessions, the flat plan may be the cheaper monthly cost. Skip Maximum Usage when it only sounds bigger. Buy it only when your task log shows repeated long multi-file sessions that would exhaust a smaller tier. Avoid guessing from a headline when you cannot see the plan price in your account.

The durable trick is one task, one bill, one decision. Choose the plan that leaves more margin in your week.