Muse Code Cost Calculator

Every rate is quoted per million tokens, so the formula behind this Muse Code cost calculator is short:

cost = (fresh_input / 1,000,000 × input_rate)
     + (cached_input / 1,000,000 × cache_rate)
     + (output / 1,000,000 × output_rate)

The ratio surprises people arriving from a chat assistant. In chat, output dominates the bill; inside an agent loop it inverts. Muse Code re-sends its working context — file contents, tool results, the running plan — every turn, so fresh input tokens outnumber generated ones many times over. Cached input therefore moves Muse Code cost more than the headline output rate.

Where these rates come from

Rates here are as reported by pricing write-ups and cross-checked on 6 August 2026 against Meta's Muse Spark pricing and rate limits page. Both tiers, their rate limits and the training clause are explained on Muse Code pricing; this page assumes a tier is chosen and only does arithmetic.

The reference turn used below

Worked examples use one Muse Code turn of 60,000 input and 3,000 output tokens — a mid-sized step where Muse Code reads several files, runs a command and writes a patch. Your turns run smaller on a tidy repository and much larger on a sprawling one.

Two spreads drive every Muse Code cost estimate: caching cuts Standard input roughly 8x, and Contributor cuts input 12.5x, output 21.25x and cached input 75x against Standard. The cheaper tier is paid for with a training grant over your prompts and completions — read Muse Code privacy before chasing the bottom row.

These multiply the reference turn; they are not invoices. Run your own turn sizes through the Muse Code cost calculator above. Real months land between the cold and cached columns, because a first read is always cold and later turns often are not.

Meta describes a Muse Code run that optimised GPU kernels on NVIDIA Hopper across more than 1,000 tool calls and as much as 24 hours. Price 1,000 reference turns and the risk takes a shape:

• Standard, cold context — $87.75. One unattended overnight session.
• Standard, cached input — $21.75. Same work, context mostly re-read from cache.
• Contributor, cold context — $6.60. The cheap tier, paid for with a training grant.

That first line explains why the loudest complaint about Muse Code cost is not the rate but the missing ceiling. A commenter on the Hacker News launch thread described being "one oopsie away from a $100 bill", and a thousand-call Muse Code run at Standard rates is exactly that arithmetic — on the flagship demo, not an abnormal run.

Treat it as an illustration built from the reference turn, not a published invoice — Meta has not disclosed token counts for that case study. It is still the right order of magnitude to budget against.

1. Shrink the context before the loop starts

   Fresh input is the biggest line in nearly every turn, so the cheapest optimisation is a smaller working set: point Muse Code at a subdirectory, not a monorepo root. Drafting with /plan, then attacking that draft with /grill, costs two turns and saves dozens.

2. Keep sessions warm so input bills at the cache rate

   Cached input costs $0.15 per million on Standard against $1.25 fresh, which is where an 8.8¢ turn becomes a 2.2¢ turn. Batch related work into one long Muse Code session rather than reopening Muse Code per errand.

3. Cap the card, because the platform will not

   Muse Code billing sends email alerts but enforces no limit, so the only real ceiling is external: a virtual card carrying its own monthly limit, as developers in the launch discussion describe, or an OpenRouter key with a per-key spending limit — setup notes sit on Muse Spark API access.

4. Choose the tier per repository, not once

   Contributor cuts Muse Code cost roughly 13x on a cold turn in exchange for training rights over prompts and completions — fine for a throwaway side project, disqualifying for client code. The clause is read line by line on Muse Code privacy.

Rival prices are as reported for August 2026. Below roughly 10 Muse Code turns a working day, token billing beats any $20 seat; above about 40 warm turns a day a flat seat wins unless you sit on Contributor. Side-by-side: Muse Code vs Claude Code, Muse Code vs Codex, Muse Code vs Cursor and Muse Code vs GitHub Copilot.

• Muse Code PricingBoth tiers, rate limits and data terms.
• Muse Spark 1.2Context window, throughput and benchmarks.
• Muse Spark APIPer-key spending limits via OpenRouter.
• Muse Code PrivacyWhat the Contributor discount costs.
• Best AI Coding AgentsThe field ranked by price.
• Muse Code Guide HomeInstall steps and comparisons.