Muse Wonder costs $4 per million input tokens, $20 per million output tokens, and $0.20 per million cached input tokens on the Meta Model API. Reasoning tokens bill at the output rate. There is no long-context premium across the 1M-token window, no batch tier, and no free credits. Web search adds $2.50 per 1,000 queries. Muse Wonder Full has no API price at all; it comes bundled with Muse Code subscriptions at $50 or $200 a month.
Those are the list prices. This post works out what they mean for real requests, where the cost hides, how Wonder compares with the models priced around it, and what the Muse Code and Muse app subscriptions buy. For the model itself, read the Muse Wonder explainer; for making the calls, the API guide.
Muse Wonder API pricing table
| Item | Price |
|---|---|
| Input | $4.00 per 1M tokens |
| Cached input | $0.20 per 1M tokens (95% off) |
| Output, including reasoning tokens | $20.00 per 1M tokens |
| Web search tool | $2.50 per 1,000 queries, plus tokens |
| Long-context premium | None; same rate up to 1,048,576 tokens |
| Batch discount | None listed |
| Zero data retention | Same price as Standard, approval required |
| Free tier or credits | None |
Meta’s own steering context is not billed and does not appear in usage. Cached tokens show up in input_tokens_details.cached_tokens; reasoning tokens in output_tokens_details.reasoning_tokens. Both come back on every response, which is how you audit a bill.
For scale, Muse Spark 1.3, the previous Meta model, is $1.25 / $4.25 / $0.15. Wonder is 3.2x the input price and 4.7x the output price for a four-point gain on the Artificial Analysis index (48 to 52).
Worked examples
Three request shapes, priced at list with no caching:
| Request | Input tokens | Output tokens | Cost |
|---|---|---|---|
Short answer, low effort |
2,000 | 800 (incl. 300 reasoning) | $0.024 |
Document review, medium effort, 60-page lease |
45,000 | 6,000 (incl. 3,500 reasoning) | $0.30 |
Agent run, high effort, 40 tool calls |
400,000 cumulative | 30,000 (incl. 18,000 reasoning) | $2.20 |
The third row is where the model lives, and where the bill grows. An agentic loop resends the growing conversation on every turn, so a 40-step run can push 400,000 input tokens through the meter even though no single prompt is large. Caching turns most of that into $0.20 tokens if your prefix is stable; without it, the same run at $4 is the $1.60 input line above.
What a task costs in practice: the Artificial Analysis numbers
List price per token and price per task are different things. Artificial Analysis ran its Intelligence Index on every frontier model at maximum reasoning effort and reported the cost per task:
| Model | List price (in / out) | Cost per index task |
|---|---|---|
| GPT-6.1 Sol | $2 / $10 | $0.72 |
| Gemini 4 Argon | $2 / $10 | $1.99 |
| GPT-6 Astra | $10 / $50 | $3.26 |
| Muse Wonder | $4 / $20 | $4.52 |
| Claude Sonnet 5.5 | $2 / $10 | $5.46 |
| Claude Opus 5.5 | $4 / $20 | $5.98 |
| Claude Fable 5.1 | $10 / $50 | $7.63 |
Wonder is cheaper per task than both Claude models and more expensive than GPT-6 Astra, which lists at 2.5x Wonder’s rate. The reason is token volume. Wonder consumed 7.69 million input tokens across the index, two to seven times more than the OpenAI and Google models, and about 54,000 output tokens per task. Its $4.52 broke down as $3.43 input (of which $1.92 cache writes, $1.44 cache hits, $0.07 uncached) and $1.08 output.
Two lessons. Wonder leans hard on the cache, so a workload that defeats caching will cost well above $4.52 per comparable task. And Astra’s $10 input rate buys fewer, terser tokens; cheap tokens and cheap tasks are not the same thing. Measure your own tasks before you switch.
Reasoning effort is a price dial
Every effort level above minimal adds reasoning tokens at $20 per million. Meta’s guidance is to use the lowest level that gives acceptable results, and none is rejected with a 400. On the example lease review above, moving from medium to max could plausibly triple the reasoning share and take the request from $0.30 to around $0.60, with no change to the input. Log reasoning_tokens by request type for a week and you will know your real per-effort rates.
Muse Code plans: where Muse Wonder Full lives
Muse Code is Meta’s terminal coding agent. It can be billed per token, but most people use a subscription:
| Plan | Price | Usage vs Everyday | Models |
|---|---|---|---|
| Everyday Usage | $5 / month | 1x (10 to 50 prompts per 5 hours) | Muse Spark, Muse Wonder, Muse Image |
| High Usage | $15 / month | 5x | Same |
| Power Usage | $50 / month | 20x | Adds Muse Wonder Full, early features, larger uploads |
| Maximum Usage | $200 / month | 80x | Same as Power |
Two things to know before subscribing. Subscriptions default to Contributor mode, which lets Meta use your prompts, code and outputs to train future models in exchange for more usage; you can switch it off and future content stops being used. And the $5 plan is the cheapest way to touch Muse Wonder with a fixed monthly cost, since the API has no free tier. Muse Wonder Full has no per-token price anywhere; the $50 Power plan is its floor.
Consumer Muse app
The Muse personal agent app is free to use, but Muse Wonder is rolling out to its subscribers first, on $20 and $100 monthly tiers. Meta has not said when or whether free users get it. If you only want to try the model, the API at a few cents per call is cheaper than either subscription.
How Muse Wonder compares on price
| Model | Input | Output | Cached input | Context |
|---|---|---|---|---|
| Muse Wonder | $4 | $20 | $0.20 | 1M |
| Claude Opus 5.5 | $4 | $20 | $0.20 cache read | 1M |
| Claude Sonnet 5.5 | $2 | $10 | $0.10 | 1M |
| Gemini 4 Argon | $2 / $4 after intro | $10 / $20 | varies | 1M |
| GPT-6 Astra | $10 | $50 | $1 (90% off) | 1M |
| Claude Fable 5.1 | $10 | $50 | $1 | 1M |
| Muse Spark 1.3 | $1.25 | $4.25 | $0.15 | 1M |
Wonder sits exactly on Opus 5.5’s rate card, undercuts Astra and Fable 5.1 by 60%, and costs twice Sonnet 5.5 and Argon. On the independent index it scores 52, below Sonnet 5.5’s 56 at half the price. The case for Wonder at this price is the agentic work Meta’s benchmarks highlight, long coding runs and legal tasks, not general intelligence per dollar. The Opus 5.5 head-to-head goes deeper on the same-price comparison.
Track cost per request in Apidog
Pricing pages tell you rates; your usage blocks tell you costs. In Apidog, set up the Responses request with MODEL_API_KEY as a secret variable, then add a post-processor script that reads usage and computes the dollar figure:
const u = pm.response.json().usage;
const cached = u.input_tokens_details?.cached_tokens ?? 0;
const cost = ((u.input_tokens - cached) * 4 + cached * 0.2 + u.output_tokens * 20) / 1e6;
pm.environment.set("last_cost_usd", cost.toFixed(5));
pm.test("cache is working", () => pm.expect(cached).to.be.above(0));
Run the same prompt at low, medium and max effort as three requests, and the environment variable gives you the per-effort cost on your own task. Put the scenario in CI with the Apidog CLI and a price change on Meta’s side shows up as a failing assertion, not a surprise invoice. Download Apidog to set it up.
Frequently asked questions
Is there a batch API discount? Meta’s pricing page lists none for Muse Wonder. Background responses (background: true) are a convenience for long jobs, not a discounted tier.
Does the 1M context cost extra? No. Meta charges the same per-token rate regardless of how much of the window a request uses.
How much does Muse Wonder Full cost per token? Nothing is published. It is bundled with Muse Code Power ($50) and Maximum ($200) and not sold through the API.
Is Muse Wonder cheaper than Claude Opus 5.5? Identical list price. Per task on the Artificial Analysis index, Wonder came in at $4.52 against $5.98 for Opus 5.5, because it used fewer expensive output tokens.
Next step
Price one real task three ways: your current model, Muse Wonder at medium, and Muse Wonder at max. Use the script above so the numbers come from usage, not from a rate card.
