---
title: "Meta Muse Wonder Costs Exactly What Claude Opus 5.5 Costs"
description: "Meta's first frontier model costs exactly what Claude Opus 5.5 costs per token. Here is how it prices against GPT-6 Astra and Sol on real workloads, and what the model card says about its benchmarks."
author: "Jahanzaib Ahmed"
date: 2026-10-13
category: "implementation"
readingTime: "9 min read"
tags: ["ai news", "meta", "muse wonder", "llm-pricing", "ai-agents"]
canonical: https://www.jahanzaib.ai/blog/meta-muse-wonder-api-pricing
source: https://www.jahanzaib.ai
---
# Meta Muse Wonder Costs Exactly What Claude Opus 5.5 Costs

Meta released Muse Wonder on October 12, its first model that it benchmarks directly against Claude Opus 5.5 and GPT-6 Astra, and priced it at $4 per million input tokens and $20 per million output tokens on the [Meta Model API](https://dev.meta.ai/docs/pricing-rate-limits). That is the exact rate card of Claude Opus 5.5, down to the $0.20 cached input price. If your agents run on Opus 5.5, you can A/B test Meta Muse Wonder at the same price.

The bigger sibling, Muse Wonder Full, is not on the API at all. It lives only inside subscriptions to Muse Code, Meta's coding agent CLI, and those subscriptions default to a mode where Meta trains on what you send.

## Meta's October 12 release in developer terms

The [Muse Wonder research post](https://research.meta.ai/blog/muse-wonder) introduces two models. Muse Wonder is available today through the Meta Model API and Muse Code. Muse Wonder Full is available to Muse Code subscribers. Meta says Wonder will also power its consumer agent, Muse, starting with paying subscribers, and that both were trained at its Prometheus data center in Ohio. [TechCrunch](https://techcrunch.com/2026/10/12/meta-releases-its-first-frontier-model-muse-wonder/) calls it the company's first attempt at a frontier model since Alexandr Wang restructured Meta's AI efforts.

Meta pitches two capabilities. Active Memory is carrying context about a user across conversations and pulling back the parts relevant to the current task. Active Focus is staying on a hard task, trying other approaches, asking for clarification and checking in before important actions. Both are claims about long running [agentic loops](https://www.jahanzaib.ai/glossary/agentic-loop).

The developer facts, from [Meta's models documentation](https://dev.meta.ai/docs/models):

-   Model ID `muse-wonder-1.0`, marked as in preview.
-   Text, image, video, audio and PDF in; text out.
-   A 1,048,576 token [context window](https://www.jahanzaib.ai/glossary/context-window), with reasoning effort from `minimal` to `max`.
-   Served on the Responses, Chat Completions and Messages endpoints (OpenAI style and Anthropic style request formats). Meta says Wonder and Muse Spark take the same requests on the same endpoints, so moving between them is a model ID change.
-   Rate limits of 3,000 requests and 4,000,000 tokens per minute.

![Diagram of Muse Wonder availability: Meta Model API runs Muse Wonder in Standard mode, while a Muse Code plan adds Wonder Full and defaults to Contributor mode](https://cdn.sanity.io/images/qajb7q5q/production/713d4657f4f627cc5c8ec90b4274bf7d0d48f244-1376x768.png?w=1200&q=75&auto=format&fit=max)

_Only the API path keeps you on Standard mode by default; a Muse Code plan is the sole route to Wonder Full, and it starts in Contributor mode until you switch it off._

## Meta Muse Wonder pricing next to its rivals

Meta's model card compares Wonder with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. Here are their list prices, each read from the vendor's own page this morning ([Anthropic pricing](https://platform.claude.com/docs/en/about-claude/pricing), [GPT-6 Astra](https://developers.openai.com/api/docs/models/gpt-6-astra), [GPT-6.1 Sol](https://developers.openai.com/api/docs/models/gpt-6.1-sol)). I added GPT-6.1 Sol because TechCrunch compared Wonder's price with Sol's.

| Model | Input per 1M | Cached input per 1M | Output per 1M |
| --- | --- | --- | --- |
| Muse Wonder | $4.00 | $0.20 | $20.00 |
| Claude Opus 5.5 | $4.00 | $0.20 | $20.00 |
| Claude Fable 5.1 | $10.00 | $0.25 | $50.00 |
| GPT-6 Astra | $10.00 | $1.00 | $50.00 |
| GPT-6.1 Sol | $2.00 | $0.10 | $10.00 |

So the TechCrunch line that Wonder costs roughly what GPT-6 Sol costs doesn't hold up. Both [GPT-6 Sol](https://developers.openai.com/api/docs/models/gpt-6-sol) and GPT-6.1 Sol list at $2 input and $10 output, half of Wonder. It's also a big step up from Meta's own Muse Spark, which costs $1.25 input and $4.25 output: Wonder is 3.2x Spark on input and 4.7x on output.

Two rules change the totals on long prompts. Meta charges no long context premium on Wonder. OpenAI prices any GPT-6 Astra or Sol prompt over 272K input tokens at 2x input and 1.5x output for the whole request. Anthropic includes the full 1M window at standard rates on Opus 5.5 and Fable 5.1.

## Two real workloads priced on each model

I priced two workloads in code. The first is one long document pass: 600K input tokens, uncached, and 8K tokens out. The second is an agent run of 40 turns, each carrying 110K cached tokens of context plus 10K fresh, and writing 3K tokens. Cache write fees are left out: Anthropic and OpenAI bill cache writes separately, while Meta's price table lists no write charge, so the Opus and GPT rows would rise slightly once you include them.

| Model | 600K token document pass | 40 turn agent run |
| --- | --- | --- |
| Muse Wonder | $2.56 | $4.88 |
| Claude Opus 5.5 | $2.56 | $4.88 |
| Claude Fable 5.1 | $6.40 | $11.10 |
| GPT-6 Astra | $12.60 | $14.40 |
| GPT-6.1 Sol | $2.52 | $2.44 |

On the long document, Wonder costs about a fifth of Astra, because Astra's 272K rule doubles its input rate, though GPT-6.1 Sol still comes in a few cents cheaper even with the same rule applied. On the agent loop, the gap to Astra narrows but stays near 3x, while GPT-6.1 Sol does the same run for half of Wonder. I covered why [prompt caching](https://www.jahanzaib.ai/glossary/prompt-caching) rates decide most agent bills in [my GPT-6 Sol cost breakdown](https://www.jahanzaib.ai/blog/gpt-6-sol-luna-pricing-prompt-caching-agent-costs), and the same logic applies here: Wonder's $0.20 cached rate is what keeps the agent column reasonable. If you want to run your own token mix, the [AI agent cost calculator](https://www.jahanzaib.ai/tools/ai-agent-cost-calculator) does this per model.

## How to read Meta's benchmark claims

TechCrunch reports that Meta puts Wonder and Wonder Full level with Fable 5.1 and GPT-6 Astra, and ahead of all three rivals on MemoryBench. Before you lean on that, read the methodology section of the [Muse Wonder model card](https://research.meta.ai/static/muse-wonder-model-card). Four lines in it limit what the chart can tell you.

-   **Max effort everywhere.** Every model was scored at maximum reasoning effort. Few production agents run at max effort on every call, because it spends more output tokens, and output is the $20 line.
-   **Best of several sources.** Where no comparable third party score exists, Meta uses "the highest of our internally validated run, self-reported score, or official leaderboard score."
-   **Rivals run best effort.** Meta writes that its own runs of competitor models "may not be optimally tuned for proprietary models."
-   **Internal tests.** MemoryBench, FocusBench and the Personal Superintelligence Index are Meta's own benchmarks. The SWE-Marathon row compares models in different agent products, which the card itself says is "not a controlled common-harness comparison."

The numbers may be right. They show a ceiling under favorable conditions, and the card's own limitations list says performance varies with prompts, tools and the agent harness. My read: the memory claims are the interesting part, and they're also the part least checked by anyone outside Meta. Your own eval set decides this, not the chart.

## Wonder Full, training defaults and the preview label

First, Wonder Full is not an API model. The [availability table](https://dev.meta.ai/docs/models) lists it only for Muse Code's Power Usage and Maximum Usage plans, and says that setting reasoning effort to max on Wonder does not select Full. If the headline scores you care about came from Wonder Full, your API calls aren't hitting that model.

Second, check the training setting. On the API, Wonder runs in Standard mode only, which Meta says keeps your prompts, code and responses out of model training. But the [Muse Code subscription docs](https://dev.meta.ai/docs/muse-code/subscriptions) say subscriptions default to Contributor mode, which lets Meta train future models on your content in exchange for more usage. You can turn it off in settings. For client code or anything under an NDA, turn it off before the first session, not after.

Third, the model is in preview, and Meta's [newsroom post](https://about.fb.com/news/2026/10/introducing-muse-wonder-models/) also says a full report with more safety evaluations will be "available soon". I wouldn't put a customer facing workflow on a preview model whose safety report hasn't landed. Meta's model card also recommends human confirmation before irreversible actions in agentic deployments, which I'd treat as a requirement for any agent touching money, email or records. I wrote about the risks of Meta's consumer agent in [my post on Muse and personal agent security](https://www.jahanzaib.ai/blog/meta-muse-personal-ai-agent-security-architecture).

## Who should test Muse Wonder this week

If you run long agent loops on Opus 5.5, you're the obvious tester. The price is identical, so an A/B run costs nothing extra, and Meta's docs list a Messages endpoint alongside Responses and Chat Completions. Meta's [Messages API page](https://dev.meta.ai/docs/protocols/messages) shows the Anthropic SDK pointed at `https://api.meta.ai`, though that page still names only Muse Spark models, so confirm `muse-wonder-1.0` responds there before you plan a migration around it.

If you read very long documents on GPT-6 Astra and want to stay at that capability tier, the 272K rule makes Wonder worth a look on cost alone. If you're on GPT-6.1 Sol for volume work, there's no cost case to move; Wonder would double the bill. And if your stack is built around one vendor's SDK, a router that sends each task to the cheapest model that passes your evals beats switching wholesale. I compared the options in [how to choose an LLM router for AI agents](https://www.jahanzaib.ai/blog/llm-router-ai-agent-routing).

The test I'd run before trusting any new model in an agent: replay 30 to 50 real production traces, score them on the rubric the current model faces, and compare cost per completed task. A model that's cheaper per token but needs two extra tool calls to finish is the more expensive one. More on how I pick vendors in [OpenAI vs Claude for small business agents](https://www.jahanzaib.ai/blog/openai-vs-claude-business-ai-agents).

[Video: Introducing Muse Wonder & Muse Wonder Full: frontier models for agentic tasks](https://www.youtube.com/watch?v=4ZwVP0-ylZU)

_Meta's own two and a half minute launch video, with its researchers describing Active Memory and Active Focus._

## A test plan before you switch

1.  Pull 30 or more real traces from your current agent and replay them on `muse-wonder-1.0` at the reasoning effort you actually use, not max.
2.  Compare cost per completed task, including retries, against your current model.
3.  If anyone on your team signs up for Muse Code, switch off Contributor mode before they open a client repository.
4.  Keep production traffic where it is until Meta publishes the full safety report and drops the preview label.

If you would rather have that evaluation done and the agent built and monitored for you, that is what I do: [see how I build AI agents](https://www.jahanzaib.ai/agents).

## Frequently asked questions

### How much does Meta Muse Wonder cost?

On the Meta Model API, Muse Wonder costs $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens. There is no surcharge for long prompts, and [Meta's pricing page](https://dev.meta.ai/docs/pricing-rate-limits) adds $2.50 per 1,000 web search queries on top of tokens. Those are the same token rates Anthropic charges for Claude Opus 5.5.

### Can I use Muse Wonder Full through the API?

No. Meta's documentation lists Muse Wonder Full as unavailable on the Model API. It is only offered inside Muse Code on the Power Usage and Maximum Usage subscription plans. Setting reasoning effort to max on the standard Muse Wonder model does not switch you to Wonder Full.

### Does Meta train on data sent to Muse Wonder?

On the Model API, Muse Wonder runs only in Standard mode, where Meta says your prompts, code, files and responses are not used for training. Muse Code subscriptions are different: they default to Contributor mode, which allows training in exchange for more usage, and you can switch it off in your settings.

### How big is the Muse Wonder context window?

Meta's documentation lists a 1,048,576 token context window for Muse Wonder, with text, image, video, audio and PDF accepted as input and text as output. Pricing is the same per token whether a request uses a small or a very large share of that window.

### Is Muse Wonder better than Claude Opus 5.5?

Meta's model card shows it competitive with Opus 5.5, Fable 5.1 and GPT-6 Astra, but every score was taken at maximum reasoning effort, some benchmarks are Meta's own, and competitor runs were best effort. Since the prices match, the practical answer is to replay your own agent traces on both and compare cost per finished task.

## Related

- [GPT-6.1 Sol Halves the Cache Rate and Leaves Every Other Price Alone](https://www.jahanzaib.ai/blog/gpt-6-1-sol-pricing-prompt-caching-migration)
- [GPT-6 Sol Costs Half of Opus 5.5 Per Token. On an Agent Loop, It Doesn't.](https://www.jahanzaib.ai/blog/gpt-6-sol-luna-pricing-prompt-caching-agent-costs)
- [Meta's Muse and the Personal AI Agent Security Problem](https://www.jahanzaib.ai/blog/meta-muse-personal-ai-agent-security-architecture)

---

Canonical HTML version: https://www.jahanzaib.ai/blog/meta-muse-wonder-api-pricing