Meta Muse Wonder: Benchmarks, Price and Best Uses (2026)

Meta Muse Wonder is the first Meta model priced like a frontier model, and on Meta’s own scorecard it trades wins with Claude Opus 5.5 rather than pulling clear. It costs $4 per million input tokens and $20 per million output tokens on the Meta Model API, the same list price as Opus 5.5. In Meta’s 14-row benchmark table, its higher-capability sibling Muse Wonder Full posts the top score in 6 rows and Muse Wonder takes one more, MemoryBench. Claude Opus 5.5 still tops 6 rows and GPT-6 Astra one.

The independent view is less flattering. On the Artificial Analysis Intelligence Index, Muse Wonder Full scores 56: below Claude Opus 5.5 at 58 and level with Claude Sonnet 5.5 at 56, a model Meta’s own chart leaves out. Muse Wonder, the version you can buy by the token, scores 52: level with OpenAI’s GPT-6.1 Sol and below Claude Sonnet 5.5, both of which cost half as much.

Meta introduced Muse Wonder and Muse Wonder Full on 12 October 2026 as its most capable models, built for personal and professional agents: long coding projects, legal and financial documents, health questions and tasks that run for hours. Muse Wonder is on the Meta Model API and in Muse Code from launch day. Muse Wonder Full, the higher-capability variant, comes only with the $50 Power Usage and $200 Maximum Usage plans of Muse Code, Meta’s command-line coding tool. Meta says Muse Wonder will reach its Muse app later, subscribers first, but by 10 November 2026 it had not named a date; it has said nothing about Full there.

Below you will find Meta’s full benchmark table next to the first independent data from Artificial Analysis and Vals AI, what a task really costs, the data setting every finance team should change first, and which model to pick for which job.

Meta Muse Wonder: the short answer

Both new Meta models against Claude Opus 5.5, which costs the same, and GPT-6 Astra from OpenAI:

Top Meta model

Muse Wonder Full

Meta

API price per 1M (in / out)Plans only

AA Intelligence Index56

Vals Finance Agent V259.1

DeepSWE 1.179.0

OSWorld V2.1 (partial)78.4

On the API

Muse Wonder

Meta

API price per 1M (in / out)$4 / $20

AA Intelligence Index52

Vals Finance Agent V256.6

DeepSWE 1.178.8

OSWorld V2.1 (partial)72.6

Same price

Claude Opus 5.5

Anthropic

API price per 1M (in / out)$4 / $20

AA Intelligence Index58

Vals Finance Agent V258.6

DeepSWE 1.174.2

OSWorld V2.1 (partial)81.8

GPT-6 Astra

OpenAI

API price per 1M (in / out)$10 / $50

AA Intelligence Index53

Vals Finance Agent V253.5

DeepSWE 1.174.1

OSWorld V2.1 (partial)75.8

Scores at max reasoning effort from Meta’s table, except the Artificial Analysis (AA) Intelligence Index, which is independent. Teal marks the best score in each row. Muse Wonder Full has no per-token price; it comes with Muse Code plans from $50 a month. Prices are per million input and output tokens.

The short version: with Meta Muse Wonder, Meta now has a real frontier model, priced exactly like Opus 5.5. By Meta’s numbers Muse Wonder Full is strongest on legal work and very long coding jobs, level on finance research, and behind Anthropic on computer use, terminal coding and the independent index.

Should you switch? If you already pay for Claude Opus 5.5, the token prices are identical; Muse Wonder cost less per task in the first Vals and Artificial Analysis runs but scored lower, so switch only where your own tests show a gain, starting with legal and document work. If you pay for GPT-6 Astra or Claude Fable 5.1, Muse Wonder costs 60% less per token and is worth a trial, alongside Claude Sonnet 5.5, which scores higher on the independent index at half Muse Wonder’s price.

If your team uses AI in a chat app rather than through code, there is nothing to switch to yet. And if you try Muse Code, turn off Contributor mode before your first prompt.

Prefer to watch? Here is Meta Muse Wonder in 50 seconds: the two models, the price, where it leads, where Claude Opus 5.5 still wins, and the fine print.

Muse Wonder vs Muse Wonder Full: who gets which

Meta ships two variants under one name, and they reach you in different ways:

One detail trips people up. Setting reasoning effort to max on Muse Wonder does not switch you to Full; Meta’s models docs say so plainly. They are separate models, and the only door to Full is a Muse Code subscription, used through the Muse Code command line tool.

In practice, developers building their own agents get Muse Wonder through the API, while people who want Meta’s best model for their own coding and agent work pay for Muse Code. Nobody can call Muse Wonder Full from their own app.

Meta Muse Wonder benchmarks: Meta’s full table

Meta tested both models at max reasoning effort against three rivals, also at max: Claude Fable 5.1 and GPT-6 Astra, which both list at $10 and $50 per million tokens, and Claude Opus 5.5. Here are all 14 rows, with what each test measures and where Meta’s model card says each result came from:

Where Muse Wonder Full leads

Muse Wonder Full’s leads on Harvey LAB-AA (13.1 against 8.6 for GPT-6 Astra), Vals Finance Agent V2 (59.1 against 58.9 for Fable 5.1), AutomationBench-AA (45.4 against 41.6), DeepSWE (79.0 against 74.2) and SWE-Marathon (57.5 against 46.3, with Full at xhigh) all rely on outside sources. Its sixth win, the Personal Superintelligence Index (59.1), is Meta’s own composite, and Muse Wonder adds MemoryBench at 85.6.

The legal lead is the most striking: 13.1 is more than three times Opus 5.5’s 4.2. Keep the scale in mind, though. Harvey LAB-AA counts a task only when the work product meets every criterion and passes the hallucination checks, so even the leader fully completes about one task in eight. The finance lead, by contrast, is 0.2 points, a tie, and Vals’ public board ranks Wonder Full fifth at 59.06.

Where Claude Opus 5.5 still wins

Opus 5.5 also tops six rows: HealthBench Pro (65.6 against 62.3), JobBench (75.4 against 75.3, a tie in practice), OSWorld partial credit (81.8 against 78.4), Terminal-Bench 4.0 (59.6 against 57.1), ProgramBench (70.5 against 59.0) and the Artificial Analysis Index (58 against 56). ProgramBench is the widest gap: rebuilding a whole program from scratch is still clearly Anthropic’s ground.

GPT-6 Astra takes GDP.pdf, 34.2 against 33.2. Claude Fable 5.1 has the best all-or-nothing OSWorld score in the table, 43.1. Meta left that cell unshaded, and its table shows no Opus figure for that measure.

The fine print in Meta’s model card

• Score selection. Where a comparable third-party or provider-reported result exists, Meta uses it. Otherwise it uses the highest of its own validated run, the self-reported score or the official leaderboard score.
• Rival runs are best effort. Meta says its own runs of rival models use common settings where practical, but prompts, tools and runtime may not be optimally tuned for them.
• Refusals score zero and stay in the count.
• SWE-Marathon is not a controlled comparison. GPT-6 Astra’s result comes from Codex and Fable 5.1’s from Claude Code, and Muse Wonder Full ran at xhigh while every other cell is at max.
• External coding results keep each vendor’s harness, so the DeepSWE row mixes Datacurve, OpenAI and Anthropic runs.
• Almost everything is at max effort, the most expensive setting; the one exception Meta marks is Full on SWE-Marathon at xhigh. Meta’s table has no results at the lower efforts many teams will run day to day.
• Results can move with the setup. The model card itself warns that performance varies with serving configuration, prompts, tools and the agent harness.

How far apart are Muse Wonder and Wonder Full?

Meta’s developer blog says Muse Wonder brings comparable performance to Full, and Mark Zuckerberg’s launch post called Wonder the fast frontier model and Full the full model. The benchmark gaps tell a more mixed story:

They are level on DeepSWE (79.0 against 78.8) and MemoryBench (85.5 against 85.6), but Full is 6 to 9 points ahead on the agent tests that reward persistence: FocusBench, ProgramBench and AutomationBench-AA. For finance and document work the gap is small: 2 to 3 points on Vals Finance Agent V2, GDP.pdf and Harvey LAB-AA. The API model is the sensible place to start; Full’s bigger lead shows up on long agent and coding runs.

Meta Muse Wonder vs independent results

Most rival figures in Meta’s table match the public leaderboards and the rivals’ own reports. A few differ, several strong rivals are missing, and some of Meta’s top results for its own models are not on public boards yet.

What matches

• Artificial Analysis: the AutomationBench-AA, Terminal-Bench 4.0 and Intelligence Index figures, and the rival scores on Harvey LAB-AA, all match AA’s live leaderboard. AA also confirms Muse Wonder Full’s own 45.4 on AutomationBench-AA‘s strict measure (AA’s headline partial-credit score is 71.3%) and 57.1 on Terminal-Bench 4.0.
• Vals AI: Vals’ public board, updated on 12 October, lists Wonder Full at 59.06 and Muse Wonder at 56.64, matching Meta’s 59.1 and 56.6. The rival scores match too: Fable 5.1 58.88, Opus 5.5 58.59 and GPT-6 Astra 53.54.
• DeepSWE: the rival scores match what the vendors report, including Opus 5.5 at 74.2 and Fable 5.1 at 67.4.
• HealthBench Pro: Opus 5.5 at 65.6 and Fable 5.1 at 62.1 match Anthropic’s system cards.

Where other published figures differ from Meta’s table

Smaller differences sit in other rows. On Terminal-Bench 4.0 Meta uses AA’s runs, while the vendors report Opus 5.5 at 66.4 at xhigh (64.8 at max) from Anthropic and Astra at 57.9 from OpenAI, and OpenAI’s Astra page lists Fable 5.1 at 55.8. OpenAI puts Astra’s DeepSWE 74.1 at xhigh, not max, where it scored 73.2. Anthropic’s re-evaluation gives Fable 5.1 42.8 on OSWorld binary, against Meta’s 43.1. ProgramBench and Harvey LAB come in different metrics and versions across sources, so the vendors’ own figures cannot be compared with Meta’s.

None of these overturns the headline, and they cut both ways. The 48.7% in the Claude Opus 5.5 system card would hand Opus 5.5 the OSWorld binary row too, while AA’s version of GDP.pdf puts Muse Wonder Full first on a row Meta gave to GPT-6 Astra. Anthropic’s 70.3 for Astra on HealthBench Pro comes from the Claude Sonnet 5.5 system card, and OpenAI’s own figures come from its GPT-6 Astra launch page and its GPT-6.1 Sol post.

The rivals Meta left out

Meta’s chart ranks Muse Wonder Full second behind Opus 5.5. AA’s own leaderboard also lists Claude Sonnet 5.5 at 56.00, a hair above Muse Wonder Full’s unrounded 55.80, so by model family Full is third, not second. Sonnet 5.5 is not the only rival missing from Meta’s comparisons:

Two names stand out. Claude Sonnet 5.5 matches or beats Muse Wonder Full on five of these seven tests at half Muse Wonder’s per-token price, $2 and $10, though it trails on GDPval-AA (1838 against 1857) and on Vals (58.10 against 59.06). Gemini 4 Argon leads both the finance research and the business workflow tests, but Google has only opened it to trusted cyber defenders so far, which may be why Meta left it out. Even Meta’s own older Muse Spark 1.2 (60.60) and Muse Spark 1.3 Max (59.96) rank above Muse Wonder Full on the Vals board.

Still waiting for public leaderboards

• Harvey LAB-AA, 13.1 and 10.0: AA’s public Harvey page has no entry for the new models yet. Its leader there is Grok 4.7 at 9.4%.
• DeepSWE 1.1, 79.0: Meta credits Datacurve, but Datacurve’s public board, last updated on 22 September, has no entry for either new model.
• SWE-Marathon, 57.5: the benchmark’s site says its new leaderboard is coming soon.
• MemoryBench, FocusBench and the Personal Superintelligence Index: Meta’s own internal tests.

So far the confirmed independent numbers come from two places. Artificial Analysis confirms the Intelligence Index, AutomationBench-AA and Terminal-Bench 4.0, plus its own GDP.pdf run (32.6, first place), which differs from the Surge figure Meta used. Vals AI confirms Finance Agent V2: 59.06 for Wonder Full and 56.64 for Muse Wonder. AA has published no speed or latency data yet, and the model has no entry on Arena. We will add both models to our AI leaderboard, which ranks models on independent tests only, as those results arrive.

Active Memory and Active Focus, in plain words

Meta built the launch around two abilities it says personal and work agents need.

Active Memory: remembering what matters

Active Memory means carrying important context about you across conversations and pulling back the right details when they become relevant. Meta measures it with MemoryBench, an internal test of recall across long conversations of 64,000 to 100,000 tokens, with distractor conversations mixed in and later turns that contradict or update the facts. An AI judge marks each answer.

Muse Wonder scores 85.6 and Muse Wonder Full 85.5, against 81.5 for Opus 5.5, 81.4 for GPT-6 Astra and 79.0 for Fable 5.1. For a finance team, the update part is the useful bit: when a client corrects a figure three messages later, the model should use the corrected figure, not the first one.

Active Focus: keeping at a hard task

Active Focus is Meta’s name for persistence: trying other approaches, working through blockers and adapting when things change, while asking for clarification when needed and checking in before important actions. Meta measures it with FocusBench, an internal test of personal agent tasks in areas such as housing, travel, IT and finance, where the agent has to work through real-world blockers over several interactions.

FocusBench is where the two variants differ most: Full scores 54.2 and Muse Wonder 44.9, behind both GPT-6 Astra (47.7) and Opus 5.5 (45.8), with Fable 5.1 at 36.6. If persistence on messy, multi-step tasks is what you need, that gap matters. Meta built and ran both tests, so they show what it optimised for.

Meta Muse Wonder pricing and what a task costs

Muse Wonder costs $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens, in Standard mode only. That is 3.2 times Muse Spark 1.3 on input and about 4.7 times on output, a clear jump into frontier pricing. The full rates are on Meta’s pricing page.

Meta’s AI chief Alexandr Wang told CNBC that API pricing would be meaningfully below the leading OpenAI and Anthropic models. That holds against GPT-6 Astra and Claude Fable 5.1, which list at $10 and $50, but not against Claude Opus 5.5, which lists at exactly the same $4 and $20 on Anthropic’s pricing page.

The fine print on the price

• No long-context premium. The rate is the same whether a prompt is short or fills most of the 1,048,576-token window. GPT-6 Astra and GPT-6.1 Sol, by contrast, charge 2x input and 1.5x output above 272K input tokens.
• Caching is automatic. When the start of a prompt matches a cached prefix, those tokens cost $0.20 per million instead of $4. There is no flag to set, and Meta lists no charge for writing to the cache, unlike Anthropic.
• Reasoning is billed as output. Thinking tokens cost $20 per million like any other output, so higher effort raises the bill fast. Effort runs from minimal to max on Meta’s reasoning settings.
• Web search costs extra: $2.50 per 1,000 searches, on top of tokens. The steering prompt Meta adds to every request is not billed.
• No batch discount and no free credits are published.
• Rate limits: 3,000 requests and 4 million tokens per minute, shared by every key in a team.
• Billing: charged when your balance hits a threshold and on the 1st of each month. There is no hard spending cap, only email alerts, and no refunds.

What a task costs: four worked examples

Token prices only become useful when you turn them into jobs. Here is the arithmetic at list prices:

Two cautions. First, the model is wordy. Artificial Analysis says it generated 130 million tokens while running the Intelligence Index, against a median of 85 million, and puts the average cost of one index task at $4.52. AA’s launch post on X adds that this is about 54,000 output tokens per task, and gives $5.98 per task for Claude Opus 5.5 and $5.46 for Claude Sonnet 5.5.

Second, token prices hide retries and thinking. On Vals Finance Agent V2, Muse Wonder costs $2.67 per test, against $9.22 for Opus 5.5 and $0.76 for Muse Spark 1.3 Max. Judge cost per finished task on your own work, not price per token.

Budget for thinking, too. The invoice example assumes 2,000 output tokens per check. If thinking adds 20,000 more to each check, the month rises from $125 to $525, and Meta offers no hard spending cap, only email alerts. Set alerts and test a lower effort level before you commit.

Muse Code plans: the only way to Muse Wonder Full

Muse Code subscription prices in US dollars a month, from Meta’s Muse Code page. Usage is shown as a multiple of the Everyday Usage plan.

Muse Wonder is on every plan. Muse Wonder Full needs Power Usage or Maximum Usage, and Maximum Usage is new with this launch. The plans work only through the Muse Code command line tool; any other API key you create is billed pay as you go. Upgrades take effect at once and are prorated, there are no refunds when you cancel, and Maximum Usage subscribers who hit the limit must wait for it to refresh.

Meta does not say how fast Muse Wonder Full uses up an allowance, so there is no fair way to turn $50 a month into a token count yet. One setting matters more than the price: every plan starts in Contributor mode, which lets Meta train on your prompts, code, files and the model’s answers in exchange for more usage. Turn it off in Muse Code settings before your first prompt, because switching off only protects what you send afterwards. Both Wonder models still work with it off; you get less usage, and Meta does not say how much less. Muse Code on a pay-as-you-go API key runs Muse Wonder in Standard mode only.

What Meta Muse Wonder means for finance and accounting teams

Meta pitches the models at law, finance and health, and its newsroom post names making sense of someone’s finances as an example. The model card lists domain knowledge work in finance, including preparing drafts for review, as an intended use. Here is what the evidence supports so far.

The finance numbers, read carefully

Three things stand out. Vals now lists Wonder Full at 59.06, confirming Meta’s figure, but that ranks fifth on its public board: four models score higher, including Meta’s own cheaper Muse Spark 1.2 and Muse Spark 1.3 Max at under $0.80 a test, against $2.39 for Wonder Full. On AA’s finance tasks, Muse Wonder Full completes fewer objectives than Opus 5.5, GPT-6 Astra, Sonnet 5.5 or Gemini 4 Argon. And AA’s raw page data also shows it breaking more of the tasks’ guardrail rules: 71 violations across the 120 finance tasks, against 53 for Opus 5.5 and 31 for Sonnet 5.5.

Its best finance-adjacent results are narrow: first on AA’s GDP.pdf by 0.4 points and within 9 Elo of Opus 5.5 on GDPval-AA, though about 80 Elo behind Opus 5.5 on AA-Briefcase’s multi-week projects. Reading contracts, filings and statements and drafting from them is where we would test it first.

Your data: the setting to check first

• On the API, Meta does not train on your content. The model runs only in Standard mode there, and Meta’s Model API terms commit it not to use Standard-mode content to train its models.
• Muse Code subscriptions start in Contributor mode. In that mode Meta can use your prompts, code, files and the model’s responses to train future models, in exchange for more usage. Switch it off in settings before client or company data goes in. Switching off only covers future content, and the terms say not to put confidential or personal information into Contributor mode at all.
• Retention has no fixed period. Meta keeps content and usage data as needed to run the service, meet legal duties and review safety and abuse. Even in Standard mode Meta may keep and process your content to detect abuse and to develop, evaluate and improve its safety and security systems, and it may use usage data such as tool-call sequences and logs to improve its products.
• Zero data retention is not self-serve. Zero data retention is for qualified accounts through Meta’s sales team, may vary by model, and switches off file uploads, web search and background requests. Content flagged for policy violations can still be kept for up to two years.
• No SOC 2 or ISO certification is named. Meta’s data commitments cite TLS 1.2 or newer encryption in transit, PCI-DSS for card data and an internal security management system. If your clients or auditors need a SOC 2 report, ask Meta’s sales team before you sign.
• Some terms are in your favour. Meta indemnifies customers of its Standard services against third-party IP claims over unmodified outputs, when paid for with money rather than promotional credits; it does not cover Contributor mode. Customers in the European region contract with Meta Platforms Ireland.

File limits that matter for finance

The Model API’s file handling list has no spreadsheet or Word formats. It takes images, video, audio, PDF, and plain text or JSON. PDFs are read for text up to page 100 and as page images up to page 50. In practice that means sending spreadsheet data as text in the prompt or as a PDF, or working through tools, and splitting long annual reports before you upload them.

How we would test it

1. Pick 20 finished tasks with known right answers: a reconciliation, a variance commentary, a contract summary, a figure pulled from a filing.
2. Run them through Muse Wonder on the API and through the model you use today, at the same effort level, and compare accuracy and cost per finished task.
3. Include the update test: change a figure partway through a conversation and check which one the model uses.
4. Give it the current rules. Meta has not published a knowledge cutoff, so supply the rule text instead of trusting what the model remembers.
5. Let a calculator or your accounting system produce the numbers, and use the model to read, draft and check.
6. Keep people approving payments, filings and anything a client sees. Meta’s own model card recommends human confirmation before irreversible actions.

For prompts and setups that already work with other models, see our Claude for finance and accounting guide, our AI prompts for finance and our overview of AI in finance and accounting.

Muse Wonder safety: what Meta’s charts and model card show

Meta calls Muse Wonder its most aligned model to date, with its strongest safeguards yet against chemical, biological, radiological and nuclear misuse and against cyber misuse, while keeping defensive cyber work available. The full safety report is still to come; Meta says it will be available soon. Its launch charts compare the two new models with Fable 5.1 and GPT-6 Astra:

Fable 5.1 still refuses slightly more biology and cyber misuse requests. Where the Muse models stand apart is defensive security work: Muse Wonder complies with 94.5% of cyber defense requests in Meta’s test, against 15.7% for Fable 5.1. Muse Wonder Full also has the lowest personal information violation rate in the chart, 2.2%. That is still about one case in 45 on Meta’s test, so keep an agent’s access to client data narrow.

The last row matters most for agents that read outside material such as supplier emails, invoices or web pages: in Meta’s test, about one attack in ten still got the model to run code it should not have. Give such agents read-only access where you can, and keep a person approving any payment, filing or message they prepare.

The model card’s risk assessment

• Chemical and biological risk. Without mitigations, Meta does not rule out Muse Wonder Full reaching the high-risk threshold of its Superintelligence Scaling Framework in this domain. It applies safety training and system-level classifiers, and rates the risk moderate or lower with them in place.
• Cyber and loss of control. Moderate or lower even before mitigations, by Meta’s assessment.
• Cyber skills. Meta’s model card reports 88.4% on CyberGym, a cybersecurity benchmark, alongside the cyber misuse safeguards above.
• Honesty and misalignment. 87.3% on MASK, which tests honesty under pressure, and a 0% harmful action rate on Agentic Misalignment.
• Known limits. It can give inaccurate answers and make errors in multi-step reasoning on unfamiliar problems, and it is weaker at visual and 3D spatial reasoning. It is not intended for people under 18.

How to start using Muse Wonder

If your team does not write code, there is no point-and-click app for using Muse Wonder at work yet: it was still not in the Muse app on 10 November 2026, and Muse Code runs in a terminal. A developer can connect it to your tools through the API or OpenRouter. To try it in a browser first, Meta’s model page has a Try in playground button. Here is how to get started:

1. Sign up for the Meta Model API. Sign-up is self-serve at dev.meta.ai. You must be 18 or older and in a supported country, and you add a card and business details on the Billing page before your first request. Companies can use a managed Business and Enterprise account with a work email. The API is not offered in restricted territories such as China, Russia, Iran and Pakistan; Meta’s geographic use policy lists the rest.
2. Point your existing code at it. The API works with the OpenAI and Anthropic SDKs: point the OpenAI SDK at https://api.meta.ai/v1 or the Anthropic SDK at https://api.meta.ai, and set the model to muse-wonder-1.0. The Responses, Chat Completions and Messages formats all cost the same. Reasoning effort runs from minimal to max, and none returns an error. The API overview has a quickstart.
3. Or install Muse Code. One line installs the muse command: curl -fsSL https://dev.meta.ai/install.sh | sh on macOS or Linux, or irm https://dev.meta.ai/install.ps1 | iex in Windows PowerShell. Run muse in a project folder and sign in. To subscribe, sign in to Muse Code with your Meta Model API account, then pick the plan under Subscriptions in Meta’s Accounts Center. Before you open a client folder, turn off Contributor mode in settings, since every subscription starts with it on. The Muse Code docs still name Muse Spark 1.2 as the default model, so choose the model you want in its settings. If you are weighing coding agents, our Claude Code tutorial covers Anthropic’s equivalent.
4. Through OpenRouter. Meta’s model page also lists OpenRouter as a way to swap the model into existing tools; OpenRouter lists it at the same $4 and $20, and Vercel AI Gateway lists it too. Calls routed through a gateway follow that gateway’s own data and retention terms, and Meta’s zero data retention does not apply.
5. In the Muse app: not yet. Meta says Muse Wonder will soon power the Muse app, subscribers first, and had given no date by 10 November 2026. On 6 November 2026 Meta said it is rolling out specialised versions of its models to the Muse app, trained with learnings from Muse Wonder; it has not said these are Muse Wonder itself. Meta’s help page lists the app’s paid plans as Power at $20 a month and Max at $100 a month, and CNBC reports those are the plans that will get it. The app launched in the US and Canada and reached Japan on 23 October 2026 and Singapore on 27 October 2026; our guide to the Muse personal agent covers the app itself.
6. Test before you switch. Run 20 of your own past tasks at medium and max effort and compare the output and cost with what you use today.

Meta Muse Wonder vs Claude Opus 5.5, Sonnet 5.5 and GPT-6 Astra: which to use

A simple rule: if you already run Claude Opus 5.5, the list prices are identical, though Vals’ first cost-per-test figures put Muse Wonder at $2.67 against $9.22 for Opus 5.5, so compare cost per finished task on your own work. If you pay for GPT-6 Astra or Claude Fable 5.1, Muse Wonder costs 60% less per token and is worth a trial, but put Claude Sonnet 5.5 in the same trial: it scores higher on the independent index at half Muse Wonder’s price.

For the rivals, see our Claude Opus 5.5 review and Claude Sonnet 5.5 review, our GPT-6.1 Sol guide, our Gemini 4 Argon review, our GPT-6 Astra vs Claude Fable 5.1 comparison and our Claude Haiku 5.5 guide for cheap bulk work.

How Muse got here: from Spark to Wonder

By Meta’s own account, it began a ground-up overhaul of its AI work fourteen months ago. The result has been a shift from the open Llama models to closed Muse models built by Meta Superintelligence Labs. The timeline shows how fast it has moved since Muse Spark in April:

Two threads run through it. Meta has gone closed: Meta Muse Wonder is hosted only, and Muse Glimmer is still the only open-weight Muse model, while the promised open weights for Muse Spark are not out yet. And Meta has climbed the price ladder, from Muse Spark 1.3 at $1.25 and $4.25 to $4 and $20 now. CNBC and the New York Times report the model’s internal codename was Watermelon, and the Times reports that a pause in pretraining in July delayed the release. For the consumer side, see our guide to the Meta Muse AI agent and the Muse app launch post.

Key takeaways

• Should you switch? From Claude Opus 5.5, only where your own tests show a gain, since the token prices are identical. From GPT-6 Astra or Claude Fable 5.1, Muse Wonder is 60% cheaper per token and worth a trial next to Claude Sonnet 5.5.
• Meta released Muse Wonder and Muse Wonder Full on 12 October 2026. The API model costs $4 input and $20 output per million tokens, the same as Claude Opus 5.5; Full comes only with Muse Code plans at $50 or $200 a month.
• In Meta’s own 14-row table, Muse Wonder Full tops 6 rows and Muse Wonder 1, against 6 for Claude Opus 5.5 and 1 for GPT-6 Astra.
• On the independent Artificial Analysis index, Muse Wonder Full scores 56: below Opus 5.5 at 58 and level with Claude Sonnet 5.5, which Meta left out.
• Vals AI now confirms Meta’s finance figure (59.06 for Wonder Full, fifth on its board), but 13.1 on Harvey LAB-AA, 79.0 on DeepSWE and 57.5 on SWE-Marathon are not on public leaderboards yet.
• For finance, Gemini 4 Argon (65.40) and Meta’s own cheaper Muse Spark 1.2 both score higher on Vals, and on AA’s finance tasks Muse Wonder Full trails Opus 5.5, Sonnet 5.5 and GPT-6 Astra. Document work looks like its best fit.
• Meta does not use API content to train its models, but Muse Code subscriptions start in Contributor mode, where it does. Turn it off before client data goes in.
• Muse Wonder, the API model, is wordy: 130 million tokens on AA’s index against an 85 million median, and $4.52 per index task, still below the $5.98 AA gives for Claude Opus 5.5.
• Meta does not rule out high chemical and biological risk without mitigations, and its full safety report is still to come.

Please note: we update this page from time to time, but AI models, prices and benchmark results change quickly, and there can be a gap between a change and our update. Check Meta’s official pages before you act. This guide helps you compare and estimate; it is not legal, tax or investment advice, and ReconScribe is not responsible for decisions made on it.

Meta Muse Wonder: frequently asked questions

Related reading

• Meta Muse AI agent explained: features and privacy
• Claude Opus 5.5: benchmarks, pricing and how it compares
• Claude Sonnet 5.5: benchmarks, pricing and how it compares
• Claude Haiku 5.5: benchmarks, pricing and best uses
• GPT-6.1 Sol: pricing, benchmarks and our verdict
• GPT-6 Astra vs Claude Fable 5.1
• Gemini 4 Argon: benchmarks and our verdict
• AI leaderboard 2026: every major model ranked
• Claude Code tutorial for finance teams
• AI prompts for finance
• Claude for finance and accounting
• AI in finance and accounting

Sources

• Meta AI Research: Muse Wonder
• Meta Newsroom: Introducing Muse Wonder and Muse Wonder Full
• Muse Wonder model card
• Muse Wonder model page (Meta AI Developers)
• Meta developer blog: Build with Muse Wonder
• Meta Model API: Models
• Meta Model API: Overview
• Meta Model API: Pricing and rate limits
• Meta Model API: Reasoning
• Muse Code subscriptions
• Muse Code
• Muse Code docs
• Meta Model API terms of service
• Meta Model API: Zero data retention
• Meta Model API: Data commitments
• Meta Model API: File handling
• Meta Model API: Geographic use policy
• Meta Superintelligence Scaling Framework
• Meta Newsroom: Introducing Muse Spark
• Meta Newsroom: Introducing Muse
• Meta AI Research: Muse Glimmer
• Artificial Analysis: Muse Wonder Full
• Artificial Analysis: Muse Wonder
• Artificial Analysis: model leaderboard
• Artificial Analysis: AutomationBench-AA
• Artificial Analysis: Terminal-Bench 4.0
• Artificial Analysis: Harvey LAB-AA
• Artificial Analysis: GDP.pdf
• Artificial Analysis: GDPval-AA
• Vals AI: Finance Agent v2
• Anthropic: Claude Opus 5.5 system card
• Anthropic: Claude Sonnet 5.5 system card
• Anthropic: Claude Sonnet 5.5
• Anthropic: Claude API pricing
• OpenAI: Introducing GPT-6 Astra
• OpenAI: GPT-6 Astra model page
• OpenAI: Introducing GPT-6.1 Sol
• Google: Gemini 4 Argon