Meta Muse Wonder tested: what it does, what it costs and how it compares

The answer in brief

Muse Wonder is Meta’s most capable AI model family, released on 12 October 2026 by Meta Superintelligence Labs. It comes in two versions. Muse Wonder is open to developers through the Meta Model API at $4 per million input tokens and $20 per million output tokens (US list prices, checked 13 October 2026). Muse Wonder Full, the stronger version, runs only inside Meta’s Muse Code coding agent, on plans costing $50 or $200 a month.

Independent testers rank Wonder Full among the strongest models available, though not at the top. It scores 56 on Artificial Analysis‘s Intelligence Index, behind Anthropic’s Claude Opus 5.5 (58), level with Claude Sonnet 5.5 (56) and ahead of OpenAI’s GPT-6 Astra (53), and places fifth of 47 models on the Vals Index of finance, coding, legal and tax work. The standard Wonder, the one developers can actually call, scores 52 and places tenth on Vals. On Artificial Analysis’s hallucination test, both versions are far less likely than the Anthropic and OpenAI models we compare to give a wrong answer when they should admit they do not know, though Meta concedes that visual and 3D reasoning remain weak.

Our own blind tests, run on 13 October 2026 and marked the next day, added Google’s Gemini 4 Argon. The standard Wonder trailed Opus and Astra on writing: it never came first on any of six short briefs, marked twice, and its longer blog post came last, while Wonder Full’s drafts were among the best. In design, the standard Wonder made the best dashboard, but its first 3D scene failed to load. In code, the standard Wonder matched Wonder Full, passing every check on a build task and fixing all 12 bugs we planted, in less time. Gemini cost the least wherever we could compare bills, at an introductory price that Google doubles on 1 December 2026, but its short writing came last in every brief.

For most UK small businesses and marketers there is nothing to switch to yet. The Muse app that will carry Wonder is limited to the US and Canada, and the other ways in are developer tools. Teams that build on Opus or Astra should try Wonder on long coding or document-heavy agent work and keep their current model for writing. Wonder is the first model Meta has built for the frontier, 18 months after its Llama 4 release disappointed, and it comes close without making Meta the leader.

What is Muse Wonder?

Muse Wonder is the second generation of Meta’s Muse models and, as TechCrunch puts it, “the company’s first attempt at a frontier model”. Its predecessor, Muse Spark, was built to be small and fast. Wonder is larger: Mark Zuckerberg described the series on X as “a much larger and more advanced class of models”.

The project’s internal codename was Watermelon, according to CNBC. Meta trained both models at Prometheus, its gigawatt-scale data centre in central Ohio, scaling up the pretraining and reinforcement learning approach it set out with Spark.

Neither model has downloadable weights, and Meta has not disclosed their size. That continues Meta’s break with its Llama years, when anyone could run its models on their own hardware. The one open exception in the Muse family is a smaller model, Muse Glimmer, released under the Apache 2.0 licence, which we tested on agency tasks.

What do you get with each version?

Why does Muse Wonder matter for Meta?

Wonder is the first public test of whether Meta’s AI spending can buy a place at the frontier. Llama 4 was widely panned in April 2025. In June that year Meta made Scale AI’s chief executive, Alexandr Wang, its chief AI officer as part of a $14.3 billion deal, and it reorganised its AI work that summer.

The new unit’s first model, Muse Spark, tied for fourth on Artificial Analysis’s index when it launched in April 2026, Reuters reported. In July, Wang reportedly told staff that the Watermelon model had caught up with OpenAI’s GPT-5.5, without saying on which benchmarks (The New Stack has a summary). This launch puts published numbers behind that claim.

Meta’s advantage is reach. Its apps serve more than 3.5 billion people, according to Reuters, and Muse already runs on WhatsApp as well as its own app and website. A stronger model inside that funnel is what Meta is betting on.

What is the difference between Muse Wonder and Muse Wonder Full?

Muse Wonder is the model developers can build on. It is available through the Meta Model API as muse-wonder-1.0, inside Muse Code and on OpenRouter. The API accepts the same request formats as OpenAI’s and Anthropic’s developer tools, so most existing agent setups need only a new address and key.

Muse Wonder Full is the stronger variant, and it has no API: you can use it only through Muse Code, Meta’s command-line coding agent for macOS and Windows. Most of the headline scores in Meta’s launch belong to Full.

People who use the Muse agent app will meet Wonder there. Meta’s announcement says Wonder will soon power Muse, starting with subscribers on its $20 and $100 monthly plans, though Zuckerberg’s launch post told people they could subscribe and use it today, so the two launch messages disagree on timing. The free tier stays on Muse Spark because Wonder costs more to serve, and Zuckerberg says Meta hopes to fit Wonder-level ability into a Spark-sized model.

Can UK teams use Muse Wonder?

Developers can, though the Muse app has not reached the UK yet. Meta describes Muse as available in the US and Canada (checked 13 October 2026). The Meta Model API terms point to a separate Geographic Use Policy, last updated on 12 October 2026, which bars the API in a list of territories that includes Russia, China and Iran. The UK is not on that list, and the only model with a country ban of its own is Muse Image, in India (checked 14 October 2026). From the UK on 13 October 2026, we called the standard Wonder through OpenRouter and used Wonder Full on a Muse Code Power plan. The API terms also include extra terms for consumers in the UK and EU, so read them, and the data terms below, before you build client work on Wonder.

How much does Muse Wonder cost?

Through the API, Muse Wonder costs the same as Claude Opus 5.5 and well under half the list price of GPT-6 Astra or Claude Fable 5.1. All prices are US list prices, checked 13 October 2026. Web search adds $2.50 per 1,000 queries, and Meta charges the same token rate however much of the one-million-token context a request fills.

Wang had said the API would be priced well below OpenAI and Anthropic, according to Fello AI’s account of his CNBC interview. That holds against GPT-6 Astra and Claude Fable 5.1. It does not hold against Claude Opus 5.5, which costs exactly the same, or Sonnet 5.5, which costs half as much and scores higher on both of the independent indexes covered below. The token price also understates Wonder’s cost per job. On Artificial Analysis’s index, a task cost $4.52 with Wonder, less than with Claude Opus 5.5 ($5.98) or Sonnet 5.5 ($5.46) but more than with GPT-6 Astra ($3.26). Artificial Analysis put that cost down to heavy input use: 7.69 million tokens per task.

Wonder Full has no token price. It comes with Muse Code’s $50 Power Usage plan, which gives 20 times the allowance of the $5 Everyday Usage plan, or the $200 Maximum Usage plan at 80 times. The cheaper $5 and $15 plans do not include it.

Read the data terms before you sign up. API traffic runs in Meta’s Standard mode, where your prompts are not used for training, and Meta does not offer Wonder at the discounted Contributor rate it sells for Muse Spark. Muse Code subscriptions default to Contributor mode, which gives you more usage from your plan in exchange for letting Meta train on your content. You can switch it off at any time, but that covers only what you send afterwards. Qualified organisations can ask for zero data retention, but that setting turns off server-side conversation state, background requests, file uploads and built-in web search, which are the features a memory-heavy agent leans on. Our Muse Code guide covers its cost, data settings and safe set-up.

What can Muse Wonder actually do?

Meta pitches Wonder on two abilities: remembering what matters about you, which it calls Active Memory, and persistence on hard tasks, which it calls Active Focus. Active Memory carries details across conversations and pulls them back when they become relevant. Meta measures it with MemoryBench, an internal test described in the model card and built from conversations of 64,000 to 100,000 tokens (roughly 50,000 to 75,000 words), laced with distracting exchanges and later corrections the model has to track.

Active Focus means that when a task gets hard, the model is meant to try another route, recover from errors and keep going, asking for clarification when it needs to and checking in before anything important. Meta tests this with FocusBench, an internal evaluation of personal tasks in housing, travel, IT and finance, and with SWE-Marathon, an outside benchmark of very long software projects.

Around those sit specialist knowledge in law, finance and health, coding, and computer use across browsers, desktop and mobile apps. Meta’s demonstrations are deliberately everyday: planning a Saturday run in a browser around someone’s preferences, designing and ordering custom merchandise from photos on an Instagram profile, solving a wedding seating chart with awkward constraints, and building a playable 3D kitchen game from images.

The early praise comes from people close to Meta. Shopify’s chief executive, Tobi Lütke, wrote on X that he “was struggling to find anything to throw at it that it couldn’t do, even when other top models failed”. He is not a neutral tester: Fello AI points out that he sits on Meta’s advisory group and that Shopify’s catalogue is built into Muse.

Meta is open about the weak spots. Wang conceded on X that the team had not trained Wonder on 3D demo games and urged people to try it on real coding work, and the model card lists visual and spatial reasoning, 3D especially, among its known limitations.

How does Muse Wonder compare with GPT-6 Astra and Claude Opus 5.5?

At the same price as Claude Opus 5.5, the standard Wonder wins five of Meta’s 14 comparisons and loses nine. Wonder Full does better against Opus, winning eight to six, but you cannot buy it by the token. Meta tested against Claude Fable 5.1, Claude Opus 5.5 and OpenAI’s GPT-6 Astra, all at maximum reasoning effort. The 14th row, the Artificial Analysis index, is charted in the next section.

Several leads are slim: Wonder Full beats Fable 5.1 on finance by 0.2, and Opus beats Wonder Full on JobBench by 0.1. The standard Wonder’s win over Opus on GDP.pdf is by 0.2 points, while its losses include 20.5 points on ProgramBench and 9.2 on OSWorld.

Wonder’s clearest strengths are long-horizon coding and legal work. Its DeepSWE score comes from Datacurve, a third party that ran both Wonder and Opus, which makes it the cleanest win on the card. On Harvey’s legal benchmark Wonder more than doubles Opus, though even the top score is 13.1: under that strict grading, every model still fails most tasks.

Meta’s model card is open about the limits of its own table. Where a comparable third-party or provider-reported result existed, Meta used it; otherwise it took the highest of its own run, the rival’s self-reported score or the official leaderboard score, and it calls its own runs of rival models best effort. The SWE-Marathon row mixes different agent tools, and Wonder Full’s score there was recorded at xhigh effort, against the max setting Meta used elsewhere. Two rows, the Personal Superintelligence Index and MemoryBench, are Meta’s own tests, and a language model grades MemoryBench.

What do independent tests show?

Independent testers broadly confirm Meta’s picture: Wonder Full sits just below the top, and the standard Wonder in the pack behind it. Artificial Analysis runs every model through the same ten evaluations, which makes its Intelligence Index the fairest single check on Meta’s table.

On version 4.3.2 of the index, Wonder Full scores 56, against 58 for Claude Opus 5.5 and 56 for Claude Sonnet 5.5, and it beats GPT-6 Astra, Gemini 4 Argon and Claude Fable 5.1. Artificial Analysis ranks it fourth of the 690 models it lists. Meta’s own comparison leaves Sonnet 5.5 out. The standard Wonder scores 52, a point behind Astra’s 53. In its launch analysis, Artificial Analysis found Wonder Full strongest at agentic office work and scientific reasoning: it narrowly leads every model on Artificial Analysis’s own run of GDP.pdf, at 33%, where Meta’s table had put Astra ahead, ranks second on GDPval-AA and tops the CritPt science test at 33%.

Wonder’s clearest independent lead is in admitting what it does not know.

AA-Omniscience counts how often a model gives a wrong answer when it should have admitted it did not know. By that measure, Wonder bluffs far less often than the Anthropic and OpenAI models it is compared with. At maximum effort, Artificial Analysis puts Wonder’s rate at 26.5% and Wonder Full’s at 23.0%, against 47.0% for Claude Sonnet 5.5, 51.3% for GPT-6 Astra, 58.6% for Claude Opus 5.5 and 72.6% for Claude Fable 5.1 (checked 13 October 2026). For an agent reading your lease or paying your bills, that matters more than a point on a leaderboard.

Vals AI, which tests models on agentic work in finance, coding, law and tax, puts Wonder Full fifth of 47 models on its index, ahead of every OpenAI model but behind Gemini 4 Argon, Claude Sonnet 5.5, Claude Opus 5.5 and Claude Fable 5.1.

The coding results pull in different directions. On Artificial Analysis’s Coding Agent Index, which runs each model inside its maker’s own coding tool, Wonder Full in Muse Code scores 68, level with Claude Code running Sonnet 5.5 and ahead of Claude Code running Opus 5.5, Google’s Antigravity CLI running Gemini 4 Argon and OpenAI’s Codex running Astra. On Vals’ Terminal-Bench run, where every model uses the same simple agent, Full trails Opus and Astra and just edges Gemini 4 Argon, though Opus’s score includes attempts its provider passed to a fallback model; counting those as failures, Vals puts Opus level with Full. ProgramBench is starker. Vals counts a program as rebuilt only when it passes every hidden test, and on that measure Wonder Full rebuilt one of 200 programs, against 37 for Opus. Meta’s table counts a run that passes at least 95% of the hidden tests, a looser bar, so its scores are far higher and the two cannot be compared directly. Vals also logged six hours for Full’s run, against under two and a half for Opus’s. Whichever bar you prefer, one rebuild in 200 is reason enough to test Full on your own code before trusting Meta’s coding numbers.

On Design Arena, where members of the public vote between two anonymous designs built from the same prompt, the standard Wonder ranked second of all models for websites on 13 October, after 2,595 head-to-head votes. It was ahead of GPT-6 Astra and Claude Opus 5.5, though the top seven sat within 30 Elo points of each other. On 3D design it ranked sixth, behind both rivals, which fits Meta’s own warning about 3D work.

No independent writing benchmark had scored either version by 13 October. EQ-Bench’s creative writing and long-form writing leaderboards listed Muse Spark models, and Muse Glimmer on creative writing, but no version of Wonder, and Arena, whose public votes rank models on writing among other things, had added Wonder to its Agent Arena without publishing a score. Meta lists writing and editing among the model’s intended uses, yet none of the evaluations in its model card measures writing either. That gap is why we ran our own writing tests, along with design and coding.

Wonder is quick once it starts, but slow to start. On Artificial Analysis’s measurements, the standard Wonder generated 87.9 tokens a second, a little above the median for its price bracket, but at maximum effort took 19.27 seconds to start replying, against a median of 5.52 seconds, a wait that repeats on every step of an agent’s work.

How did Muse Wonder do in our own tests?

The standard Wonder matched Wonder Full on code and made the best dashboard, but trailed Opus and Astra on writing, where Wonder Full’s drafts were among the best. We tested Muse Wonder ourselves on 13 October 2026, with our own OpenRouter credits and a Muse Code Power plan, on writing, web design, a coding build and a bug hunt. Claude Opus 5.5 and GPT-6 Astra took only the writing and design tests, and that evening we added Google’s Gemini 4 Argon, which tops the Vals Index, to every test. Its costs in our tables are at Google’s introductory price, which doubles on 1 December 2026.

The results compare the setups as we ran them: Wonder Full always worked in Muse Code with its own tools, while the other models answered the writing and design briefs in a single reply through OpenRouter, as Gemini did with its code.

How did we score the tests?

Every entry appeared under a letter, with no model name, in a random order that changed from set to set. For each brief we gave every entry a mark out of 10 and a place from first to fifth, judging which we would publish, send or ship with the fewest changes; the blog posts also got marks for accuracy against the fact sheet and for writing. Places could not tie, so where two entries shared a mark, the place records which we preferred. We read the drafts as text and opened the web pages live in a browser, at desktop and phone width, dragging the 3D scenes to orbit them. One editor marked everything, on 14 October, so the writing and design results are one experienced reader’s judgement and a small sample. The code and the page faults were marked by hidden tests and automated checks.

Does Muse Wonder write as well as Opus, Astra and Gemini?

In our blind marking the standard Wonder wrote worse than Opus and Astra and better than Gemini, while Wonder Full came close to the best. Across six short briefs marked in two rounds, we picked GPT-6 Astra’s draft as the best six times out of 12, Wonder Full’s four times and Claude Opus 5.5’s twice; the standard Wonder and Gemini 4 Argon never came first, and Gemini’s draft came last every time. Wonder Full also had the highest average mark, 8.0 out of 10, against 7.9 for Astra, 7.8 for Opus, 7.2 for the standard Wonder and 5.4 for Gemini.

We gave Muse Wonder, Claude Opus 5.5, GPT-6 Astra and Gemini 4 Argon the same six briefs: a blog opening, a LinkedIn post, a customer apology email, a product description, a short explainer and the opening of a short story. Each brief gave the topic or the facts, asked for British English and a length, and said nothing about style or tone, so the drafts show each model’s default voice. All four ran through OpenRouter, with each request sent to the model maker’s own service at the default reasoning setting. We ran every brief three times: 72 drafts, $0.84 in all. Wonder Full is not on the API, so it wrote two drafts of each brief in Muse Code, at its default high effort and with its own set-up.

We marked two rounds, pairing the API models’ first and second runs with Full’s two sets, and left the third API run unmarked, though the em dash, length and spelling checks cover all 84 drafts. Wonder Full’s four wins were the product description in both rounds and the second-round blog opening and LinkedIn post. Opus won the first-round LinkedIn post and the second-round explainer, and Astra took the other six, including both apology emails and both story openings.

On length and spelling the five were indistinguishable: all 84 drafts landed within the requested length, and none used any of the American spellings we checked for. Gemini leaned hardest on stock phrases. Words from our list of 44, such as “crucial”, “boasts” and “excited to announce”, turned up 18 times in its 18 drafts, against two to four for each of the other API models and once in Wonder Full’s 12. Wonder’s sentences varied least in length.

The apology email showed the clearest difference. Astra was the only model whose two emails invented nothing. Opus’s first email made up names and a tracking number, both of the standard Wonder’s emails added details the shop never gave, such as a “by Friday” deadline, and both of Gemini’s left fill-in placeholders unfilled. Sticking to the facts in a brief is a different skill from admitting ignorance, which AA-Omniscience measures.

Wonder also thought hardest before writing, spending about three times as many reasoning tokens as Opus (hidden working that is billed as output), so at the same list price it cost twice as much per draft. It was still quick: at default settings it finished a typical draft in a median 6.1 seconds, against 7.1 for Opus and 7.4 for Astra. Gemini was quicker again, at 5.0 seconds, and the cheapest, at 0.4 cents a draft, or about 0.8 cents at Google’s December price.

Can Wonder Full write a longer blog post?

Yes. In our blind marking its post came second, behind Claude Opus 5.5’s and level on marks with GPT-6 Astra’s, which we placed third, while the standard Wonder’s came last. Short briefs may hide what a model does over a longer piece, so we set a longer job: a blog post of about 900 words about Muse Wonder Full itself, for UK small-business owners and marketers. Each model got the same brief and a one-page fact sheet drawn from this article, with orders to add nothing that was not in it.

Opus’s post read best: accurate, readable and well judged for its audience, with a practical, fair verdict. Wonder Full’s was accurate and complete, with a clear structure, if a touch list-heavy, and Astra’s had useful framing and a handy table of the plans, with some padding. Every post scored 9 for accuracy against the fact sheet in our reading, so the writing decided the order. Gemini’s used Title Case headings and put its benchmarks under the verdict, and the standard Wonder’s ran longest, at 1,299 words, and was too repetitive to publish without heavy cuts.

How good is Muse Wonder at web design?

Good on the dashboard and poor on the 3D scene, in our test. In our blind marking the standard Wonder made the best dashboard, but its first 3D scene failed to load, and Wonder Full came second on the landing page and the 3D scene and last on the dashboard. Claude Opus 5.5 made the best landing page and GPT-6 Astra the best 3D scene.

We gave Muse Wonder, Claude Opus 5.5, GPT-6 Astra and Gemini 4 Argon three design briefs through OpenRouter at default settings, and gave Wonder Full the same briefs in Muse Code: a landing page for an invented Bristol pottery studio, the orders dashboard of an invented online tea shop, and an interactive three.js scene of a lighthouse at dusk. Each brief asked for one HTML file that works at any width from 360 to 1,440 pixels; the landing page and dashboard briefs also listed the content and set basic accessibility rules. Muse Code gave Full its built-in design guidance, plus a three.js guide for the 3D scene, and let it edit and recheck its page over 9 to 39 model calls.

The standard Wonder’s dashboard came first, polished and warm with a clear hierarchy and a strong phone layout, ahead of Gemini’s, which earned the same mark for its status filters and stock bars; Wonder Full’s came last, its phone header taking too much space. Opus’s landing page had the strongest visual identity, and Astra’s came last, its small grey text looking faint and its fonts unfinished. Astra’s 3D scene was the most polished, with clear waves and a visible boat, and the standard Wonder’s came last because a script error stopped it before anything was drawn.

The automated checks found little to separate the pages: none scrolled sideways or left out content, and only Astra’s dashboard and Gemini’s two pages drew accessibility faults. Both Meta models scattered em dashes through their copy: Wonder’s landing page used 14 and Full’s 17, against three each from Astra and Gemini and none from Opus.

Gemini’s pages did better than its writing. Its landing page was clean and readable, though a little generic, with a typo (“oroatmeal”) and a 2025 date in the footer. Its dashboard was very usable, but it showed figures the brief never gave, such as 14 orders awaiting dispatch and a £35 free-post threshold. Its 3D scene loaded and orbited, with a dramatic sunset and a bold beam.

We ran the 3D brief twice through the API and marked each model’s first attempt. Wonder’s second scene loaded, as did all six other API scenes and Full’s single attempt. On first attempts through the API, Wonder was the quickest designer on all three briefs and Gemini the cheapest, at 8 to 21 cents a page; the table gives the landing-page figures. At Google’s December price, Gemini’s dashboard would have cost 43 cents, against 31 cents for Wonder’s.

Is Wonder Full better at coding than Wonder?

Not on the job we gave it. In Muse Code, on a Power plan, we asked each model to build a Python tool that checks a shop’s product spreadsheet and turns it into a Google Merchant Center feed, with rules for prices in three currencies, stock wording, barcode check digits and links, plus a JSON error report and its own unit tests. A hidden set of 40 checks, which no model saw, marked the work. A tool that rejected every row would have passed 13 of them. Muse Code runs only Meta’s models. Version 1.3.4 of Google’s Antigravity CLI did not offer Gemini 4 Argon on our Google AI Pro plan and would not accept it with an API key either, and Google’s pricing page says the model is not on that plan. So we gave Gemini the same task through OpenRouter as a single prompt at high reasoning effort. It could not run code or see a test result, and had to reply once with every file.

Every run passed all 40 checks, including one from Meta’s smaller Muse Spark 1.3 and both of Gemini’s single replies. The differences were time and effort: Wonder Full needed 19 model calls and 6.2 minutes, the standard Wonder about half that time. Gemini’s one slip came in its own tests. In its second reply, one line of the test file carried a stray space, so Python refused to run the file. A single reply gave Gemini no chance to run its tests and catch the slip. A clearly specified job of a few hundred lines is within reach of every model we tried, so Full’s extra power bought nothing here.

Can the standard Wonder fix bugs as well as Wonder Full?

Yes, on the job we set. In Muse Code both fixed all 12 bugs we planted in every run, and Gemini 4 Argon, writing its fixes in a single reply, fixed all 12 when the rules were written down. Meta’s own table puts the standard Wonder four to nine points behind Full on three of its four coding benchmarks, so we built a harder test: a small invoicing library for UK sole traders, about 380 lines of Python plus 110 lines of tests, with 12 bugs planted in it. Some are plain coding mistakes, such as a database query built from a customer’s name and a file name that lets “../” escape its folder. Others need UK knowledge: a tax year that starts on 1 April instead of 6 April, VAT worked out on the price before a discount instead of after it, and a £10,000 debt given the £70 fixed recovery sum instead of £100. A hidden test for each bug, which no model saw, decides whether it was really fixed.

We ran the job in two versions. In the first, the docstrings and README spelled out the rules the code should follow. In the second, we deleted those rules and left the code, so a model had to know the UK rules itself. In Muse Code, Wonder and Wonder Full had to find and fix the bugs, add tests and leave everything else working. Gemini got the same fix job through OpenRouter as a single prompt holding the whole library, and replied once with every file it changed.

Neither Meta model broke any of the behaviour our hidden tests protect, and Full was again the slower of the two. Full’s leads over Wonder in Meta’s table come on bigger jobs, such as ProgramBench’s programs built from scratch and SWE-Marathon’s very long projects.

Is Muse Wonder safe?

Meta’s own tests say yes, with safeguards in place, though for a business the more practical risk is privacy. Meta calls Wonder its most aligned model so far, with its strongest safeguards yet against chemical, biological, radiological, nuclear and cyber misuse. It says it deliberately kept the model able to help security teams find flaws in their own code. Meta says a full report with further safety evaluations will follow.

The model card goes further than the announcement. Tested without safeguards, Meta could not rule out Wonder Full reaching its high-risk threshold for chemical and biological threats, so it added safety training and system-level classifiers. With those in place, it rates the remaining risk as moderate or lower in every domain it tracks.

The headline safety results are strong on paper. Wonder Full refused 98.7% of high-risk biology prompts in SecureBio’s BioTIER test, took harmful actions to advance its goals in 0% of an agentic misalignment test, and scored 87.3% on MASK, which checks whether a model stays honest under pressure. The cyber results show how capable it is in that field: in ExploitGym, a hacking test, it captured 92 flags, though only 21 came through the flaw the test intended; it found its own way to the other 71.

A chart in Meta’s launch post puts Muse Wonder’s rate of mishandling personal information in Meta’s own test at 3.4%, and Wonder Full’s at 2.2%, against 4.6% for GPT-6 Astra and 6.8% for Claude Fable 5.1 (Opus 5.5 is not shown). For an agent designed to read your messages and calendar, even 3.4% is roughly one test case in 30 handled badly.

Meta also lists limitations beyond 3D. The models can make errors in multi-step reasoning on unfamiliar problems, and results vary with the tools and agent harness around them. They are not meant for anyone under 18, and Meta advises human confirmation before any irreversible action.

Should your business switch to Muse Wonder?

For most UK small businesses and marketers, not yet. The Muse app that will carry Wonder is limited to the US and Canada, and the API and Muse Code are developer tools.

If you already build on Claude Opus 5.5, the evidence does not justify switching either. Wonder costs the same per token and trails Opus on nine of the 14 rows in Meta’s own table and on both independent indexes, while Claude Sonnet 5.5 scores higher on both indexes at half the price. Wonder does deserve a trial for long-running coding agents, where its DeepSWE and SWE-Marathon results are strongest, and for legal or document-heavy agent work. Its low rate of made-up answers on Artificial Analysis’s test makes it worth testing on research.

Teams on GPT-6 Astra would pay 60% less for input and output tokens with Wonder, and 80% less for cached input, though not always less per job. Wonder cost more than Astra per Artificial Analysis index task and per short writing draft, and less on the long blog post and on our design jobs.

If cost is the attraction, test Google’s Gemini 4 Argon too. It leads the Vals Index, and in our tests it cost the least wherever we could compare API bills. Writing code in a single reply, without running it, it passed all 40 build checks and fixed every bug when the rules were written down. Those bills were at Google’s introductory price, though. From 1 December 2026 Gemini costs the same per token as Wonder, and at that rate it would still have been cheapest on every job we could compare except the dashboard. Its short writing came last in every brief we marked, though its dashboard came second, so the savings come with weaker copy.

Wonder Full is a different proposition. It is the best model Meta has built, but you can use it only inside Meta’s coding tool, on Meta’s data terms, with no way to put it into your own product. In our tests it was slower than the standard Wonder on code and no more accurate, though its writing was among the best, with the highest average mark on the short briefs and second place on the long post, and its design work was uneven. Independent coding results are mixed: inside Muse Code it scored level with Claude Code running Sonnet 5.5 on Artificial Analysis’s Coding Agent Index, but it fell far behind Opus on Vals’ ProgramBench run. Try it on your longest coding job before you pay for a second month, because its leads in Meta’s table come on long projects and our shorter jobs showed no gain over the standard Wonder.

For everyday users, the real question is Muse. Wonder should make the agent better at the multi-step chores it was built for, from bookings to forms, and it reaches paying subscribers first. The privacy trade-off of handing an agent your inbox and payment details stays the same whichever model runs underneath.

In the 18 months since Llama 4, Meta has gone from also-ran to contender. Until Wonder Full is open to developers and the Muse app reaches the UK, most teams should test it rather than switch to it.