October 12, 2026, according to a post by developer Max Blade
A developer gives two AI models the same request: build me a lawn-mowing game, big-studio quality, with a beer and a cigar.
Claude Opus 5.5 comes back with an estate with trees, shadows and a small map in the corner. The bill: $99.43. Muse Wonder, Meta's new model, comes back with green grass, a fence and a steering wheel. The bill: $4.11.
Both games work. One looks like a real game, the other like a rough draft of one. And on that very same day, Meta announced that its model belongs in the same league as the best in the world.
How can both be true at once? That is the question this article sets out to answer.
First, the context. Muse Wonder is Meta's first big model, and it came out on October 12, 2026. Until now, Meta trailed: its competitors, Anthropic and OpenAI, had far stronger models. With Wonder, Meta claims it has closed the gap. Mark Zuckerberg published a score table in which the new model beats Claude Opus 5.5, the model considered the strongest today, on most of the tests.
I read Meta's announcements, the technical document it published on the model, and the tests developers ran in the first 24 hours. I did not run Wonder myself, and I flag along the way what is fact and what is my opinion.
In short
- Muse Wonder is Meta's first big AI model. It is far better than Meta's earlier models and close to the leaders, but it has not overtaken them.
- There are two versions. The strong one, Wonder Full, is the one in the headlines, and it is locked inside a Meta subscription. The version businesses can connect to their own systems is weaker.
- The price jumped. Wonder costs the same as Claude Opus 5.5, roughly three to five times Meta's previous model.
- My opinion: Meta built a model that finishes the task without failures and cheaply, not a model that produces the best result. For routine tasks that is enough. For work where quality is the product, not yet.
What Muse Wonder is, and how it differs from Wonder Full
Meta's earlier models, the Muse Spark family, were light, fast and cheap. Wonder is a different family: a much larger model, built for long and complex work. The launch comes 14 months after Zuckerberg rebuilt the company's AI organization. For anyone following Muse, Meta's personal agent, this is the piece that was missing: until now the agent ran on a light model.
The first thing to understand is that Wonder comes in two versions, and they are not sold the same way:
Muse Wonder
The version anyone can buy
- Pay per use, and connect it to any system or agent
- $4 per million input tokens and $20 per million output tokens. A token is the unit of text a model bills by
- The weaker of the two
- According to Meta, what you send to it is not used for training
Muse Wonder Full
The version in the headlines
- Available only inside Muse Code, Meta's coding tool, on a $50 or $200 monthly subscription
- Cannot be connected to your own systems
- The stronger of the two
- On the subscription, your content is used for training by default, and you can turn that off
Two new features are worth knowing by name. Active Memory: the model remembers important details about you from one conversation to the next. Active Focus: when it gets stuck it tries another route instead of giving up, and it stops to ask for approval before an important action. According to Meta, Wonder will soon reach the Muse app as well, paying subscribers first. And if you are in Israel: Muse is not available there yet.
Zuckerberg's table: what it shows, and what it doesn't
AI companies grade models the way a school grades students: with tests. Each test covers a different subject, such as writing code, legal work or operating a computer, and the score is the percentage of tasks the model completed.
The table Zuckerberg published is that kind of report card, with 14 subjects and five students: the two versions of Wonder and three models from the competition. Here is the report card when you ask only one question, who got the top score in each subject:
Who leads each of the 14 tests in Meta's table
Wonder Full leads 6
- Coding on long projects
- Coding on very long projects
- Business process automation
- Legal work
- Financial analysis
- Meta's composite index
Claude Opus 5.5 leads 6
- Building software from scratch
- Command-line work
- Operating a computer
- Health questions
- Professional office tasks
- The independent general index
Others lead 2
- Memory: the standard Wonder
- Working with documents: GPT-6 Astra
My own count from the table Meta published on October 12, 2026.
Six against six. For a lab that was far behind until six weeks ago, that is a real achievement. But when you read the technical document behind the table, three things change the picture.
First: the top student is not the one for sale. All six wins belong to Wonder Full, the closed version. The standard Wonder, the one a business can connect to its agent, leads in a single subject: memory.
Second: in some subjects Meta is also the teacher. Meta wrote the memory test itself, and the grade on it was given by another AI model. The composite index is Meta's too. And on the very long coding test, Meta's own document notes that this is not a fair comparison: each model was tested inside a different tool.
Third: two students are missing from the report card. Claude Sonnet 5.5 and GPT-6.1 Sol cost half as much as Wonder, and Meta left them out of the comparison. On the general index from Artificial Analysis, an independent body that runs every model under the same conditions, Sonnet 5.5 gets exactly the same score as Wonder Full.
What does hold up? Two wins were measured by third parties and by a clear margin: coding on long projects and business process automation. Those are the two areas where I would take Wonder seriously.
This is Zuckerberg's original post, with the table as Meta designed it:
Mark Zuckerberg's post on X with the score table, October 12, 2026
And below it, the same table with plain-language test names. In each row the top score is marked in green, and a tag marks the tests Meta measured itself:
The table Meta published, in plain language: every model's score on every test
| Test | Wonder Full | Wonder | Claude Fable 5.1 | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|---|
| Personal and professional agents | |||||
| Meta's composite index Measured by MetaPersonal Superintelligence Index | 59.1 | 54.3 | 49.9 | 54.5 | 55.2 |
| Legal workHarvey LAB-AA | 13.1 | 10.0 | 6.4 | 8.6 | 4.2 |
| Financial analysisVals Finance Agent V2 | 59.1 | 56.6 | 58.9 | 53.5 | 58.6 |
| Health questions Measured by MetaHealthBench Pro | 62.3 | 60.5 | 62.1 | 65.3 | 65.6 |
| Business process automationAutomationBench-AA | 45.4 | 39.1 | 32.1 | 41.6 | 41.2 |
| Working with documentsGDP.pdf | 33.2 | 30.8 | 27.6 | 34.2 | 30.6 |
| Professional office tasksJobBench | 75.3 | 70.5 | 73.1 | 66.1 | 75.4 |
| Operating a computer Measured by MetaOSWorld V2.1 | 78.4 | 72.6 | 80.7 | 75.8 | 81.8 |
| Memory Measured by MetaMemoryBench | 85.5 | 85.6 | 79.0 | 81.4 | 81.5 |
| Coding | |||||
| Command-line workTerminal-Bench 4.0-AA | 57.1 | 52.0 | 52.0 | 59.1 | 59.6 |
| Coding on long projectsDeepSWE 1.1 | 79.0 | 78.8 | 67.4 | 74.1 | 74.2 |
| Coding on very long projects Not a controlled comparisonSWE-Marathon | 57.5 | 53.1 | 45.6 | 42.5 | 46.3 |
| Building software from scratchProgramBench | 59.0 | 50.0 | 55.5 | 52.5 | 70.5 |
| General | |||||
| The independent general indexArtificial Analysis Index | 56 | 52 | 53 | 53 | 58 |
The top score in the row. "Measured by Meta": the test was written or run by Meta itself. Source: Meta's launch announcement, October 12, 2026. The plain-language test names and the markings are ours. For OSWorld the partial score is shown, and all models were measured at the highest thinking setting.
Look at the legal row: even the leader succeeds on only 13 tasks out of 100. No model yet produces a legal document you can rely on without a lawyer.
How many months behind is Meta? Fewer than I thought
My first impression, shared by some of the critics, was that Wonder shows the level the frontier was at four to six months ago. I checked that against one index that runs every model under the same conditions, the Artificial Analysis Intelligence Index, and the numbers narrow the gap.
Artificial Analysis index (version 4.3): the score, and when the model was released
Source: Artificial Analysis, Intelligence Index v4.3, all models at the highest effort level. Release dates from the makers' announcements.
Wonder, the version in the API, sits between Opus 5 from late July and Fable 5.1 from early September. So on this index it is at the level the frontier reached six to eleven weeks ago, not half a year ago. Wonder Full matches Sonnet 5.5, which came out two weeks before it.
So where is the "a generation behind" feeling justified? In three places the general index glosses over:
- Factual knowledge. On the AA-Omniscience test, Wonder scores 25 and Wonder Full scores 30, against 46 for Opus 5.5 and 43 for Astra. On the other hand, the hallucination rate fell from 33% in Spark 1.3 to 23% in Wonder Full, the second lowest among the strong models after Gemini 4 Argon at 15%. The model knows less, and admits it more often.
- 3D and visual design. More on that in the next section.
- Judgment. A developer who tested it on three coding projects wrote on Hacker News that the intelligence is there, but the taste and the decisions are not there yet, and that the model kept repeating the same round of checking and fixing. That is one person's report.
Two games, $4 against $99
This is the comparison from the opening, by developer Max Blade:
Max Blade's post and comparison video on X, October 12, 2026
His conclusion is blunt: Muse is far from the frontier on 3D tasks such as games and models for 3D printing. BridgeMind, which runs a recurring test of building a lava lamp, reported that it got a red cylinder, with no glass and no lava:
BridgeMind's post on X, October 12, 2026
In the limitations section of the model card, Meta itself writes that visual and spatial understanding, and 3D tasks in particular, is a known limitation of Wonder. These two tests hit exactly the spot Meta flagged in advance. That does not refute its table, because none of the 14 rows tests 3D. It shows what the table does not test.
Now for the bill. On the Artificial Analysis index, an average task costs $4.52 on Wonder and $5.98 on Opus 5.5, which makes Wonder 24% cheaper. In Max Blade's game the gap is a factor of 24. The price per token is identical, so where does the gap come from? My explanation: from the amount of work. On a task with a defined end, both models do roughly the same amount of work. On an open-ended task like "a big-studio quality game", Opus keeps polishing and polishing, and Wonder decides it is done much earlier. This is a conclusion from a single game, so it is a hypothesis.
My opinion: Meta built a model that finishes, not a model that excels
Here I move from facts to interpretation.
I think Meta deliberately optimized for getting the task done without failures, not for the best possible result. For example: say you ask for a zombie game. Opus will work for an hour on every detail, and you will get an excellent result and a high bill. Wonder, by my estimate, will finish faster and with a much smaller bill. You will get a zombie game that works, probably without many bugs, and maybe even with some thought about what would make it fun. But it will be the most basic version of what you asked for.
Three reasons this looks deliberate to me, not accidental:
- Meta's customer is not a developer. Meta is late to the race, and its advantage is the billions of users of Facebook, Instagram and WhatsApp. The official statements also talk about an assistant, not a programmer: Zuckerberg wrote about "personal superintelligence for everyone", the first index in the table is called the Personal Superintelligence Index, and a researcher on the team explained in the launch video how much effort went into having the model remember what matters to you. The demos were a wedding seating chart, a running plan for Saturday and personalized gifts picked from Instagram photos. Those are tasks where "clean and working" is worth more than "perfect".
- Meta is the one paying for the compute. A personal assistant running for hundreds of millions of people has to be cheap to run. A model that works for an hour on every request is a product a public company cannot hand out for free.
- The features it chose to highlight. Memory across conversations, trying again when stuck, asking for approval before an important action, and fewer hallucinations. Those are the traits of an assistant you can rely on, not of a super-programmer.
And this is not just my interpretation. The day before the launch, in an interview on David Senra's podcast, Alexandr Wang, who leads Meta's AI lab, said that in early 2026 the industry saw coding agents as the most important product, and that Meta decided to step out of that race and focus on a personal agent. On launch day, on the Cheeky Pint podcast, he added that people who use Muse to write code are a minority of its users. In the same conversation, Zuckerberg said that discretion and personality matter more in a personal agent than in a coding agent for developers. In other words, the gap with Opus in programming is not a glitch that will be fixed in the next version. Meta chose it.
There is also a side to this that the benchmark tables don't measure. In the reactions I've been seeing online since the launch, the same line keeps coming up: Muse doesn't deliver results like Opus, and yet it's pleasant to work with. People describe natural language and a sense that it understands what they mean, more than other models do. This is my impression from those reactions, not measured data, but it lines up with what Zuckerberg said about personality. In a personal assistant, one you talk to every day, that counts for no less than the score.
In my view this is the right direction for today's market. Most of the tasks an AI agent does in a business on a normal day are of this kind: answer, summarize, update, schedule. On tasks like these there is nothing to polish, so there is no reason to pay for polish.
Still, three things don't fit this story for me:
- The price doesn't match the strategy. A model built for efficiency is sold at Opus 5.5's price per token. More on that in the next section.
- The good version is locked inside a developer tool. If the goal is a personal assistant, it is odd that Wonder Full is available only in Muse Code.
- Privacy. According to data Meta itself published, as reported by Handy AI and on Hacker News, Wonder mishandled personal information in about 3% of cases. For a model that is supposed to read messages and calendars, that is a lot. We covered this at length in the article on privacy in Muse.
The price went up: per token and per task
Muse Spark was a bargain model. Wonder no longer is.
Sources: Meta (API price list) and Artificial Analysis (cost per task). If the term is new to you, we explained what a token is in a separate article.
What sets the bill is not only the price per token, but how many tokens the model consumes to finish a task. According to Artificial Analysis, Wonder reads 28% more input tokens per task than Spark 1.3, and input is 76% of its cost.
Average task cost on the Artificial Analysis index, in dollars
Source: Artificial Analysis, October 12, 2026. There is no figure for Wonder Full, because it is not sold per token.
The chart shows three things. Wonder is 24% cheaper than Opus 5.5 per task, not orders of magnitude cheaper. Sonnet 5.5, which costs half as much per token, ends up more expensive than Wonder per task. And if efficiency is the point, the model that should worry Meta is GPT-6.1 Sol, at 72 cents per task.
What about the subscription? Max Blade wrote that the $4.11 run used up his quota on the $50 plan, and that he upgraded to the $200 plan to finish. That is one user's report, and Muse Code quotas reset every five hours, so there is no way to know what he had already used before that. Before you commit to a subscription for Wonder Full, run one task and check how much of the quota it ate.
Who Muse Wonder is for, and who it isn't for yet
A good fit
- Tasks with a clear end: fill in, update, summarize, schedule
- Business process automation, where it leads in third-party testing
- Long coding work on an existing project
- Tasks where "works" is enough and no polish is needed
Check first
- Work that depends on precise factual knowledge
- A Muse Code subscription: by default, your content is used for training
- Heavy use of Wonder Full: check how much of the subscription quota each task consumes
- Hebrew and other languages: Meta writes that the model was not tested in every language
Not yet
- 3D, games and visual design. A limitation Meta itself states
- Work where quality and polish are the product itself
- Legal documents without human review: even the leader scores 13 out of 100
- Sensitive customer information on the subscription in its default mode
Is it worth trying on your next task?
Five questions about one task you are thinking of giving it. After the result, run the task once and check the bill.
Where Muse Wonder stands in Sigma AI Ranking
We added Muse Wonder to Sigma AI Ranking, and for now it appears there without a score, marked "Awaiting data". Our ranking weighs only open sources that measure every model with the same method, and one day after the launch none of them has published a result yet. Epoch AI has already listed the model and has not published a score, the latest LMArena ranking is from October 8, and Terminal-Bench has no submission yet. The price and the context window are already there.
We did not want to derive a score from Meta's table. When the sources publish results, the score will be added in the update that follows. Until then, this is the reference point:
Sigma AI Ranking: weighted score, as of October 13, 2026
Source: Platin Digital's Sigma AI Ranking, which weighs open sources (Epoch AI, LMArena, Terminal-Bench and official price lists). Muse Spark 1.3 is in 21st place, mainly because of a low score for agent work: 58.6.
In our ranking, Spark 1.3 scored 93.7 on general capability and 90.6 on reliability, and fell short on agent work. If Wonder is what Meta says it is, that is where the jump should show up. Why we built our own ranking, and why it relies only on open sources, is explained in the article on Sigma AI Ranking.
What this means for a business owner
On the Artificial Analysis index, Wonder Full is tied for second place with Sonnet 5.5, at 56 against 58 for Opus 5.5. The upshot: models for serious work now come from four companies, Anthropic, OpenAI, Google and Meta. For anyone running an AI agent in a business, that means three things:
- "Good enough" and "the best" are priced differently. On the only open-ended task that was measured, the gap was a factor of 24, and on the general index it was 24%. For routine work there is no reason to pay the difference.
- Choose a model per task. A top-performing model for work where quality is the product, and an efficient model for everything else. I wrote about this in the article on Anthropic's comeback and Claude Opus 5.5: within three months the leading model changed twice.
- Don't buy based on the vendor's table. Check who measured each row, and test on one real task of your own, with a stopwatch and the bill in front of you.
Platin Digital is an OpenAI Select Partner, and we build the agents we set up with the models and tools that best fit each business, from OpenAI, Anthropic, Google and others. Wonder is joining our list of models we test on routine tasks. For tasks where quality decides, I am staying with Opus 5.5 for now.
Further reading: the full guide to Meta Muse, Muse for Small Business, Muse vs. OpenAI's dots and ChatGPT, Gemini or Claude for your business.
Frequently asked questions
What is Muse Wonder?
Muse Wonder is Meta's family of large models, released on October 12, 2026 in two versions: Muse Wonder, sold through the API, and Muse Wonder Full, the stronger version, available only on a Muse Code subscription. Both accept text, images, video and audio, with a context window of one million tokens, and were built for agent work: long tasks, code and tool use.
How much does Muse Wonder cost?
In Meta's API, Muse Wonder costs $4 per million input tokens, $20 per million output tokens and $0.20 for cached input. That is the same price as Claude Opus 5.5, and 3.2 to 4.7 times the price of Muse Spark 1.3. Muse Wonder Full has no per-token price: it is included in the Muse Code subscriptions at $50 and $200 a month.
Is Muse Wonder better than Claude Opus 5.5?
On the independent Artificial Analysis index, Opus 5.5 leads with 58, against 56 for Wonder Full and 52 for Wonder. In Meta's table, Wonder Full is above Opus 5.5 in eight rows out of fourteen, mainly in long coding work and business automation, and Opus 5.5 leads in health, computer use, Terminal-Bench and ProgramBench. In one published comparison, building a 3D game, Opus 5.5 delivered a much better result at 24 times the price.
Why isn't Muse Wonder Full available in the API?
Meta has not published a reason. In practice, Wonder Full is available only inside Muse Code, Meta's coding tool, on the Power and Maximum plans. What that means for anyone building a product or an agent: the model you can connect to your systems is the standard Wonder, and most of the standout numbers from the launch belong to the version you cannot connect.
Can I use Muse Wonder in Israel?
The Wonder API is sold through Meta and through OpenRouter. The Muse app itself is still not available in Israel, and Meta has not published a date. Full details are in the article on Muse in Israel.
Is my data used to train Muse Wonder?
It depends on how you access it. According to Meta, content sent through the pay-per-use API is not used for training. On a Muse Code subscription, the default is a mode in which Meta may use prompts, code and files to improve future models. You can turn this off in the settings, and turning it off applies only to what is sent afterward.
Sources
- https://research.meta.ai/blog/introducing-muse-wonder/
- https://dev.meta.ai/resources/blog/build-with-muse-wonder
- https://research.meta.ai/static/muse-wonder-model-card/
- https://dev.meta.ai/docs/muse-code/subscriptions/
- https://x.com/finkd/status/2109689305073614918
- https://x.com/finkd/status/2109689307288240154
- https://artificialanalysis.ai/articles/muse-wonder
- https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v4-3
- https://x.com/ArtificialAnlys/status/2109697083519447118
- https://techcrunch.com/2026/10/12/meta-releases-its-first-frontier-model-muse-wonder/
- https://venturebeat.com/technology/metas-new-muse-wonder-and-muse-wonder-full-models-achieve-frontier-performance-but-only-the-weaker-one-is-available-via-api
- https://the-decoder.com/metas-new-frontier-model-muse-wonder-full-breaks-into-the-ai-top-tier/
- https://www.bloomberg.com/news/articles/2026-10-12/meta-says-new-muse-model-reaches-ai-frontier-after-14-month-push
- https://handyai.substack.com/p/model-drop-muse-wonder
- https://news.ycombinator.com/item?id=50059115
- https://x.com/_MaxBlade/status/2109723440710066646
- https://x.com/bridgemindai/status/2109733498114417013
- https://x.com/theo/status/2109724741234028639
- https://openrouter.ai/meta/muse-wonder
- https://epoch.ai/benchmarks
- https://www.itiger.com/news/1153188847
- https://cheekypint.substack.com/p/muse-wonder-zuckerberg-and-wangs
