Matthew Berman's first Muse Wonder test starts with a working Rubik's Cube, then quickly turns into a less tidy story: subscription upgrades, HTTP 429 errors, a playable game with visual glitches, and 3D scenes that do not meet their photorealistic brief. Meta's October 12 launch is significant, but this video is more useful as an early reality check than as a winner-takes-all model ranking.
Muse Wonder and Muse Wonder Full are Meta's new frontier-oriented models for coding and long-running agents. Matthew's Rubik's Cube is the strongest visible build in this short test; his other game and 3D outputs need more work. Wonder is available on the Meta Model API and in Muse Code; Wonder Full is limited to qualifying Muse Code subscriptions. Independent benchmark results show a substantial leap over Muse Spark, but neither benchmark scores nor a few demos establish which model is best for your own job.
Watch Matthew Berman Test Muse Wonder
Credit: The hands-on tests and opinions are Matthew Berman's. Product access, specifications, and pricing below come from Meta's model page and documentation. Benchmark context comes from Artificial Analysis and Meta's model card.
Muse Wonder and Wonder Full Are Different Products
Muse Wonder
Where it runs
Available on the Meta Model API as muse-wonder-1.0 and in Muse Code. Meta's documentation calls the API model a preview. Its context window is about one million tokens; it accepts text, images, audio, and video and returns text.
What it costs
Meta lists $4 per million input tokens, $0.20 per million cached input tokens, and $20 per million output tokens on the API. A subscription is a different billing route, subject to usage limits. Max reasoning effort is still Wonder, not Wonder Full.
Muse Wonder Full
Where it runs
At launch, only in Muse Code on the Power Usage and Maximum Usage subscription plans. Meta does not offer Full as a pay-per-token Model API model. It shares the one-million-token context and multimodal input, text-output profile.
What to check
Meta's model page lists Power Usage at $50/month and Maximum Usage at $200/month. Those are plan prices, not a guarantee of uninterrupted Full usage. Check the current quota, model selector, and data mode before moving an important workload.
Meta calls Active Memory the model's ability to retrieve relevant context across conversations, and Active Focus its ability to keep working through blockers and ask for clarification. Those are Meta's capability descriptions; Matthew's brief video does not independently verify long-term memory or hours-long agent reliability. Meta says Wonder will power more of the consumer Muse experience later, beginning with subscribers, so current Muse Code access should not be confused with a full consumer rollout.
What Matthew's Four Demos Actually Show
1. Rubik's Cube
Visible result
He rotates faces, scrambles the cube, and runs a solve. The face colors appear consistent through the interaction, and he says the build took about four minutes.
Practical limit
A working, attractive demo is a useful signal for interactive coding. It is one creator run, not a repeatability test or proof that the solver handles every valid state.
2. Obstacle-course game
Visible result
The Fall Guys-inspired game runs, but Matthew points out flickering and clipping. He says other models produced stronger versions for this particular brief and was still iterating on this one.
Practical limit
Judge game feel, collision behavior, and stability after several runs, not just a screenshot or a first launch. A later revision could change the result.
3. Three.js boat and ocean
Visible result
Water reflections and time-of-day controls impress him. He also sees white artifacts near the boat, a weak shadow, and wind controls that affect water speed more clearly than wave size.
Practical limit
The brief asked for hyperrealism. Interactivity and attractive reflections do not make the finished scene physically convincing. Inspect the rendered scene and controls separately.
4. Watermelon impact
Visible result
The browser-based 3D impact is watchable, but the image blurs and does not satisfy Matthew's hyperrealistic target. He suggests testing a different production environment next.
Practical limit
This does not prove web graphics cannot be photorealistic; it only shows that this output missed this brief. A fair comparison would hold the prompt, tools, revision budget, and review criteria constant.
At 03:20, Matthew also shows an unrelated and unsettling sentence that appeared in one model response. It is worth logging as an anomalous output, not interpreting as evidence about the model's intentions or treating a single incident as a measured failure rate.
Strong Scores, but Read the Benchmark Labels
Artificial Analysis reports Intelligence Index v4.3 scores of 52 for Wonder and 56 for Wonder Full at maximum reasoning effort, versus 48 for Muse Spark 1.3. The same release snapshot puts Claude Opus 5.5 at 58. These are index points, not percentages of tasks solved. The independent assessment also notes Wonder Full's gains on agentic knowledge work and coding, while finding weaker general knowledge than some rivals.
Meta's model card lists DeepSWE 1.1 at 79.0% for Full and 78.8% for Wonder. It identifies the Muse scores as third-party Datacurve results, but its comparison table mixes external runs, provider-reported figures, and some internal runs. On OSWorld V2.1, the displayed 78.4 for Full and 72.6 for Wonder are partial-reward scores; they are not strict task-completion percentages. Meta separately lists lower binary completion figures. Its Personal Superintelligence Index and FocusBench are internal evaluations. The model card is unusually explicit about these methods, so read the metric and harness before repeating any rank claim.
The independent cost-per-task analysis estimates $4.52 per Intelligence Index task for Muse Wonder at max effort, compared with $0.72 for GPT-6.1 Sol in that evaluator's setup. That is a benchmark-specific workload, not what every request will cost. Wonder Full has no comparable API cost-per-task number because Meta has not put it on the per-token API.
Price, Quota, and the Data Setting
Matthew says he started on a $15/month Muse Code plan, upgraded to $50 to reach Wonder Full, hit repeated 429 errors, then tried the $200 plan and still encountered access trouble. This documents his launch-day experience, not a stable quota for all customers. A 429 can reflect plan limits or service-side rate limiting; the video does not isolate which applied to every failed request. Meta's subscription docs say users who reach a plan limit can wait for the refresh or upgrade, while its API docs describe separate rate limits and retry guidance.
There is a second setting worth checking before testing client work. Meta says Muse Code subscriptions default to Contributor mode, which provides more usage in exchange for allowing submitted prompts, code, files, and responses to help train future Meta models. Subscribers can turn it off; Standard mode does not use that content for training, according to Meta's model documentation. Confirm the account setting and applicable terms before sending proprietary code or private data.
Who Should Test It Now?
If you build interactive software or run long coding tasks, Wonder is worth a controlled test against your current model. Start with an authorized, non-sensitive project and a clear acceptance test. Measure completed work, human corrections, time, and cost, then repeat after quota resets. Wonder Full may be attractive if you already use Muse Code's higher plans; the API version is the relevant comparison for production applications that need direct model calls. For a business-critical workflow, keep a fallback model and a human approval step until you have evidence that output quality and availability hold up on your own tasks.
Jump to Matthew's Test
| Time | Segment | Time | Segment |
|---|---|---|---|
| 00:00 | Introducing Muse Wonder | 01:03 | Plans and quota errors |
| 01:30 | Obstacle-course game | 01:54 | Boat and ocean scene |
| 02:43 | Watermelon impact | 03:20 | Anomalous response |
| 04:14 | Pricing and model access | 05:35 | Benchmark chart |
| 08:09 | Independent index and task cost | 09:43 | Value judgment and verdict |
Business Idea: A One-Workflow Model Bake-Off
Businesses do not need a generic “best AI model” report. They need to know whether one repetitive task can be completed accurately, affordably, and reliably under their own approval rules. A small, evidence-backed comparison is a clearer service than selling access to whichever model topped this week's chart.
Test one real workflow
Turn a model comparison into a seven-day business pilot.
Sources and Limits
Hands-on source: Matthew Berman's video. Official Meta sources: launch announcement, research overview, model card, model and plan page, and subscription/data-mode documentation. Independent evaluator: Artificial Analysis launch assessment. Pricing, quotas, and leaderboards can change. The four creator demos are not a controlled comparative benchmark; model-card results should be read with their stated metrics and harnesses.