If Muse Wonder felt weaker than the launch page promised, the explanation may be simple: you tested the API model, not Muse Wonder Full. Full sits behind Muse Code plans and scored 56 on the independent intelligence index, while the per-token API model scored 52.
Muse Wonder full vs API: which model does each score belong to?
Muse Wonder Full scored 56 on Artificial Analysis's Intelligence Index v4.3, and the model served through the standalone API scored 52. Full is available only inside Muse Code subscriptions, while the API model carries per-token pricing. Meta announced both together in its launch post dated October 12, 2026.
That split is the single fact that resolves most of the confusion around the launch. Meta's announcement described state-of-the-art performance on agentic tasks, and the charts behind that claim describe Full. Developers who signed up for API keys on day one received the other model.
The speaker in the source video put it plainly: if the model felt weaker than the launch page promised, you most likely tested the weaker one. Hacker News commenters reached the same conclusion, pointing out that the marketing numbers belong to Full.
| Variant | Score (v4.3) | Access | Pricing |
|---|---|---|---|
| Muse Wonder Full | 56 | Muse Code subscription | No per-token price |
| Muse Wonder (API) | 52 | Standalone API | $4 per million input tokens, $20 per million output |
| Claude 5.5 | 58 | API | Same list price as the API model |
The table above reports scores from Artificial Analysis as described in the video; the index version matters, and the next section explains why.
Who ran the benchmark, and on which index version?
Artificial Analysis ran the benchmark on Intelligence Index version 4.3, which bundles 10 evaluations into one score. Both Muse Wonder variants ran at maximum effort on that version. Meta's own scorecard uses different tests, so the two sets of numbers answer different questions.
Version drift can outweigh the gaps everyone argues about. Fable 5.1 appears at 66 in an earlier Artificial Analysis article and at 53 on the current release page because the index version changed between those publications. That is a 13-point swing, more than six times the two-point gap between Full and Claude 5.5.
A score quoted without its index version is unusable for comparison. The same applies to effort level and to whether a laboratory or an independent tester produced the number.
Meta's own SWE Marathon figures, as VentureBeat reported them, put Full at 57.5 against Opus 5.5 at 46.3. On the independent index, Opus leads overall. Both statements can be true because the tests differ.
A developer-community write-up counted 14 Meta benchmarks and concluded the API model loses to Opus 5.5 on nine of them. Treat that count as secondhand; the video presenter did.
Why benchmarks can win in finance and still feel weak in coding
The index favors office, finance, and science tasks, and Muse Wonder Full landed in the top three on six of 10 evaluations. That is a real result on the measured mix. It says less about a quick image-to-HTML attempt, which resembles few items on the list.
Full took first place on crypt and GDP.pdf, second on GDPval, and third on three further tests. Most of those are document and analysis work. An index is a report card with 10 exams squeezed into one grade, and weights decide how much each exam counts.
Individual coding reports in the launch thread described a gap. One tester converted an image to HTML and found the model closer to Sonnet or Haiku than Opus 5.5. Another, working on a React front end and Node back end, placed it below GPT-6 Astra, Opus 5.5, and Fable.
Those are single tests from individual users, so they say little about all coding or all vision work. The video presenter made the same caveat and noted that Artificial Analysis does publish a separate coding index.
On that coding agents index, run inside Muse Code at maximum effort, Full scored 68.
How much does effort level move a score?
Effort level moves a score by as much as six points inside a single model, which is larger than the gap between Muse Wonder Full and Claude 5.5. Artificial Analysis lists Fable 5.1 at 53 on max effort, 51 on high, 49 on medium, and 47 on low.
Both Muse Wonder runs used the maximum setting, and the token counts show the cost of that choice. Full produced 140 million output tokens during the index run against a median of 88 million, roughly 59% more. The API model wrote 130 million against 85 million.
A chart that omits its effort setting cannot be compared fairly with one that names it. When a model's score sits two or four points from a rival, the setting can decide the ranking.
Artificial Analysis's page for Full lists 56 at max effort, fourth among 689 models, and states that Full has no publicly available API. That line, more than any chart, tells you what you are buying.
What the API model costs per task and per token
Artificial Analysis measured the API model at $4.52 per index task, against $3.26 for GPT-6 Astra and $1.99 for Gemini 4 Argon. Output speed is 115 tokens per second, and the first token takes about 20 seconds against a platform average of 5.52 seconds.
Subtracting Astra's cost from the Muse model's leaves $1.26 more per task, and the API model ranks 103rd of 112 in its price tier. Muse Wonder Full has no per-token price because it exists only inside Muse Code, whose plans run from 50 to 200 dollars a month.
The list price of the API model, $4 per million input tokens and $20 per million output tokens, matches Claude 5.5. Equal list price does not mean equal cost per finished task.
Time to first token matters in agent loops, where a 20-second wait lands on every turn. The video presenter framed that delay with a joke about rereading your prompt, and the number stands on its own.
Neither of us can reduce that latency by choosing a different effort setting on the API; the measured figure is what Artificial Analysis recorded for the index run.
What Muse Code subscribers and API developers should each believe
Muse Code subscribers get Muse Wonder Full at 56, roughly two points behind Claude 5.5 on the independent index as of mid-October 2026. API developers get the 52 model, which trails Opus 5.5, Sonnet 5.5, Fable 5.1, and GPT-6 Astra on the same version.
The subscription also defaults differently on data use. Muse Code starts in a contributor mode that lets Meta use prompts and code to improve its models, while the API defaults to standard terms with no training use. The video presenter drew that distinction from Meta's documentation; check the current terms before you rely on either default.
Neither model has downloadable weights, according to VentureBeat's reporting. That means no local deployment path and no inspection of the two variants to see how they differ.
Meta has not said whether the gap comes from size, training, or a different checkpoint. The scores remain the only public evidence that two distinct models exist behind one name.
A one-minute check for any model launch chart
Three questions settle most launch-chart confusion: which model the number belongs to, which effort level produced it, and who ran the test. Each has a documented example behind it in this launch.
- Name the exact variant. Muse Wonder Full and the API model share a brand but sit at 56 and 52.
- Name the effort level. Fable 5.1 alone moves six points between low and max.
- Name the operator. A laboratory's own scorecard and an independent index answer different questions.
- Name the index version. Fable 5.1 moved 13 points between two Artificial Analysis publications when the version changed.
After those four, run your own task on the model you can actually reach. Both the index and the Hacker News tests measure someone else's work, not your codebase.
FAQ
- Is Muse Wonder Full the same model as the API version? No. Full scored 56 on Artificial Analysis's Intelligence Index v4.3 and is available only through Muse Code subscriptions. The standalone API serves a separate model that scored 52 on the same index version.
- Which Muse Wonder model is better for coding? Full has the stronger published figures, including 68 on Artificial Analysis's coding agents index run inside Muse Code at maximum effort. The API model trails several rivals on the general index, and individual coding reports favored competitors over both.
- Why does Muse Wonder Full have no per-token price? Full is not offered through the API at all. Artificial Analysis's page lists it as having no publicly available API, so the only way to use it is a Muse Code plan priced between 50 and 200 dollars a month.
- Does effort level change a model's benchmark score? It can change it by several points. Fable 5.1 scored 53 at max effort and 47 at low effort on the same index, a six-point spread that exceeds the gap between Full and Claude 5.5.
- Should I trust Meta's SWE Marathon numbers over the independent index? They measure different tasks. Meta's reported figures put Full ahead of Opus 5.5 on SWE Marathon, while the independent index puts Opus ahead overall. Use the benchmark that resembles your workload.
Turning launch material into articles you can defend
A launch like this one produces hours of video, a pile of charts, and a comment thread full of first-hand tests. The useful move is to keep the three checks intact when you write it up: which model, which effort level, who ran the test.
Teams that publish explainers, comparisons, and technical breakdowns often already have that material recorded. Skala Blog turns a public YouTube transcript into a private draft you review before anything goes live, with the source links and figures kept where a reader can check them. A shared vocabulary around tools like Crazystack TypeScript and community projects covered by voices such as Dev doido makes those drafts easier to adapt for a developer audience.
If a video on Crazystack or a conference talk already carries the explanation you want to publish, Skala Blog can draft the article from it while you keep editorial control.
Fork this article
Start a new branch from the same video, shaped your way. You keep the credit; the original keeps the attribution.
A fork in another language is filed as a translation of this article, so the two pages point at each other. You can unlink it later from the editor.
0/240
You are creating
- Format
- For
- Language
- Source
- Your angle
You will be asked to sign in before it is generated.
Buy credits