Muse Wonder Full scores 56 on the Artificial Analysis Intelligence Index and Muse Wonder scores 52, against 58 for Claude Opus 5.5, 56 for Claude Sonnet 5.5, and 53 each for GPT-6 Astra, Gemini 4 Argon and Claude Fable 5.1. Meta’s own model card tells a stronger story: leads on long-running coding, legal work, business automation and memory. Both are true, because they measure different things, and the gap between them is the most useful fact in this post.
Here is every published number, Meta’s and independent, with what each one tests, where Wonder leads, where it trails, and which figures nobody outside Meta can reproduce. For the model itself, see the Muse Wonder explainer.
Meta’s model card table
Meta compared Muse Wonder Full, Muse Wonder, GPT-6 Astra and Claude Opus 5.5 at maximum reasoning effort. The rows, via VentureBeat’s report on the card:
| Benchmark | What it measures | Wonder Full | Wonder | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|---|
| Artificial Analysis Intelligence Index | Ten-eval composite | 56 | 52 | 53 | 58 |
| Harvey LAB-AA | Legal agent tasks | 13.1 | 10.0 | 8.6 | 4.2 |
| AutomationBench-AA | Business process automation | 45.4 | 39.1 | 41.6 | 41.2 |
| JobBench | Professional work | 75.3 | 70.5 | 66.1 | 75.4 |
| DeepSWE 1.1 | Software engineering | 79.0 | 78.8 | 74.1 | 74.2 |
| SWE-Marathon | Long-running coding runs | 57.5 | 53.1 | 42.5 | 46.3 |
| Terminal-Bench 4.0-AA | Terminal agent tasks | 57.1 | 52.0 | 59.1 | 59.6 |
| HealthBench Pro | Health questions | 62.3 | 60.5 | 65.3 | 65.6 |
Read across and the shape is clear. Wonder Full wins where a task is long and agentic: SWE-Marathon by 11 points over Opus 5.5, DeepSWE by five, legal by a factor of three. It loses on terminal work, on health, and on the composite index. Even on Meta’s own table, Opus 5.5 is the more intelligent model in aggregate; Wonder Full is the more persistent one.
Two caveats Meta itself states. Competitor runs were best-effort, with settings that may not have been optimal for those models. And some coding comparisons used different tooling, so those rows are less controlled than the rest.
Meta’s internal benchmarks: memory and focus
Three figures in Meta’s launch material come from tests only Meta can run.
| Benchmark | Wonder Full | Wonder | GPT-6 Astra | Claude Opus 5.5 | Claude Fable 5.1 |
|---|---|---|---|---|---|
| MemoryBench | 85.5 | 85.6 | 81.4 | 81.5 | 79.0 |
| FocusBench | 54.2 | 44.9 | 47.7 | 45.8 | 36.6 |
MemoryBench hides facts inside 64K to 100K-token conversations, adds distractors, and changes some facts halfway through; the model has to recall the current version. It is the evidence for the Active Memory feature. FocusBench measures whether a model keeps pursuing a goal through blockers, the Active Focus feature. Note that the smaller Wonder scores below Astra and Opus on FocusBench; only Full leads. A third internal figure, the Personal Superintelligence Index, has no published competitor comparison.
These are plausible tests of real behaviors, and they are also exactly the behaviors Meta trained for. Until someone outside Meta can run them, treat them as a description of the training target.
The independent scores: Artificial Analysis
Artificial Analysis ran both models through its Intelligence Index v4.3 at max effort. The composite and the parts:
| Evaluation | Wonder Full | Wonder | Muse Spark 1.3 | Notable comparison |
|---|---|---|---|---|
| Intelligence Index | 56 | 52 | 48 | Opus 5.5 58, Sonnet 5.5 56, Astra / Argon / Fable 5.1 53, Grok 4.6 44 |
| Terminal-Bench 4.0 | 57% | 52% | 33% | Gemini 4 Argon 57% |
| GDPval-AA (Elo) | 1857, #2 | n/a | 1681 | Opus 5.5 1866, Sonnet 5.5 1838 |
| AA-Briefcase (Elo) | 1729, #3 | n/a | 1581 | |
| CritPt | 33%, #1 | n/a | n/a | Nearest rivals 31 to 32% |
| GDP.pdf | 33%, #1 | n/a | n/a | Nearest rivals 31 to 32% |
| SciCode | 63%, #3 | n/a | n/a | |
| AA-Omniscience (knowledge) | 30 | 25 | 25 | Opus 5.5 46, Astra 43, Argon 42, Sonnet 5.5 32 |
| Hallucination rate | 23% | 27% | 33% | Gemini 4 Argon 15% |
| Coding Agents Index | 68 | 63 | 54 | Claude Code with Sonnet 5.5 68, Opus 5.5 66 |
Three things stand out.
The jump from Spark is real. Eight points on the composite in six weeks, and Terminal-Bench nearly doubled. Artificial Analysis called it a material jump in intelligence over previous Meta models.
Knowledge is the weak spot. An Omniscience score of 30 against 46 for Opus 5.5 means Wonder knows fewer facts, though its low hallucination rate means it is less likely to invent them. For retrieval-heavy products, that argues for grounding with search or your own documents rather than relying on recall.
Several leads are narrow. CritPt and GDP.pdf are one to two points ahead of the field, and the GDPval lead over Sonnet 5.5 is within confidence intervals. The Coding Agents Index tie with Claude Code on Sonnet 5.5 is the most relevant number for anyone choosing an agent runtime, and it belongs to Full, not the API model.
The row where Meta and Artificial Analysis disagree
Meta’s model card lists Wonder Full at 59.0 on an Artificial Analysis Intelligence Index row. Artificial Analysis publishes 56. The difference likely comes from index versions, run dates, or Meta’s own evaluation harness, and neither side has explained it publicly. When you see 59 quoted, it is Meta’s figure; when you see 56, it is the independent one. The same gap shows up on AutomationBench: Meta’s table has 45.4 and 39.1, Artificial Analysis reports 71% and 68% on what appears to be a different version of the test. Quote the source with the number.
Tokens and cost behind the scores
Scores at max effort come with a token bill. Artificial Analysis measured about 54,000 output tokens per task for both Wonder models, against 62,000 for Argon, 78,000 for Fable 5.1 and 27,000 for GPT-6 Astra. On input, Wonder used 7.69 million tokens across the index, Full 6.44 million, two to seven times more than the OpenAI and Google models. That is why Wonder’s cost per task landed at $4.52, below Opus 5.5 at $5.98 but above Astra at $3.26 despite Astra’s 2.5x higher list price. The pricing post works through what that means for a real workload.
What to test yourself
Benchmarks tell you where to look, not what to ship. For the agentic strengths Meta claims, the right test is your own long task: a 30-step refactor, a multi-document review, a tool loop with deliberate failures injected. In Apidog, set up the same prompt and tool definitions against Muse Wonder, Opus 5.5 and whatever you run today, assert on the structured output, and compare usage on each. The tool-call test harness and the non-deterministic agent testing guide cover the setup. Download Apidog to run the three side by side.
Remember that the API serves Wonder, not Full. Any benchmark headline you read about Meta’s frontier model probably describes the one you cannot call.
Frequently asked questions
Is Muse Wonder the best model available? Not on the independent composite. Opus 5.5 scores 58 to Full’s 56 and Wonder’s 52. Wonder Full leads specific agentic benchmarks: SWE-Marathon, DeepSWE, Harvey LAB, FocusBench.
Which benchmarks can I trust? Artificial Analysis runs are reproducible and third-party. Meta’s cross-model table is Meta-run with best-effort competitor settings. MemoryBench, FocusBench and the Personal Superintelligence Index are Meta-only.
How does it compare with Gemini 4 Argon? Argon scores 53 on the index to Wonder’s 52 at half the price, ties Full on Terminal-Bench at 57%, and hallucinates less (15% vs 27%). Our Argon benchmarks post has its full table.
Does max effort change the model? No. Effort changes how long Wonder reasons; it never turns muse-wonder-1.0 into Muse Wonder Full.
Next step
Pick the one benchmark closest to your workload, SWE-Marathon for long coding runs or Omniscience for factual recall, and let it decide whether Wonder gets a trial. The composite number will not.
