ModelsArchitectures & capability
Meta launches Muse Wonder models, keeping top-scoring Wonder Full out of its API
Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full, which score 52 and 56 on Artificial Analysis's Intelligence Index. Developers can call Wonder at $4/$20 per million tokens. The stronger Full model is limited to paid Muse Code plans.
From the Simon Willison article of October 12, 2026 and cited sources: On Artificial Analysis's Intelligence Index v4.3, Wonder Full scored 56 and Wonder scored 52, both at max reasoning effort. [2] [3]
SOURCES[2] Artificial Analysis: Muse Wonder article artificialanalysis.ai[3] VentureBeat: Only the weaker one is available via API venturebeat.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- On Artificial Analysis's Intelligence Index v4.3, Wonder Full scored 56 and Wonder scored 52, both at max reasoning effort. [2] [3]
- Original source
- Simon Willison article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 12, 2026 article published
- Covered
- Oct 15 edition
- Evidence
- 9 sources across 8 sites
- 1Simon Willison: Release llm-meta-ai 0.2simonwillison.net
- 2Artificial Analysis: Muse Wonder articleartificialanalysis.ai
- 3VentureBeat: Only the weaker one is available via APIventurebeat.com
- 4Meta Newsroom: Introducing Muse Wonder and Muse Wonder Fullabout.fb.com
- 5TechCrunch: Meta releases its first frontier model, Muse…techcrunch.com
- 6GitHub: simonw/llm-meta-ai 0.2github.com
Reach: not measured yet
On Artificial Analysis's Intelligence Index v4.3, Wonder Full scored 56 and Wonder scored 52, both at max reasoning effort. Muse Spark 1.3 scored 48 about six weeks earlier. Wonder Full improved on all ten component evaluations and cut measured hallucination to 23%, down from 33%. VentureBeat reports that Claude Opus 5.5 still leads at 58. [2] [3]
Access is split. Muse Wonder is on the Meta Model API at $4 input, $0.20 cached input and $20 output per million tokens. It has a 1M-token context, accepts text, image, speech and video input, and outputs text. Wonder Full is available only on $50–$200 monthly Muse Code plans. Neither model's weights can be downloaded. Simon Willison's llm-meta-ai 0.2 plugin added support on launch day, including max reasoning effort. [2] [3] [4] [7] [1] [6]
Running costs are high. At max effort, Wonder averaged about 7.69M input tokens per Artificial Analysis task, or $4.52, compared with $1.60 for Spark 1.3. That makes caching important. Time to first token is 41.4 seconds, against a 4.04-second median for its price tier, which suits background agents better than interactive use. [2] [7]
Meta's state-of-the-art claims on LAB, FocusBench and SWE-Marathon come from Meta itself and are published without numbers. VentureBeat notes that Meta's benchmarking mixes results run under inconsistent conditions. By default, Muse Code lets Meta train on user submissions, and zero-retention options limit the memory features. [8] [3]
Artificial Analysis scores put Meta near the frontier, and Wonder's API price matches Opus 5.5.
Read the full assessment
Implication: agent builders gain another credible vendor, but the strongest model requires a subscription and the slow first token favors background workloads.
Executive brief
Meta's new Muse Wonder Full model scored 56 on Artificial Analysis's Intelligence Index, up 8 points from Muse Spark 1.3 (48), which shipped about six weeks earlier. That puts Meta Superintelligence Labs close to the frontier. The most capable version, Wonder Full, is not sold through the API: only paying Muse Code subscribers can use it. Developers can call the standard Muse Wonder at $4/$20 per million input/output tokens. Simon Willison updated his llm-meta-ai plugin to support it on launch day and noted that Meta is releasing roughly one model a month.
What changed and event timeline
Muse 1.2 released
This is the earlier baseline in Willison's pelican-drawing comparison. The date comes from his post, and no primary Meta source for it appears in the reviewed sources.
Muse 1.3 released
Willison calls it "Muse 1.3." Artificial Analysis appears to mean the same model when it says "Muse Spark 1.3" with a six-week gap before Wonder.
Meta announces Muse Wonder and Wonder Full
Meta calls them its "most capable" models, built for agentic work. They were trained at the Prometheus data center and include "Active Memory" and "Active Focus" features.
More detail
Artificial Analysis publishes its scores
On Intelligence Index v4.3, Wonder Full scores 56 and Wonder scores 52, both at max reasoning effort. Wonder Full improved on all ten component evaluations.
Llm-meta-ai 0.2 ships
The release notes say "New models. Willison generated pelicans at six effort levels.
More detail
Capabilities and access
- Muse Wonder: available through the Meta Model API and Muse Code (Meta).
- Muse Wonder Full: only on Muse Code plans costing $50–$200/month, with no API access (VentureBeat).
- Weights: both models are proprietary and can't be downloaded or self-hosted (VentureBeat).
Read the full section
- Muse Wonder: available through the Meta Model API and Muse Code (Meta). It costs $4 per million input tokens, $0.20 for cached input and $20 for output (Artificial Analysis). It has a 1M-token context window, accepts text, image, speech and video input, and outputs text only (AA model page).
- Muse Wonder Full: only on Muse Code plans costing $50–$200/month, with no API access (VentureBeat).
- Weights: both models are proprietary and can't be downloaded or self-hosted (VentureBeat).
- Reasoning effort: six levels, from minimal to max (Willison).
Technical analysis for researchers and developers
- Architecture: not disclosed. Meta says only that the models use "larger, stronger base models" and scaled pretraining plus reinforcement learning (Meta research blog).
- Cost per task: at max effort, Wonder used about 7.69M input and 54k output tokens per Artificial Analysis task, costing $4.52 versus $1.60 for Spark 1.3 (Artificial Analysis).
- Latency: time to first token is 41.4s, against a 4.04s median for its price tier. Output speed is about 154 tokens/s (AA model page).
Read the full section
- Architecture: not disclosed. Meta says only that the models use "larger, stronger base models" and scaled pretraining plus reinforcement learning (Meta research blog).
- Cost per task: at max effort, Wonder used about 7.69M input and 54k output tokens per Artificial Analysis task, costing $4.52 versus $1.60 for Spark 1.3 (Artificial Analysis). Heavy input use means caching matters a lot for cost.
- Latency: time to first token is 41.4s, against a 4.04s median for its price tier. Output speed is about 154 tokens/s (AA model page). That suits background agents more than interactive use.
- Reproducibility: Wonder Full's index score comes from runs inside Muse Code, so outside teams can't reproduce it through the API (Artificial Analysis).
Claims and evidence
- Index scores of 56 (Full) and 52 (Wonder): these come from Artificial Analysis, a third-party evaluator.
- Willison says Artificial Analysis rates Wonder (Max) as equivalent to GPT-6.1 Sol (Willison).
- "State-of-the-art" results on LAB, FocusBench and SWE-Marathon: these are Meta's own claims, and the blog post gives no numbers (Meta research).
Read the full section
- Index scores of 56 (Full) and 52 (Wonder): these come from Artificial Analysis, a third-party evaluator. VentureBeat reports that Claude Opus 5.5, at 58, scores higher and that Sonnet 5.5 ties Wonder Full (VentureBeat).
- Comparisons conflict. Willison says Artificial Analysis rates Wonder (Max) as equivalent to GPT-6.1 Sol (Willison). TechCrunch says the models are comparable to Fable 5.1 and GPT-6 Astra (TechCrunch). Artificial Analysis's own model page ranks Wonder (Max) 9th of 216 (AA).
- "State-of-the-art" results on LAB, FocusBench and SWE-Marathon: these are Meta's own claims, and the blog post gives no numbers (Meta research).
- MemoryBench lead: reported by TechCrunch. No independent corroboration was found.
Context and prior work
Muse Wonder is Meta Superintelligence Labs' first model billed as frontier-class, following the low-latency Muse Spark line and Muse Agent (TechCrunch). Willison describes them as possibly Meta's first "Fable class" models and judges the month-to-month improvements, at least in his pelican-drawing test, to be material (Willison).
Read the full section
Muse Wonder is Meta Superintelligence Labs' first model billed as frontier-class, following the low-latency Muse Spark line and Muse Agent (TechCrunch). It also continues Meta's move away from open weights: the smaller Muse Glimmer stays downloadable, but the Wonder models do not (VentureBeat). Willison describes them as possibly Meta's first "Fable class" models and judges the month-to-month improvements, at least in his pelican-drawing test, to be material (Willison).
Limitations, safety and contested findings
- Hallucination: Artificial Analysis measures 23% for Wonder Full and 27% for Wonder, improved from 33% for Spark 1.3 (Artificial Analysis).
- Safety documentation: Meta cites its "Superintelligence Scaling Framework" and CBRN/cyber safeguards, but the full safety report is still "forthcoming" (Meta research).
- Benchmark methodology: VentureBeat says Meta's benchmarking "deserves scrutiny" because it mixes internal, vendor and third-party results run under inconsistent conditions (VentureBeat).
Read the full section
- Hallucination: Artificial Analysis measures 23% for Wonder Full and 27% for Wonder, improved from 33% for Spark 1.3 (Artificial Analysis).
- Safety documentation: Meta cites its "Superintelligence Scaling Framework" and CBRN/cyber safeguards, but the full safety report is still "forthcoming" (Meta research).
- Benchmark methodology: VentureBeat says Meta's benchmarking "deserves scrutiny" because it mixes internal, vendor and third-party results run under inconsistent conditions (VentureBeat).
- Data use: Muse Code defaults to "Contributor mode," which lets Meta train on submissions. Zero-retention options restrict the memory features (VentureBeat).
Business and practitioner implications
- Pricing: Wonder's API price matches Opus 5.5 and is below GPT-6 Astra (VentureBeat).
- Subscription lock-in: keeping Wonder Full inside Muse Code means the best version comes through a subscription, not the API.
- Before adopting: check Contributor-mode defaults, budget for the long time to first token, and re-run evaluations each month given how fast Meta is releasing models.
Read the full section
- Pricing: Wonder's API price matches Opus 5.5 and is below GPT-6 Astra (VentureBeat). That makes Meta a credible fourth vendor to consider for agent workloads.
- Subscription lock-in: keeping Wonder Full inside Muse Code means the best version comes through a subscription, not the API. CNBC frames the launch as a test of whether users will pay (CNBC).
- Before adopting: check Contributor-mode defaults, budget for the long time to first token, and re-run evaluations each month given how fast Meta is releasing models.
- Quick trial: the
llm-meta-ai0.2 plugin offers a fast way to try the models across reasoning levels (GitHub).