ModelsArchitectures & capability
Meta releases Muse Wonder and Muse Wonder Full, its first frontier-class AI models
Meta's Muse Wonder is available through its API at $4/$20 per million tokens. The stronger Wonder Full is limited to Muse Code subscriptions. Both target long-horizon coding and persistent memory, but they rank below Claude Opus 5.5 on the Artificial Analysis index.
From the TechCrunch AI article of October 12, 2026 and cited sources: Meta split access between the two models. [1] [3]
SOURCES[1] Meta releases its first frontier model, Muse Wonder techcrunch.com[3] AI Weekly: Meta releases Muse Wonder aiweekly.co
AT Pulse research deskPublished · Updated · 4 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- TechCrunch AI 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 13 edition
- Evidence
- 9 sources across 9 sites
- 1Meta releases its first frontier model, Muse Wondertechcrunch.com
- 2Meta has some Wonder Full new AI modelstheverge.com
- 3AI Weekly: Meta releases Muse Wonderaiweekly.co
- 4Unite.ai: Meta unveils Muse Wonder seriesunite.ai
- 5DEV Community: Muse Wonder costs what Opus 5.5 costs…dev.to
- 6The Decoder: Muse Wonder Full breaks into the AI top tierthe-decoder.com
Reach: not measured yet
Meta split access between the two models. Muse Wonder is available through the Meta Model API at the same price as Claude Opus 5.5. Wonder Full is not offered through the API and comes only with Muse Code's $50 or $200 monthly tiers, so teams can't plug Meta's strongest model into their own tools. [1] [3]
Artificial Analysis rates Wonder Full at 56 and Wonder at 52. That places them ahead of GPT-6 Astra (53) and Muse Spark 1.3 (48) but behind Opus 5.5 (58). Gains show up in agentic and coding tasks: Terminal-Bench rose from 33% to 57%, and Wonder Full ties Claude Code with Sonnet 5.5 on the Coding Agents Index. Factual recall still trails Opus 5.5 by a wide margin. [6] [3]
Meta's model card uses the best of three scores for each benchmark and labels its runs of competitor models as best-effort, so its numbers are vendor-reported and favorable to Meta. Even so, Wonder loses to Opus 5.5 on 9 of Meta's 14 benchmarks. The largest gap is on ProgramBench. Wonder Full's lead on DeepSWE is Meta's own figure. [5] [6]
Meta's main pitch is agent features. Active Memory keeps context across sessions, and Active Focus keeps multi-step agents on task as they move between apps. Both are meant to power the Muse personal agent, which follows Meta's move from open Llama models to the proprietary Muse line. [3] [1] [7] [8]
Artificial Analysis places Wonder Full near Opus 5.5 but below it, with weak factual recall.
Read the full assessment
Implication: API buyers get a same-priced option that mainly makes sense for long-horizon coding and memory-heavy agents, while Meta keeps its strongest model inside its own subscription.
Executive brief
On October 12, 2026, Meta released Muse Wonder and Muse Wonder Full, its first frontier-class models. The stronger one, Wonder Full, can't be called through Meta's API. It only comes with a Muse Code subscription. The API model, Muse Wonder, costs the same as Claude Opus 5.5 ($4/$20 per million tokens) but loses to it on 9 of Meta's own 14 benchmarks. On the Artificial Analysis index, Wonder Full scores 56 against Opus 5.5's 58. Meta's selling points are long-horizon coding and "active memory." A full safety report has not been published yet.
What changed and event timeline
Wang hired after Llama 4 stumbles
Meta hired Alexandr Wang as Chief AI Officer after the Llama 4 release in April 2025 went badly. That set the direction for the Muse line ().
Muse Spark launches
This was Superintelligence Labs' first model. It is proprietary, not open-weight, and runs in Meta's apps, with API access limited to a private preview ().
Muse agent debuts
Muse is a proactive personal agent that runs on cloud virtual machines. It has a free tier plus paid tiers at $20 and $100 a month ().
Wonder and Wonder Full released
Meta calls them the next step on its "scaling ladder." Both were trained at its Prometheus data center in Ohio and will power the Muse agent ().
First third-party index scores
Artificial Analysis rates Wonder Full at 56 and Wonder at 52. Muse Spark 1.3 scored 48 on the same index ().
Capabilities and access
- Muse Wonder: available through the Meta Model API and Muse Code. Costs $4 per million input tokens and $20 per million output tokens (TechCrunch).
- Muse Wonder Full: Muse Code only, on the $50/month Power or $200/month Maximum tier (AI Weekly). Not served through the API.
- Headline features: "Active Memory," which keeps context across sessions, and "Active Focus," which keeps multi-step agents on task when they move between apps (AI Weekly).
Read the full section
- Muse Wonder: available through the Meta Model API and Muse Code. Costs $4 per million input tokens and $20 per million output tokens (TechCrunch).
- Muse Wonder Full: Muse Code only, on the $50/month Power or $200/month Maximum tier (AI Weekly). Not served through the API.
- Headline features: "Active Memory," which keeps context across sessions, and "Active Focus," which keeps multi-step agents on task when they move between apps (AI Weekly).
Technical analysis for researchers and developers
- Architecture: Meta has not disclosed parameter count, architecture, or context window (Unite.ai). That is an inference, not a disclosed fact.
- How Meta scored itself: the model card uses the highest of three numbers for each benchmark: Meta's internal run, the self-reported score, or the leaderboard score.
- Cost per task: about $4.52 per task on the Artificial Analysis suite, close to 3x Muse Spark 1.3.
Read the full section
- Architecture: Meta has not disclosed parameter count, architecture, or context window (Unite.ai). The Decoder reads the weak factual-recall scores as a sign of a smaller base model. That is an inference, not a disclosed fact.
- How Meta scored itself: the model card uses the highest of three numbers for each benchmark: Meta's internal run, the self-reported score, or the leaderboard score. Competitor runs are described as "best-effort," and reasoning-effort settings vary between benchmarks (DEV Community).
- Cost per task: about $4.52 per task on the Artificial Analysis suite, close to 3x Muse Spark 1.3. Wonder uses 28% more input tokens and 10% fewer output tokens per task (The Decoder).
Claims and evidence
Third-party measurements (Artificial Analysis), all from The Decoder: Meta's own numbers:
Read the full section
Third-party measurements (Artificial Analysis), all from The Decoder:
- Agentic work: Terminal-Bench 4.0 rose from 33% to 57%.
- Coding: Wonder Full scores 68 on the Coding Agents Index, tying Claude Code with Sonnet 5.5.
- Hallucination: the rate fell from 33% to 23%.
Meta's own numbers:
- DeepSWE 1.1: Wonder Full scores 79.0, against about 74 for Opus 5.5 and GPT-6 Astra (The Decoder).
- MemoryBench: Wonder beats Fable, Opus, and Astra (TechCrunch). No independent check of this claim was found in the reviewed sources.
- Overall standing: Meta's testing puts both models on par with Fable 5.1 and GPT-6 Astra (TechCrunch). Artificial Analysis also puts GPT-6 Astra at 53.
Context and prior work
- Break from Llama: Muse Spark ended Meta's open Llama approach. The company says it hopes to open-source later versions (Fortune).
- Agent product: the Muse agent already had user-editable Memory files and a Goals tab before Wonder launched (Built In).
- Development time: about 14 months under Wang (AI Weekly).
Limitations, safety and contested findings
- Factual recall: on AA-Omniscience, Wonder Full scores 30 and Wonder 25, against 46 for Opus 5.5 (The Decoder).
- Biggest benchmark gap: on ProgramBench, Wonder scores 50.0 to Opus 5.5's 70.5 (DEV Community).
- Safety: Meta says these are its strongest safeguards yet against CBRN and cyber misuse.
Read the full section
- Factual recall: on AA-Omniscience, Wonder Full scores 30 and Wonder 25, against 46 for Opus 5.5 (The Decoder).
- Biggest benchmark gap: on ProgramBench, Wonder scores 50.0 to Opus 5.5's 70.5 (DEV Community).
- Safety: Meta says these are its strongest safeguards yet against CBRN and cyber misuse. Only summary safety evaluations have been released so far, with the full report still to come (Unite.ai).
- Positioning: Meta's "comparable to frontier" framing and the Artificial Analysis ranking (below Opus 5.5) point in different directions.
Business and practitioner implications
- API buyers: at the same price as Opus 5.5, Wonder makes sense mainly where it wins: long-horizon coding, legal agent work, and memory-heavy agents (DEV Community).
- Lock-in: putting Wonder Full behind a subscription means teams can't plug Meta's best model into their own tools.
- Budgets: plan for higher costs per task than Muse Spark, driven by more input tokens (The Decoder).
Read the full section
- API buyers: at the same price as Opus 5.5, Wonder makes sense mainly where it wins: long-horizon coding, legal agent work, and memory-heavy agents (DEV Community).
- Lock-in: putting Wonder Full behind a subscription means teams can't plug Meta's best model into their own tools.
- Budgets: plan for higher costs per task than Muse Spark, driven by more input tokens (The Decoder).
- Before adopting: run your own evaluations, since many of the headline results are Meta's own best-of-runs numbers.
The source trail.
Sources (9)
Meta releases its first frontier model, Muse Wonder
Article text retrieved; extracted text may omit tables or interactive elements.
techcrunch.comMeta has some Wonder Full new AI models
Related coverage; assess separately
theverge.com